SprintpcbGroup

Head-Up Display PCB Design: Thermal, High-Speed and Heavy Copper Trade-Offs for TFT, DLP and LBS HUDs

sprintpcbgroup — Tue, 10 Nov 2026 07:01:00 +0000

The PGU Debate Is a False Question

After nearly ten years in automotive electronics, with no fewer than twenty HUD PCBs handled, I increasingly feel the so-called debate over PGU routes is in fact a false question. Many hardware teams spend months arguing over TFT, DLP or MEMS, and the board that finally comes out still heats where it should heat and still jitters where it should jitter. The problem is not which technical route you chose but whether you have a real feel for the physical limits of the PCB behind that route.

A TFT-LCD Case: When the Backlight Baked the Copper

Take the case that gave me the most headache. I took on a volume AR-HUD project, with TFT-LCD decided for the PGU side. Many people feel TFT is low in cost and has a mature supply chain, and that on the PCB it is nothing more than a video bridge chip plus several bias drivers, with no difficulty at all. The result was that when the first prototype came back, after half an hour of lighting the screen, the copper foil temperature near the backlight LEDs soared to 120 degrees, and the whole Head-Up Display PCB deformed until the light path went off directly. It was not that you had not calculated the heat path, but that you had the relationship between heat sink and copper thickness wrong: the backlight power of TFT is poured into the PCB by brute force, and you think adding a heatsink settles it, but in fact heat runs laterally in the inner layers of the multilayer board and stirs the impedance of high-speed differential pairs all up. At this time you learn that it is not a matter of casually finding a heavy copper PCB supplier to thicken the copper and call it done, and you must start again from the stackup and thermal channel, which forced me to switch temporarily to a high frequency PCB manufacturer able to make mixed dielectrics, using high-thermal-conductivity FR-4 mixed with PTFE to separate the hot zone from the high-speed zone, at the price of the unit cost rising by nearly thirty percent.

Why I Lean Toward DLP

So I actually have more goodwill toward the DLP route, not because its optical efficiency numbers look good but because its heat-generating area is naturally better isolated from sensitive signals. The DMD chip itself does not get very hot, and the real heat sources are on the PMIC of the LED driver and the DMD controller, which can be pushed directly to the board edge in PCB physical position, with a separate copper-based area carved out to serve them. A recently cooperating hdi pcb supplier helped me make a 2+N+2 stackup, burying all the high-speed LVDS traces of DLP on L2 and L5, with the surface layers almost entirely power copper and heat dissipation paths, and even running the 1.5 GHz DMD clock was clean. Some people at every turn complain that the DLP chipset is too tied to TI, but plainly speaking, as long as you understand the PCB as part of the chipset and do not force generic design thinking onto it, the margin of this set of things is actually very wide.

The MEMS Lesson

On the MEMS route, however, the losses I have taken explain the problem better. Two years ago I was set on making a compact HUD PCB and chose a PGU architecture based on a piezoelectric MEMS mirror and laser diodes. At the time I liked its small optical engine volume and thought the PCB could be made palm-sized, but when drawing the board I found that the laser diode driver of each color channel was a high-speed pulse current, with di/dt so large that even with three ground planes poured I could not press down ground bounce, and the SerDes used for pixel timing on the FPGA side lost lock all the time. Later I changed the board to four layers and it was still unstable, and finally I switched to a high speed PCB manufacturer that truly had accumulated depth in high-speed signal simulation, rebuilt an S-parameter model of the whole board, and relied on a complete signal return design to open the eye diagram. This experience cost me a full seven months, but made me completely give up the illusion that the MEMS route is simpler. The PCB for MEMS is not at all a low-speed control board in the traditional sense. It is a board mixing high-precision analog drive and high-speed digital timing, and its requirements on the process capability of an hdi pcb manufacturer are no lower than for any high-end phone motherboard.

Functional Safety Written Into Copper

There is another easily overlooked pit, the layer of functional safety. When ISO 26262 ASIL B requirements land on PCB design, it is often not enough to muddle through by adding two redundant power supplies. You have to consider physical isolation of single-point faults, for example placing the MOSFET of the safety shutdown path in an area that does not share copper with the main PMIC, or in a MEMS scheme putting the traces of the monitoring MCU and the laser driver on different layers to avoid both being finished after a short circuit. Once these constraints come up, the board’s area and layer count cannot be held down, and you can only look for a heavy copper PCB manufacturer able to do a one-stop service with mixed lamination, thick copper and laser blind via full stackup. This is not simply a supply chain problem but whether you can translate your own safety concept into a story of copper and dielectric.

In the end, my experience from these projects boils down to one sentence

Choosing a Board Factory for HUD: Copper, Loss and Stackup First

Some time ago someone asked me how to pick a factory for making a head-up display board, and my first reaction was that this question is actually more grinding than picking the display scheme itself. There are quite a few manufacturers on the market who can claim to make a Head-Up Display PCB, but when it comes to handling large-current backlight, high-speed video signals and that pitifully small heat dissipation space at the same time, few can fight. In these two years I have dealt with quite a few factories and slowly worked out a knack: do not rush to stare at the architecture differences between TFT and laser scanning, first digest copper thickness, dielectric loss and stackup structure, and the road of choosing a supplier afterward can go smoother by a large stretch.

Most teams working on HUD PCBs, as soon as the scheme comes out, are easily led by the nose by display brightness, thinking that as long as the light source is powerful enough it is fine. In fact the pits I have stepped in lie exactly in the PCB itself: when large current cannot be carried and the copper foil heats up, impedance soars, and however high the brightness, it is in vain. What impressed me most was one time trial-producing a high-speed board with heavy copper areas, where I first found a factory that advertised itself as a Heavy copper PCB manufacturer, and when the sample came back and was powered up, the copper temperature was directly hot to the touch, and the power layer inside the four-layer board simply could not spread the heat evenly. Later I switched to another, who directly mixed the thick copper core board and high-speed dielectric during lamination, and although yield was lower, the heat dissipation path was entirely different. This made me thoroughly understand that for making this kind of hybrid structure, you have to find a team that has really run a large volume as a Heavy copper PCB supplier on the production line, not just look at the multi-ounce numbers printed in a brochure.

High Frequency and High Speed Are Different Problems

There is another easily overlooked point, which is that high-frequency characteristics and the high-speed digital part are in fact a pair of enemies. A serial video stream in a HUD easily runs at several Gbps, and if the trace is slightly longer, the eye diagram collapses immediately. Many people think high frequency equals high speed and go everywhere looking for a high frequency PCB manufacturer, and the result is that the board they get has a lower dielectric constant but a serious glass weave effect, and skew takes off as soon as the differential routing is long. I now prefer to find a high speed pcb manufacturer with specialized high-speed material selection experience, who will at least discuss with you whether to go with M4 or M6 grade rather than throwing you a Rogers 4350 specification sheet and being done. After all, what truly needs strict impedance control and delay control in a HUD are those LVDS or eDP channels, and this work cannot be directly transplanted from the experience of high-frequency RF.

Volume and the HDI Requirement

As for volume, there is no need to say more: the optical engine of an AR-HUD today almost has to be stuffed into the gap of the instrument panel, and PCB area is squeezed until it cannot breathe. Several layers of power, video processing and LED matrix drive all have to be squeezed into a palm-sized place, and an ordinary through-hole board simply has no way. At this time I seriously screen a round of hdi pcb manufacturers, wanting those that can do any-layer interconnect and keep laser drilling aperture stably within 0.1 millimeter. It is not that microblind vias are omnipotent, but facing that big row of boost inductors of the TFT backlight, using HDI to shrink the switching loop to the extreme can bring EMI down by a notch. When looking for an hdi pcb supplier, I generally ask them directly whether their minimum via-in-pad can reach 0.2 millimeter, and those who answer vaguely are sure to cause problems in later cooperation.

Manufacturing Is Not a Detour Around Display Technology

Many people think I talk about these things to bypass display technology and talk manufacturing, but it is exactly the opposite. Whether you use TFT plus an LED array or DLP plus a lens group, in the end the path of current, the return of signals and the conduction of heat all have to be drawn stroke by stroke on the PCB. However fancy the technical route, if the board’s heat dissipation structure is rotten, strong light still shows you its colors. I would now rather spend my time grinding copper thickness allocation and dielectric selection with supplier engineers, which is far more substantial than arguing in a meeting room over which scheme has a few hundred percent higher brightness. In the end, in the HUD trade, the PCB looks like the least sexy link, but only when sunlight pours back and burns the screen or snow noise appears will you know that a solid board is more useful than any algorithm.

Backlight Heat, HDI Need and Supplier Qualification

When I did HUD projects for several car makers in the past, I gradually realized a fact ignored by most: everyone’s mouth is full of how to choose the TFT bias chip and which video bridge scheme to use, but what truly stalls a project or even forces a teardown and restart is often the PCB itself. This HUD thing is mounted in front of the instrument panel, and in summer with direct sunlight plus backlight LEDs running at full power, local board temperature can easily soar past a hundred degrees, and if the constant current source and boost circuit of the backlight driver use only ordinary FR4 with 1oz copper, the thermal camera scans out a sheet of red, and long-term reliability simply cannot hold. Later I simply first pinned down the capability circle of heavy copper PCB suppliers, confirming that they can make 4oz or even 6oz inner and outer layers, pulled the current carrying and heat dissipation channels of the backlight loop to the full, and only then considered chip selection. If the heat problem is not solved well, once the TFT’s liquid crystal layer overheats, response speed drops straight down, strong light shows ghosting, and user complaints all arrive.

Another underestimated point is the necessity of the HDI process on the HUD mainboard. Between the video bridge chip and the flat cable interface of the TFT, high-speed differential signals run, and to control EMI, quite a few layouts require length matching and impedance matching, and the layout is so tight that laser blind vias must be used. At this time finding a reliable hdi pcb manufacturer takes far more effort than choosing a bridge chip. I have paid for this: to be cheap I changed to a small-scale hdi pcb supplier, and the result was batch-wise hole wall fracture and picture flicker lines after high and low temperature cycling, and locating and troubleshooting tossed me for over a month. Later we wrote the requirement for a high speed PCB manufacturer directly into the supplier qualification threshold, and assessed separately the heavy copper of the backlight drive part, the HDI of the signal part, and the high Frequency PCB manufacturer for the high-frequency digital part, not allowing one factory to do everything but do everything just so-so. High frequency especially, because many HUD mainboards also have to integrate FM radio, Bluetooth or even V2X modules, and for the RF area you must find a high Frequency PCB supplier that can stably supply low-loss board material, otherwise the signal-to-noise ratio is a mess.

The pits stepped in over the years on HUD PCB selection have made me understand that however mature the backlight and TFT drive scheme, in the end everything is carried on the board. Rather than agonizing over chip datasheets for half a month, it is better to first go to the PCB manufacturer and settle the basic items: stackup scheme, copper thickness distribution, line width and spacing capability and solder mask temperature resistance. The Head-Up Display PCB is by no means a copy of ordinary automotive electronics: it demands heavy copper, fine pitch, high frequency and high temperature all at once, and with the four demands superimposed, the suppliers to choose from are few. When I do a project now, the first step is to thoroughly assess the process cross-evaluation of heavy copper PCB manufacturers and hdi pcb manufacturers, confirming that they can master both large backlight current and high-speed signals on one board, and then go on to selection, and detours are cut by a full half.

Four Needs at Once: Thermal, ASIL, HDI and High Frequency

In these two years I have handled several head-up display projects, and honestly, this field’s requirements on the circuit board are far more twisted than I imagined. Many people start by staring at optical design, feeling that once the virtual image is done well everything is fine, but what really keeps one awake at night is often that HUD PCB hidden inside the heat dissipation shell.

At that time there was a project that had me find a manufacturer able to make a Head-Up Display PCB, and at first I thought that any factory that had made automotive boards would do, but when the sample came back and was powered on, I hit a wall. The LED driver chip on the board heated directly to power derating under large current, and measuring the copper thickness, I found that the 1oz copper given by the ordinary factory simply could not withstand the sustained impact of peak current. Later I simply changed that whole block of the power loop to 3oz copper thickness and found a factory specializing as a Heavy copper PCB supplier, and only then was junction temperature pressed down. Changing to this factory also had an unexpected gain: they were themselves a Heavy copper PCB manufacturer, understood thermal management thoroughly, and had aligned via filling and the connection of thermal interface materials for me in advance, saving the effort of tearing with structural parts later.

ASIL is also a pit. Many people feel that as long as the main chip passes ASIL B, the whole board is automatically safe, but it is not so. In making high-speed differential traces, I found that board-level crosstalk and impedance mismatch also cause timing jitter, and once the video timing monitoring circuit misjudges, the system throws you a frozen frame. Later I simply handed the high-speed link to a factory working as a high speed PCB manufacturer, whose dielectric material control was more stable than ordinary FR-4, and the impedance consistency report looked at last reassuring.

The HDI part cannot be saved on either. HUDs are getting smaller, with main chip ball pitch dense, and 2-order or even 3-order buried and blind vias are standard. At first, being cheap, I found a small hdi pcb supplier, and the sample came with microvia alignment off in a mess, with uneven hole copper thickness from laser drilling, and when running OpenLDI signals the eye diagram closed directly. Later I changed to a fairly large hdi pcb manufacturer, whose laser drilling precision and copper filling process were clearly mature, link loss simulation matched measurement, and there was no more repeated prototyping.

The high-frequency part was not as fragile as I thought. Some peers insist on throwing HUD boards to outrageously expensive high-frequency board factories, piling on Rogers material at every turn, and I did not follow this trend. I used a hybrid dielectric scheme, laminating a layer of high-frequency material only on the layers that need to run high-speed serial links and keeping ordinary halogen-free FR-4 on the other layers. I gave this work to a factory that was both a high Frequency PCB supplier and a high Frequency PCB manufacturer, who could calculate the post-lamination warpage for you, and cost did not double, and yield was steady after volume production.

In the end, the HUD board is not hard because of any single technology but because there are too many hybrid processes, with heat dissipation, high speed, high-density interconnect and functional safety all squeezed into a palm-sized place. What you want to find is not a factory strongest in some single item but a team that can digest this pile of contradictions at once. My cooperating partner now is basically one factory covering everything, with both a heavy copper production line and experience in HDI and high-speed boards, so that when problems arise, there is no passing the buck among several suppliers, and for a product doing functional safety, this matters more than anything.

Thickness, Layers and Reliability on a DLP Board

A few years ago when I first came into contact with automotive head-up displays, my understanding of the PCB was still at the stage of as long as the lines connect. Later I was taught a hard lesson by reality. This HUD thing looks like it only projects an image onto the windshield, but the small Head-Up Display PCB behind it is in fact a monster integrating optics, thermal management and high-speed signal integrity.

Many people discuss TFT and DLP routes, arguing red in the face, but what I care about more is that whichever imaging technology you use, once the copper thickness and interlayer structure on the board are chosen wrongly, later batch problems can make you shed a layer of skin. Especially in a DLP scheme, the power supply and high-speed control signals of the DMD chipset are deadly sensitive to the PCB’s parasitic inductance. Once during debugging we always had the micromirror occasionally lose lock, and finally found that the copper impedance from the power layer to the DMD was too high and instantaneous current could not come up. Later I switched to a Heavy copper PCB manufacturer able to do hybrid lamination of thick copper and high-density interconnect, raised the copper thickness of the drive loop to 4 ounces, and only then did the problem calm down. After that I understood that finding the right HUD PCB supplier is ten times more important than studying the typical application diagrams in the chip manual.

The same pit was stepped in on the backlight drive part. The current of an LED matrix is not small, and the pulse width modulation frequency runs to several hundred kilohertz, and ordinary FR-4 board material drifts badly in dielectric constant at high temperature, so brightness and color coordinates jump with it. At that time a friend making high-frequency boards suggested that I directly use low-loss High frequency PCB and integrate the drive loop and heat dissipation copper into the same multilayer board. At first I thought it expensive, but later I ran the numbers, and compared with being returned by the car maker because of optical performance drift, the cost of this board material is simply negligible. When choosing a HUD PCB partner now, my first requirement is that the other side must have the dual capability of a high frequency PCB supplier and a Heavy copper PCB supplier, otherwise I will not even talk.

Miniaturization and the Real Meaning of Reliable HDI

Another easily overlooked place is miniaturization. The new generation of AR-HUD is getting smaller in volume while having a pile of functions, and PCB area is compressed to fingernail level while still having to hold a 0.3-inch DMD, PMIC and a complete set of control chips. At such density an ordinary through-hole board simply cannot route, and any-layer interconnect must be used. I have contacted quite a few factories claiming to make hdi pcb, and when the real prototypes came back, the copper filling flatness of laser blind vias was pitiful, and chip soldering yield was so low I doubted life. A truly reliable hdi pcb manufacturer must have reverence for automotive-grade reliability, and details such as hole copper thickness and stackup symmetry cannot be fudged even slightly. Later I permanently cooperated with one, whose engineers would even proactively remind me which pads’ line spacing might be affected by Z-axis expansion after reflow and suggest changing the via-in-pad layout, and this kind of experience cannot be bought just by expensive equipment.

Looking back, whether the HUD’s imaging effect is good is of course strongly related to optical design and chipset algorithms, but the final landing point of all those precise calculations is the PCB. If a board’s signal integrity collapses or the copper cracks after thermal cycling, however good the optical engine, it is wasted. So in selecting suppliers now, I would rather spend more time early assessing their high-speed and high-frequency mixed lamination capability and the maturity of their heavy copper process, than be on tenterhooks in volume production. The Head-Up Display PCB track, in the end, does not compare who draws the drawing more beautifully but who can stay stable for years under the triple stress of heat, electricity and force.

DLP in Practice: High-Density, High-Speed, High-Current Together

When I looked at the HUD circuit board for a friend’s project two years ago, I had a very deep feeling: for the DLP road, too many people do armchair strategizing, and only those who have really touched the board know how muddy the water is. Outside everyone is hyping DLP for high brightness and beautiful contrast, but once it lands on the PCB, with the dozens of high-speed differential signals of the DMD laid out, the difficulty of the whole layout jumps up by an order of magnitude directly, simply not something an ordinary six-layer board can carry. At that time I contacted no fewer than five factories advertising that they could do high-density interconnect, and those who could get a Head-Up Display PCB through a first prototype without problems could be counted on one hand.

Many people burn their energy on optics and structure and instead feel the circuit board can be made do with by any hdi pcb manufacturer, and this idea is especially harmful. The DMD’s pixel clock is not low, and if the trace is slightly mishandled, the ghosting and double images projected will make you suspect the lens group has a problem, when in fact it is that high-speed signal integrity has been smashed. And the pit of heat dissipation is deeper: the heat of that DLP optical path is concentrated in a very small area, and if the back of the DMD does not rely on heavy copper to pull the heat away quickly, brightness starts to fall in less than twenty minutes. Later I simply split the power part from the DMD substrate and found a factory that was both a high Frequency PCB manufacturer and had Heavy copper PCB supplier capability, having them pile the copper thickness to four ounces and then run fine dense traces on the outer layers. Few can do this hybrid process smoothly, and most, on hearing that both high-frequency board material and inner-layer thick copper had to be handled, declined directly.

It is not that a fully separated architecture has never worked. One project made the LED driver and logic control into two boards connected by a flexible flat cable, but the continuity requirement of the grounding system was outrageously high. Once, the grounding loop was not handled well, and the DMD reset signal was interfered with by bounce at irregular times, and at the car maker’s trial installation it lit up once and then went black, and troubleshooting took nearly a month. Since then I learned my lesson: when looking for a high-speed PCB supplier, you cannot look only at their line width and spacing capability, and you must ask clearly whether they have truly run a reference board with MIPI DSI or parallel RGB interfaces, since verbal claims are useless. Qualified hdi pcb suppliers now basically honestly put effort into stackup design and impedance consistency, but when heavy copper collides with fine lines, yield falls again.

In the end, what the DLP scheme brings to the HUD is not only a matter of a DMD chip: it nearly gathers all the difficulties of board-level design together, high density, high speed, high frequency, large current and small-area heat dissipation, and if any link cannot be held, the performance of the whole optical engine does not come out. I have seen too many schemes die on the PCB while the team was still changing lenses and software, with the direction all off. If next time someone talks to me about a DLP HUD, my first sentence will certainly be to ask: has the board factory you chose really made an automotive-grade Head-Up Display PCB? If not, do not discuss anything else yet.

Dielectric Drift, Differential Copper and Hot Spots

Take a HUD project: many engineers start by following the chip manual to place the PMIC tight against the DMD and the DLP controller, feeling that once the power supply loop parasitic inductance is low enough everything is fine. But once it actually runs, you find that what is really a headache is often that the interlayer dielectric of the PCB begins to shift after thermal cycling, and once the microstrip impedance floats, the eye diagram of high-speed differential pairs immediately closes by half. At this time, changing to a so-called high-frequency board, the cycle simply cannot make it. So when I select a high frequency PCB manufacturer now, I look not only at the nominal DK and DF values but care more about their pairing experience in glass fiber cloth and resin systems: some factories are good at communication base stations, but with the same material stackup, taken to a board type like a HUD that is irregular, thin and also needs local mixed lamination of copper thickness, they flip over directly.

There is also something that may be contrary to most people’s intuition. The DMD itself indeed has low power consumption, but the row of LED driver MOSFETs beside it eats current without any hesitation, and the current spike at the switching instant can pull local ground bounce to several hundred millivolts. If at this time you mechanically look for a heavy copper PCB manufacturer and pile copper thickness from 2oz to 4oz or even 6oz, thermal resistance goes down, but the splitting of the inner-layer power plane is destroyed into a mess, and incidentally the reference layer of the DMD’s low-speed control signals is cut into pieces. My own habit is to find an hdi pcb supplier that can do differentiated copper thickness, using laser blind vias to connect the drive loop directly to the thick copper top layer, keeping the signal layers at standard copper thickness, and using multi-order any-layer interconnect to truly separate digital ground and power ground, which is far stronger than simply adding copper. Earlier an hdi pcb manufacturer, to cut cost, pushed a stacked via scheme to me, and I did not take it, because the resonance frequency of the few micromirrors under the DMD is sensitive to mechanical stress, and the stress concentration risk of stacked vias simply cannot be explained clearly.

Back to the MEMS laser scanning line: what I think is most underestimated is the synchronization precision between the high-speed pixel clock and the mirror drive. Everyone stares at the narrow pulse current of the laser diode, as if finding a high speed PCB manufacturer to control the rising edge is enough, but in fact when the scanning frequency approaches the resonance point, the phase jitter of the mirror feedback signal causes dynamic distortion across the whole picture. At this time it is not enough to have low-loss board material, and the length-matching constraint on the PCB has to change from the DMD to the whole link from the FPGA to the mirror driver, and the stub of every differential clock pair has to be pinned within 5 mils. Many high Frequency PCB suppliers who claim they can do it cannot deliver when intra-pair skew is stuck at plus or minus 2ps, and to ship stably in volume you can only change to a factory with a harder process, and those who have not stepped in these pits simply cannot realize them.

Why LBS Is Not Ready to Scale

In recent years in the head-up display field, quite a few people regard the LBS set of things as the next-generation standard, and chatting about it at conferences they beam with delight, as if once laser plus MEMS scanning comes out, all optical difficulties are resolved. I instead feel that the deeper we dig on the PCB side, the harder it is to follow this wind. Selecting a HUD route is in essence choosing a pit for yourself, and the pit LBS digs is the least conspicuous kind.

Not to mention anything else, the substrate hurdle alone is hard to clear. A decent LBS optical engine, even if only an engineering sample, puts rather contradictory requirements on the Head-Up Display PCB. Laser drive current is not large, but timing cannot jitter, and a deviation of a few dozen picoseconds can make a pixel drift a full body position. This means the few high-speed differential signals on the board need extreme impedance continuity, and ordinary FR-4 gets as soft as rubber candy at 85 degrees in a car, and once the dielectric constant drifts, picture edges blur immediately. More troublesome is the MEMS mirror drive, which wants high slew rate and low-noise analog traces, while laser modulation is a nanosecond-level pulse current loop, and with these two kinds of signals squeezed on one board, if crosstalk handling is slightly careless, the noise is livelier than the snow of an old TV. Many that call themselves high speed PCB manufacturers, on receiving this kind of stackup requirement, still honestly send it back to change the scheme, because once the automotive temperature cycle runs, the promised interlayer alignment accuracy is all discounted.

I have also contacted several hdi pcb manufacturers who have been pushed up to the altar, and on hearing that laser microvias have to be stuffed into a 3-N-3 structure and copper blocks embedded for heat dissipation, the quotation sheet directly becomes hard to look at. It is not that their technology is not good but that LBS is too sensitive to the thermal path. The area of the laser chip is so small that the instantaneous heat flux density is higher than we imagine by more than an order of magnitude, and relying on the little thermal conductivity of FR-4 itself is simply asking an elephant to walk a tightrope. So some people turn to look for manufacturers that can make thick copper inlays, but a Heavy copper PCB supplier is usually more used to the crude approach of power modules, and asking them to precisely stuff a thick copper island into a fine dense trace layer and still keep six-sided flatness gives yields that make one tear one’s hair. Forcing the process of a Heavy copper PCB manufacturer to be compatible with fine-pitch FPGA pads, cost aside, signal integrity is another debt never paid off.

Digging further down, the laser diode itself is hard to serve. Its wavelength shifts at high temperature, so the circuit has to do feedback control of temperature compensation, which adds yet another low-noise analog loop. To make optical power monitoring and adaptive drive current adjustment stable, trace shielding and ground bounce management are far trickier than imagined. I have seen a team that, to settle one LBS driver prototype board, changed three high Frequency PCB suppliers in succession, because the board material selection was not suitable, return loss above 6GHz fell unwatchably, and the transmit end efficiency collapsed with it. In fact however capable a high Frequency PCB manufacturer is, it cannot make decisions for the system, and board material loss and the board factory’s process boundary are hard right there. Some hdi pcb suppliers can indeed press out substrate with loss tangent below 0.002, but whether this number can still be held after automotive aging, no one so far dares to pat their chest on.

That said, the lifetime dispute of the MEMS mirror itself is less of a headache for me. What truly makes me feel LBS cannot reach volume that quickly is the whole supply chain’s ability to take on this precision analog plus high-speed digital plus optoelectronic hybrid board. Finding a board factory that understands lasers, high-speed layout and automotive reliability at once is harder than finding a Tier1 willing to spend money on volume production. In the HUD PCB trade, most of the time what is competed is not who has the higher indicators but who can gather the strange and varied process windows into a controllable cost range. The LBS scheme happens to tear the process window widest on the PCB, roughly equal to inviting all kinds of failure modes to a party.

So whether policy blows warm winds or demo cars amaze, if behind it there are not a few hdi pcb manufacturers who truly settle down to make automotive-grade high-density interconnect and are willing to run in a dedicated line for LBS, then however clean the picture on the PPT, it is hanging in the air. The road of laser display is not that it cannot be walked, only that the current PCB manufacturing ecosystem is far from ready to lay a steady step for it.

An LBS Project’s Detours: Substrate Loss and Supplier Mix

Some time ago a HUD project took many detours, and looking back now, the biggest pit was not at all in laser selection or FPGA resource evaluation but in thinking of the Head-Up Display PCB too simply. The engineering team at first all dived into the LBS route, with minds full of MEMS closed-loop bandwidth and SerDes jitter, feeling that as long as the safety circuit was independent and laser power monitoring was in place, the board could be run. The drawing was drawn beautifully, with four-layer buried and blind vias and high-speed differential pairs pulled neatly, and the result was that when the board came back from an hdi pcb supplier we did not know well, after power-up the eye diagram of the laser modulation signal closed by half directly, and MEMS scanning synchronization dropped frames from time to time. After troubleshooting to the end, the problem was in substrate loss and copper foil roughness: an ordinary hdi pcb manufacturer that has not made automotive high-frequency displays will simply not care about the sensitivity of your laser drive loop to 0.1dB insertion loss, and its process window was given by consumer electronics. This made me thoroughly understand that for the LBS route to land, you must first find a board factory that can handle high frequency, high speed and heat dissipation mixed lamination at once, otherwise ground bounce and crosstalk can make the safety monitoring loop falsely trigger every other day, and even if laser emission is within the limit, the display is a mess.

Later we changed suppliers around once, and slowly worked out a knack. Do not rush to find any high speed PCB manufacturer first, because quite a few board factories advertising high speed shake their heads directly when they see you want to mix lamination of copper base and low-Dk dielectric. Once, chatting with a heavy copper PCB supplier, they were quite honest and took out a sample of local 10 ounce thick copper that they had previously made for a lidar, but once I said I wanted to fit 0.2mm pitch BGA routing and microstrip on the same board, yield immediately fell below sixty percent. What finally covered the problem was a high Frequency PCB manufacturer who came in via an RF power amplifier background, who used laser drilling plus resin plugging to control the thermal resistance and electromagnetic isolation between the heavy copper area and the high-frequency area within an acceptable range. The engineering director of this high Frequency PCB supplier told me that what the HUD PCB lacks most at present is not design guidelines but cross-process integration experience, because on one board you have to serve at the same time the heat dissipation of a hundred-watt-level laser heat source, FPGA pixel timing with sub-picosecond jitter, and a functional safety loop that needs an independent hardware shutdown path, and this requires twisting the know-how of the traditional heavy copper PCB manufacturer, the HDI board factory and the RF board factory onto one rope.

On the safety side, the pits we stepped in also had nothing to do with the chip schematic. Everyone can recite the Class 1 limit of lasers and the three-channel shutdown logic, but in actual layout, once the analog ground of the detection loop and the high-speed digital plane share a split that is not decisive enough, the laser power monitoring value jumps at the MEMS resonance frequency, causing frequent system degradation. This forced us to rethink the stackup with the hdi pcb manufacturer and stuff two extra shielding ground layers in. Now when I look at any HUD route debate, whether TFT, DLP or LBS, I am used to first asking one sentence: does the hdi pcb supplier you cooperate with have the ability to carry together the scrap rate across the three dimensions of high-speed signal integrity, thermal management and functional safety? If not, however beautiful the architecture diagram, it is only a simulated waveform in a computer.

Choosing the Supplier First, Then the Design Details

In my years in the HUD trade, I have found that many people start by diving into the debate over PGU technical routes, TFT or DLP or LBS, which can be argued over a whole day. But what truly holds up project progress is often not the route itself but whether the HUD PCB you chose can run out steadily on the volume production line. Once we were burned quite badly by an hdi pcb supplier claiming to do high-density interconnect, with laser holes drilled off and plating uniformity a mess, and the scrap rate of the traces on the LED array drive side rose straight up. Later I simply changed my thinking: first screen suppliers, then settle the design details.

In fact the environment a HUD’s PCB faces is more perverse than ordinary automotive electronics. Local high heat brought by sunlight pouring back, an imaging unit in compact space stuffed with components, plus car makers now often requiring functional safety levels raised directly to ASIL-B or even D, all of this forces you not to stare only at conventional FR-4 and ordinary through-hole processes. Our practice now is that anything involving the LED backlight drive and power management block directly uses heavy copper PCB, since thick copper foil is much more tolerant of heat dissipation pulses and has smaller thermal resistance when bonded with a metal backplate. The heavy copper pcb manufacturers in China that can truly do well are in fact just a few, and their lamination capability and copper thickness uniformity are the hard threshold.

Judging High-Speed Capability by Evidence

The high-speed signal part is another headache. Do not look at the fact that a head-up display shows only some vehicle speed and navigation arrows: an AR-HUD now has to superimpose forward perception information in real time, with MIPI or LVDS trace length often exceeding 200 millimeters, and I have seen far too many boards with a mess of eye diagrams. In these days, if you call yourself a high speed pcb manufacturer but cannot even hold 100 ohm differential impedance within plus or minus 8 percent, you simply cannot get onto the supplier list. We even met a factory that used the wrong dielectric material, with DK temperature drift outrageous, and the equalization parameters tuned in winter all went haywire in summer. So later I developed a habit: first have the high speed pcb manufacturer take out their stackup calculation sheet and the test report of a 24GHz radar board they made before, which is more useful than how many system certifications.

As for high frequency, the HUD itself is not a typical RF board, but for a period we tinkered with integrating a Bluetooth and Wi-Fi module onto the HUD PCB for wireless firmware upgrade, and those few small areas were very picky about the etching precision and surface treatment of the high frequency pcb supplier. Immersion silver or ENIG and the shape of the solder mask bridge, these subtle places directly affect the corner continuity of characteristic impedance. At a high-frequency PCB manufacturer in Shenzhen specializing in high frequency, I saw them use laser direct imaging equipment to make the lines, with line width capability stable within 3 mil, and only then felt safe handing over the coupler structure.

Working Backward From the Supply Chain

On route selection, I instead feel that working backward from the supply chain to design is the most pragmatic. You may have built in the lab a prototype with DLP’s DMD plus RGB LEDs whose brightness and contrast are heavenly, but at volume production you find that the dedicated driver chip indispensable on that head up display pcb has a lead time stretched to 30 weeks, with only two package-compatible alternatives, so what do you do? After being burned, we now, at project initiation, pull the hdi pcb manufacturer and key chip distributors to sit down together and lay out all the BOM risk of alternative schemes and the PCB process capability boundary. Sometimes you even find that choosing a low-key route, using mature commodity LEDs and conventional HDI stackup, can instead shorten the project cycle by at least a quarter, without begging the supply chain every day. In this industry, the time window matters far more than parameter leadership.

Why HUD PCB Selection Is Deeper Than It Looks

Some time ago I chatted with a friend who makes head-up displays, who told me with a pained face that their project had been stuck on the PCB for a full two months. I was not surprised at all, because I have experienced this myself and stepped in even more pits than he. Many people feel that HUD is a job of light paths and images, and that the PCB is just finding a board factory to draw a board and prototype and done, but only after running through the automotive process do you know how deep the water of HUD PCB selection is.

When I first took over an AR-HUD project, all my attention was on the optical engine, and my requirement of the PCB stopped at it works. I casually found an hdi pcb supplier that had made automotive electronics, thinking it was just first-order and second-order buried and blind vias, nothing big. The result was that when the prototype was installed and hit with a thermal camera, the temperature of the power area soared directly to a completely unacceptable degree, and the driver chip went into self-protection from time to time. Later, on opening it, copper thickness was designed too conservatively, and the heat dissipation path had never been carefully planned. At that moment you realize that for something like a HUD sealed in a small volume, heat dissipation is harder than for many control modules, so later I repeatedly told newcomers that finding a Heavy copper PCB supplier that has truly done this is not an optional item but the basic lifeline.

Another thing I understood only after paying tuition. Earlier we thought HUD’s signal rate was not that high, since the distance from the main controller to the panel was just that little, but when the image chain was run, from the serializer to the deserializer segment, the eye diagram collapsed so that it could not be looked at. We tried changing cables and tuning terminations, and finally the problem was in board material and processing. Ordinary FR4 simply cannot control the return loss and crosstalk of that high-speed segment, and then when you look back, the so-called high speed PCB manufacturer and high speed PCB supplier are far more than just saying with the mouth that they can do it. Some board factories say they have low-loss board material and yet cannot even give batch consistency data of dielectric constant, and would you dare use such a thing?

Thick Copper Meets Fine Lines

More headache-inducing is that the HUD mainboard is often not a purely digital board. As brightness and heat dissipation requirements go up, many projects use thick copper, buried copper blocks or even copper inserts to spread the heat flow path, but then the problem comes: when thick copper and fine dense traces coexist on the same board, the challenge to etching compensation and lamination process rises exponentially. I have seen some heavy copper PCB manufacturers make the line width tolerance of 3oz copper like a roller coaster, causing power internal resistance to float into a mess. You cannot measure the problem on this kind of board at normal temperature, and once it goes through high and low temperature cycling, stress arises in the copper layer from thermal expansion coefficient differences, and the probability of via cracking goes up.

So when I look at HUD PCB selection later, I do not start by discussing which display technology is better, because that is a path set first by system and optics. Whether your PGU is TFT, DLP or a scanning mirror, what really affects reliability on the PCB side is in fact some things that sound very earthy: is the board factory’s alignment accuracy enough? Can their HDI process be compatible with heavy copper inner layers? Are high-frequency materials on their production line the real thing or just for show? These questions cannot be solved by flipping through a supplier directory, and often you have to run the production line to see whether the automotive HDI cases they have actually made have experience with a mixed structure of heat density like a HUD.

Having run through more board factories, you find a rather cruel reality: those who can do well as a high frequency PCB manufacturer do not necessarily manage heavy copper, and those good at power heavy copper processes often stumble in HDI stackup and high-frequency material processing. And a HUD board happens to be the intersection of these three, with high-speed video signals, local heavy-load current and high-density interconnect all indispensable. So later when we selected suppliers, the points of examination changed completely: first see whether they have truly mass-produced multi-order HDI boards that include both thick copper and impedance control requirements, and then look at their high-speed link simulation capability and actual test data. Those who cannot do it, however low the quote, we dare not touch.

The post Head-Up Display PCB Design: Thermal, High-Speed and Heavy Copper Trade-Offs for TFT, DLP and LBS HUDs appeared first on SprintpcbGroup.

Distributed Control System PCB: Redundancy, Hot-Swap and Heavy Copper Reliability in DCS Modules

sprintpcbgroup — Tue, 10 Nov 2026 07:00:00 +0000

Is Redundancy Overrated at the PCB Stage?

Some time ago a project gave me quite a headache: the I/O modules of the DCS on site dropped signals every now and then, and after checking back and forth, I found that the redundant switching circuit on the board was simply not synchronized with the operating rhythm. This made me reconsider a question: have we deified the word redundancy already at the PCB stage? Many system integrator friends, as soon as the DCS PCB comes up, start by demanding dual redundancy, hot swap and hardware voting, as if keeping a spare channel settles everything. But I increasingly feel that for a Distributed Control System PCB that has already gone to multilayer boards, heavy copper and high-density routing, redundancy is more like a double-edged sword: it preserves availability but quietly gnaws away signal determinism in places you cannot see.

Power Switching and the Heavy Copper Supplier

Take the power switching part. To let I/O modules be plugged and unplugged live, soft start and isolation must be built separately on that four-layer or six-layer board, and heavy copper is usually needed to carry the instantaneous large current. In earlier years we always went to one fixed multilayer pcb supplier for this, whose conventional multilayer board process was very stable, but once local 4 ounce or even 6 ounce copper thickness Heavy copper PCB requirements were met, delivery began to float. Once, in a board we got, the etch factor at the copper edge of the power plane had not been controlled well, making the equivalent resistance of the switching loop higher than the design value by a margin, and as a result, during hot swap the surge blew the fuse at the front end. At that time I realized that finding a manufacturer that truly understands the heavy copper process is far more important than adding several more redundant paths in the circuit. Unfortunately many peers, when selecting, stare only at layer count and line width and spacing in the quotation of the multilayer pcb manufacturer and overlook that thick copper is not on the same order of magnitude as ordinary FR-4 in Z-axis expansion rate.

Bus Interfaces Are the Hardest-Hit Area

The bus interface side is an even harder-hit area. For the isolation of HART or fieldbus signals, in essence the same differential pair has to carry both a digital heartbeat and superimposed analog disturbance. At the PCB level this is no longer a question of how many layers, but of how delicate the stackup and return path are. I once modified a DCS main control backplane, originally an eight-layer board, in which the power layer and ground layer were too close together, and parasitic capacitance ate part of the modulated signal edge, causing devices on the intrinsically safe side to fail handshake occasionally. Later I switched to another multilayer pcb supplier, who suggested running two inner-layer signals sandwiched between a pair of same-polarity reference layers, so that the return path for high-frequency components was cleaner. After the change, the bit error rate of the whole station dropped visibly to the naked eye, but do you say this had anything to do with redundancy? No direct relation. Yet the stability users finally feel happens to grow out of these non-redundancy details.

Thermal Cycling Life of Heavy Copper Boards

There is another place particularly easy to overlook, which is the thermal cycling life of heavy copper boards. The ambient temperature in a DCS cabinet is not as constant as in an IT machine room, and in summer, cards near furnaces may stay for a long time in a mild baking state of fifty or sixty degrees. If an ordinary heavy copper pcb manufacturer uses a high-Tg but low-toughness resin system, after three or four years the bonding force between the copper layer and the substrate begins to decay. The most outrageous case I have seen was a redundant main control card at a petrochemical site that suddenly went offline, and when it was pulled out, it was not that the chip had burned but that the power copper foil, after many thermal expansions and contractions, had lifted as a whole sheet from the root of the pad. After that, when choosing a heavy copper pcb supplier again, the first thing I asked was not how many ounces you can make but whether your thick copper boards have been through thermal shock validation of more than 500 cycles.

Repeatability Over Paper Redundancy

Ultimately, hardware like DCS that runs ten years without rest needs not a design declaration of several redundancies on paper but repeatability of every lamination and every etching on the production line. A redundant circuit can be drawn extremely elegantly in the schematic, but if your multilayer pcb manufacturer cannot even keep the alignment accuracy of inner layers from floating around, the backup channel may instead become a hidden interference source. My habit now is to lower the key of redundancy a bit at the system design stage and pour energy into seemingly clumsy things like board material, copper thickness uniformity and drilling quality. The longer I do hardware, the more I feel that a DCS PCB that is honest, with no cold solder and no interlayer shorts, lets one sleep more soundly than a hundred pages of redundancy analysis report.

An Offshore Platform Lesson: Redundancy and Hot Swap on a 12-Layer Board

I used to always feel that DCS hardware had no mysticism, that having drawn several four-layer boards and used isolated power supplies would settle it, until an offshore platform project completely changed my view. At that time, because of space limits, we had to stuff the redundant controllers and I/O links of the whole distributed control system onto one 12-layer board, and after finding several multilayer board manufacturers for prototyping, problems were exposed as soon as power was applied: not that the routing did not connect, but that redundancy and hot swap had been understood as paper labels, and almost paralyzed the entire bus system during live maintenance.

On that DCS PCB, we piled up two sets of fully symmetric computing cores, thinking hardware voting could achieve gapless switching. The result was that, without pinning down the copper thickness and impedance control of the power layer at a thick copper PCB manufacturer of the same batch, the reference grounds of the two redundant boards would fight under transient current of several dozen amps, jumping between primary and backup, not at all nanosecond-level synchronization but pulling several dozen milliseconds of glitches directly on the bus arbiter. Later we honestly found a factory that specialized as a heavy copper PCB supplier, pulled two power copper layers in the six-layer stackup to 5 ounces, and also buried TVS clamps and soft-start hot-swap controllers near the gold fingers of each plug-in card, and only then absorbed the surge current and common-mode noise at the instant of insertion and removal. The fieldbus part was even more of a headache: Modbus and CAN H run mixed with 4-20mA loops, cable lengths often over a hundred meters, and if the multilayer board supplier does not know how to sandwich split-ground isolation between layers and precisely control differential pair spacing, HART signals directly punch the analog readings to pieces.

My basic judgment of DCS hardware now is: what you choose is not a multilayer board manufacturer but the fault tolerance bottom line of your site. Hot swap is never for the convenience of replacing cards but the only chance of online repair, and redundancy cannot be patched together by duplicate devices either: it needs the physical layer of the entire power distribution network and bus to bury an escape route for single-point failure in advance. With these lessons, whenever I audit the stackup report of a heavy copper PCB manufacturer, I stare tightly at board thickness uniformity and current-carrying capability, because I know that thermal fatigue of any pad can, in some unknowing live replacement, become the starting point of the whole system’s shutdown.

Redundancy Is Proven on the Backplane

After so many years in process control, I increasingly feel that redundancy in the end must show its true colors on the backplane. However lively the primary-backup switchover and voting logic of a DCS is made, as long as the physical isolation of the Distributed Control System PCB behind it is not thorough, it gives you a taste of its colors at the critical moment. A few years ago we met a case where two redundant controllers synchronized through the backplane, and the previous revision used an ordinary multilayer board, with heartbeat constantly lost, and on checking, ground bounce had pulled the common-mode voltage of the synchronization link off. Later we switched to a multilayer board supplier specializing in power management boards, changed four layers to six, and separately allocated a fully shielded layer for peer communication, which finally stopped the jitter of primary-backup synchronization. That board also carried a power redundancy OR-ing circuit, with two 48V busbars from top to bottom having to carry 30 amps continuously, and ordinary 1 ounce copper could not take the heat. We specially found a heavy copper PCB manufacturer, with inner layers directly 4 ounce copper and outer layers 2 ounce, and running full current for three days, temperature rise was only about twenty degrees, so that channels were truly independent.

The voting circuit was even more elaborate: the hardware redundancy switch in the FPGA uses differential signal voting, and even a nanosecond-level disturbance can trigger a false switch, all depending on how accurately differential trace impedance can be controlled. So later we fixed on one multilayer board manufacturer and did not let it change, who could guarantee plus or minus 5 percent impedance tolerance, with line width spacing and normal-temperature drift all calculated clearly. There is another hidden place: the surge path when the primary and backup controllers are hot-swapped, and if the backplane heavy copper is not laid well, the ground potential floating at the instant of insertion can reset the backup machine’s logic, and troubleshooting this kind of fault can make one lose a layer of skin. Later, on new systems, when selecting a multilayer board supplier, I simply had the PCB factory mark the glue flow quantity and prepreg model of every layer on the stackup drawing, and without even one of these, I would not sign the sample.

In the end, redundancy is not piled up by protocols but rammed in micron by micron with copper thickness, stackup and isolation. Those heavy copper PCB suppliers that are truly useful will not just quote you an ounce count, but knead current density, heat dissipation path and mechanical stress together into the stackup simulation, and when a board like that is installed in the back frame, when the primary and backup switch over, you will not hear the sound of relays jumping back and forth.

Physical Layer Details Behind Modular Design

The longer you do industrial control, the more you feel that what is most easily overlooked is instead the board that carries everything. Many people talk about redundancy and hot swap, opening their mouths with protocols and logic, but once it lands on the physical layer, for those high-speed signals and high-current paths running on a Distributed Control System PCB, slight mishandling puts the reliability of the whole system in doubt. I once took part in a project where I/O modules sporadically restarted on site, and after two months checking code, I finally found the DCS PCB stackup had not been laminated well, and the resonance point of the power plane happened to fall in the switching noise band, and changing to another multilayer pcb manufacturer for re-prototyping made the problem disappear directly. Since then I no longer superstitiously believe in the original manufacturer’s reference design but force myself to thoroughly understand the board factory’s process capability.

The trouble brought by modular design is more than imagined. We always think that splitting functions into independent plug-in units settles everything, but in fact once a card is inserted into the backplane, the return paths of power, ground and signal all change. Especially during live insertion, millisecond-level differences in connector pin contact create ground bounce, and once this ground bounce couples into the control logic of the switching circuit, the switching action may be falsely triggered. Many people put attention on the symmetry of signal length matching, but I feel what deserves more attention is the collapse of the power rail at the switching instant. In my design I simply made a separate supply for the voting and switching part, taken directly from the backplane’s redundant power, isolated with Heavy copper PCB traces and ferrite beads, and not sharing wet connections with the VCC of any microcontroller. Facts proved that even if the main control module loses power, the switching logic does not move at all.

Heavy Copper Supplier Reality Check

Speaking of Heavy copper PCB, in fact the factories in China that can make thick copper boards well are just a few, and when looking for a Heavy copper PCB supplier you really cannot look only at website promotion. I once used a factory claiming it could make 10oz copper thickness, and the prototype that came out had serious inner-layer copper blistering, and delivery was also delayed by half a month. Later I switched to a Heavy copper PCB manufacturer specializing in heavy power boards, who used multiple laminations plus laser drilling to bring the temperature rise of the high-current module down to an acceptable level. In an industrial site, cabinet temperature in summer easily goes to 60 degrees C, and if the power module in a DCS rack relies on ordinary copper-thickness traces, IR drop makes voltage fall outside the error range, and then neither redundancy nor switching helps.

Multilayer Supplier Selection: Transparency Over Layer Count

In choosing a multilayer pcb supplier, what I value is not how many layers they can make but the transparency of impedance control and material library. Many unbranded board factories use cores and prepregs from small domestic factories, with extremely large batch fluctuation of dielectric constant, and the result is that the impedance of differential pairs floats outrageously. Once clocks and control signals reflect seriously, timing margin at switching is eaten up. Later I fixed on one multilayer pcb manufacturer, who directly provided measured data of Rogers and Taiyao board materials and had strict process specifications for keep-out zones and back-drilled stub control. You may not believe it, but just by changing the board factory, the bit error rate of the system bus dropped by two orders of magnitude.

Remote Sensing and Make-Before-Break Switching

A few more words about power. Many people, when drawing a Distributed Control System PCB, like to connect the feedback points of all power modules close to the module outputs. This looks fine, but with long-distance backplane power supply, the actual voltage received by remote I/O modules is lower. My approach is to design a Kelvin sampling point at the far end of the backplane, using a differential line to pull voltage feedback back to the compensation pin of the power module, which is equivalent to making the power output voltage truly stable only at the load end. This change requires running an extra pair of thin lines on the multilayer pcb, but the benefit is that voltage overshoot at the switching instant can be quickly pulled in by the loop, with no inexplicable restarts.

Switching itself is also not that mysterious. I do not pursue microsecond-level handover of control, and for the inertia of an industrial process, a few milliseconds faster or slower simply does not matter. What is really tricky is the open-circuit state during switching: if path A is turned off before path B is fully on, the I/O output drops instantly to an indeterminate level, and the actuator on site may jitter once. So on the DCS PCB I made a make-before-break hardware logic, using a P-channel MOSFET and a capacitor for a simple delay, ensuring that the two control paths are both valid for a short time.

Hot Swap Is a Mechanical and Thermal Test

After doing industrial control for a long time, one deeply felt lesson is that those redundancy architectures and hot-swap designs that look high-end are in the end, when landed physically, nothing more than a reliable Distributed Control System PCB. However much time you spend pondering protocols and controllers, if the PCB itself deforms under repeated insertion, pins oxidize, or the current-carrying capacity of the power layer lets you down, the whole module is armchair strategizing.

I remember a project in which, to achieve true hot swap for the DCS I/O modules, we agonized for a long time over pin definitions and long and short pin coordination. But what finally held up progress was instead not finding a suitable multilayer pcb manufacturer. At that time the design required an 8-layer board with two thick copper layers sandwiched inside, which had to carry the inrush current at module startup. An ordinary multilayer pcb supplier shook its head directly on seeing the thick copper and inner-layer heat dissipation requirement, saying yield could not be guaranteed. Later we turned to ask a factory specializing in Heavy copper PCB, who did take the order, but when the prototype came back, the copper around those high-current hot-swap pins had blistered during reflow. This made us thoroughly understand that when designing a DCS PCB you cannot look only at logic, and you must run in with the manufacturing end in advance. Especially when a module has to be pulled out and plugged back in during operation, the mechanical stress of each insertion on pin solder joints and the pulling on the PCB’s Z-axis expansion are simply something an ordinary consumer electronics board cannot withstand. A good Heavy copper PCB manufacturer will discuss with you copper thickness distribution and the choice of glass transition temperature point and may even suggest additional copper-plated fill at pin through holes, and these details are far more practical than your calculating current limit slope on the datasheet.

Later we simply split the risk control of hot swap into two blocks: the power logic at the system level and the board reliability at the physical layer, each taking half the weight. In looking for suppliers I would rather spend more time validating their multilayer pcb process and go to the workshop to see their parameters when pressing thick copper cores, than be cheap at the selection stage. Because once a DCS module breaks on site, even the lead seal may be torn off by the user, and you do not even get a chance to explain. In the end, behind every pin marked as supporting hot swap, there must be a board that can withstand countless mechanical shocks and current shocks, and this is the hardest bone buried deepest.

HART Over 4-20mA: Isolation and Superimposed Signals

We recently did a project and were stuck by an old problem: how to squeeze HART communication onto a board that already has isolation, without letting analog signal accuracy fall. The topic is rather fine, but I think many hardware peers may have stepped in similar pits.

In the DCS systems I come into contact with, bus isolation on the backplane is in fact not new. Between every slot and the shared bus there is a bus switch or buffer for isolation, and when a module is pulled out of the board, the corresponding bus interface goes directly to high impedance, and backplane communication is completely unaffected. This holds for conventional digital buses, but once you start to consider superimposing HART on 4-20mA, the logic of the whole physical layer has to be sorted out again.

Many people think HART is just simply riding a signal on a current loop, but in fact the water here is deep. The physical layer of HART superimposes an FSK on the 4-20mA DC signal, with logic 1 corresponding to 1200Hz and logic 0 to 2200Hz, with amplitude about plus or minus 0.5mA. The problem is that the output channel of the DCS PCB you make has to do two things at once: precisely output DC current on one hand, and transparently transmit and receive the HART signal superimposed on it on the other, without the two fighting. I have seen quite a few schemes in which problems appear at the place of signal shaping and superposition, and the analog accuracy of 10 microamps simply drifts. This is not something solved by swapping an op-amp or tuning an RC, and the two domains must be completely separated from the architecture.

The structure we use now puts the HART shaping module and the analog output signal conditioning module on a daughter card, leaving isolation and control on the main card. The benefit of this is that they are separated physically, avoiding the trouble of digital ground and analog ground coupling, and when prototyping with a reliable multilayer pcb manufacturer, the control of impedance and interlayer coupling is also much more stable. Especially in making a four-layer stackup, how the power layer and signal layer are arranged directly decides whether the signal integrity of the HART superposition point can be kept. Once we tried integrating the shaping part also into the main card, and the result was that in the 8-channel dense output version, crosstalk between channels was obvious and analog output ripple directly doubled.

The Isolation Barrier and Heavy Copper Output Channels

There is also an easily overlooked point here, the isolation barrier. If you are only running an ordinary 4-20mA output, doing isolation well depends mainly on creepage distance and insulation withstand voltage. But a HART signal is an AC quantity superimposed on DC, and the high-frequency response of the isolation device directly affects the demodulation bit error rate of FSK. We used several digital isolators to replace optocouplers, with no problem in DC characteristics, but group delay inconsistency in the high-frequency range caused an abnormally high HART modem handshake failure rate. Later we were forced to reselect the heavy copper pcb supplier, because the copper thickness of high-current output channels directly affects heat dissipation and impedance uniformity, and if you do not do this end right, after the board heats up, the amplitude consistency of HART all runs off.

Honestly, my requirements for a heavy copper pcb manufacturer are much stricter now than before. Before, 2oz copper carrying two or three amps felt like nothing, but now with 4-20mA output plus HART, you also have to consider the voltage drop at the field-side terminals, and thicker copper means lower impedance but also a greater test of process flatness and lamination stress. When communicating with a multilayer pcb supplier, you have to tell him clearly which layers carry large current and which carry small-signal superposition, otherwise if he goes by the ordinary process, the board comes back and you feel puzzled even measuring the impedance deviation.

Between modules and the bus, I now prefer an isolated backplane bus architecture, letting the physical layer of every channel be completely independent. When a module is pulled out, the bus enters high impedance, which is basic, but what is truly valuable is that isolation is done thoroughly, so that even if a certain channel has a problem when superimposing HART, it will not interfere with other analog outputs in the whole rack. This is especially obvious when making a high-density DCS rack: in a 48-channel rack, if the impedance of the HART superposition point of one channel changes and couples through the power bus, the HART demodulation success rate of other channels can be pulled down by several percentage points.

Many people doing industrial control do not mention the debugging details of the HART interface, but these things are really ground out bit by bit in actual combat. On the board, impedance matching, capacitive load at the superposition point and group delay characteristics of isolation devices, if any one is not handled well, on site they turn into communication that comes and goes or analog output accuracy exceeding the limit.

DP Interfaces, Heavy Copper and the Cost of Cheap Boards

In my first few years in industrial control, I always felt that a circuit board was just that, and that finding a multilayer board manufacturer and throwing over the Gerber would be done. Later, after tumbling on several DCS projects in a row, I realized that a DCS PCB simply cannot be covered by the standard of a general PCB. Especially for boards with DP communication and analog acquisition, once routing density goes up, the process deviation of an ordinary multilayer board supplier begins to be exposed, impedance mismatch of differential pairs makes 12Mbps unstable, and stations drop on site from time to time without you being able to find the cause.

Many people think making hardware for a Distributed Control System is just piling up components, but what is actually most headache-inducing is finding a supplier who truly understands heavy copper. A DCS backplane or power distribution layer often has to pass a dozen-plus amps of current, and an ordinary multilayer board manufacturer tells you that doing 2oz copper counts as heavy copper, and when you ask whether 4oz inner layers can be made, the other side begins to hem and haw and double the quote. The Heavy copper PCB supplier I later fixed on to cooperate with said straight out that for inner-layer etching control above 4oz, few have truly long-term production line experience, and many factories subcontract after taking orders, with yield a mess. The worst case I have seen was an outer layer copper thickness nominally 4oz, but the cross-section showed uneven plating, with some places not even reaching 3oz, and this kind of board, sent to run the thermal cycling test for SIL2 certification, cracked directly.

Now the DP interface. In my early days, to save trouble when drawing boards, I routed the A/B lines of RS-485 with a bit of detour and drilled two more rows of vias, and when the DP master on site pulled a long line, the bit error rate at high speed shot up directly. Later I learned my lesson, honestly wrapping ground beside the DP differential pair and making the termination resistor pads deliberately a bit longer, to make it convenient to change resistor values to match the topology on site. These small moves are not something the Heavy copper PCB manufacturer cares about at all: you have to mark clearly that the area under the common-mode choke needs to be hollowed out, and also urge the multilayer board manufacturer not to touch your stackup structure, otherwise they secretly change the prepreg thickness for you and the characteristic impedance floats.

Now when I look back at DCS PCBs drawn by others, my first look is at power layer copper thickness and interlayer coupling of signal layers, my second look is at the length matching of DP traces, and my third is to find that multilayer board supplier’s process capability statement. Many factories that make fast boards dare to take orders, but those who can truly make stable a board like a DCS that is mixed-signal, with safety isolation, and also has to withstand SIL2/3 life requirements are just a few. Do not ask me how I know: a person who has scrapped several hundred boards, on seeing the quotation sheet of a Heavy copper PCB supplier, first turns to the process capability page to see whether they have really made 4oz buried and blind via boards with more than 6 layers. In this matter, the pits for saving money are all waiting for you in the places you cannot see.

SIL and the Physical Separation of Safety Signals

I remember a period when our team was working on a chemical plant renovation, and the control system had to be upgraded to a DCS with safety interlock. Every day we dealt with all kinds of drawings, and what turned in my head was always the same thing: how could a seemingly ordinary board withstand the diagnostic coverage required by SIL2 certification? It is not enough to write the program more rigorously, and this matter falls entirely on the hardware.

Once, chatting with a long-cooperating multilayer pcb manufacturer, they were skilled at communication backplanes, but when they heard we wanted a Distributed Control System PCB dedicated to safety loops, their engineering chief shook his head directly, saying that the copper thickness on your board has to be changed, trace spacing has to be recalculated, and power layer splitting is much more complex than usual. I fully understood what he said, because next to safety-related signals are ordinary control signals, and how to cut them physically is simply not something solved by drawing a line.

Later we switched to a Heavy copper PCB supplier specializing in industrial power and heavy copper boards, whose stackup advice was very direct: pull independent copper for the power supply of the safety channel, borrowing no return from general layers all the way from the input terminal to the relay driver. In this way a single-point fault does not easily pass through, and the diagnostic path also becomes much clearer. Of course cost goes up a notch, but that feeling that when a problem occurs it will not be muddled is worth more than the money saved.

Many people think SIL levels are posted by certification certificates. The more I see, the more I feel it is a very paranoid habit in PCB design. Like a Digital Output module we later met, to meet diagnostic requirements, each output not only had to measure current but also loop resistance and short-circuit state, and in the end this diagnostic information could not be hidden away but had to go over an independent channel to the safety logic solver. This requires the board layout to physically separate diagnostic traces from safety signal traces from the very beginning while preventing cross interference, and every step tests the board factory’s control of impedance and spacing.

Optical Isolation and a Hybrid Stackup

Speaking of isolation, fiber is indeed a good thing, and later we simply ran all data exchange between the safety controller and the general controller over optical ports. This is not only to satisfy the sentence about electrical isolation in the standard, but because we really met several cases where, when a big pump on site started, communication on copper cable dropped packets, and after a week of checking, it was found that ground plane noise had coupled over. If that was a signal on a safety loop having this kind of problem, I estimate I would not sleep soundly.

The most unforgettable, though, was working with a multilayer pcb supplier on a hybrid laminated board. One side of the board was ordinary FR4 for logic processing, and the other side used high-Tg board material running Heavy copper, specifically to withstand the high temperature and high current of relay switching. For this scheme they at first feared yield would not be good, and after trying batch after batch, they slowly figured out the lamination parameters. Thinking back now, for a DCS PCB, especially a product with a safety loop, design specifications alone are far from enough, and you have to find a board factory willing to tinker with you, who must truly digest which traces on your board are the deadly ones and which can be compromised. Once this tacit understanding is established, afterward, with diagnostics and SIL2 validation, you have much more confidence.

The Underlying Reliability of the Board

I have come into contact with many industrial control projects and slowly realized something overlooked by many: what decides the long-term stability of a DCS system is often not the controller redundancy or software diagnostics displayed on the surface but the circuit boards themselves hidden in the cabinet. Especially for a Distributed Control System PCB, many integrators stare only at parameters like I/O point count and scan cycle and rarely put thought into what the PCB’s lamination structure and copper thickness actually mean. For power distribution boards and analog acquisition boards, once current is slightly larger and heat dissipation design cannot keep up, the first to collapse are the power rails using ordinary PCB, with a voltage offset of a few tenths of a volt starting to make the acquisition of the whole channel drift, which is especially fatal in a continuously running petrochemical site.

So whenever I handle any project involving DCS hardware now, I clarify one thing from the start: I must find a multilayer pcb manufacturer that truly understands industrial electronics reliability, and not just an ordinary supplier who only presses four-layer boards. What you need is a multilayer pcb supplier who can argue clearly with you about underlying details such as the glass transition temperature of FR4, the porosity of through-hole copper filling and copper foil ductility. For example, even for the same six-layer board, the lamination order and the choice of prepreg can make the board’s Z-axis expansion differ by several times under temperature change, which directly affects the long-term life of BGA device solder joints.

Heavy Copper for Surge Survival

Another unavoidable topic is heavy copper. Many DCS relay output boards or power interface boards may carry a continuous operating current of only two amps, but at a short circuit or surge on site, instantaneous current can soar to several dozen amps. At this time a reliable Heavy copper PCB manufacturer will make copper thickness 4oz or even 6oz and use through-hole plating to ensure sidewall connection, rather than simply using thick copper to photoplot a few thick traces. I remember once replacing a batch of I/O boards in an old system: after changing the original standard copper boards to 3oz thick copper boards made by a Heavy copper PCB supplier, board burn-outs caused by field wiring short circuits almost vanished. This is not mysticism, just copper thick enough and the heat dissipation path wide enough that heat was conducted away before it had time to damage the copper layer.

Validation Beyond the 72-Hour Burn-In

Finally, I must talk about validation. Many people think running a 72-hour aging before factory shipment counts as validation, but that is only the most basic link. Real PCB-level validation has to be traced back to the bare board stage with flying probe testing, impedance TDR testing, and board-level thermal stress screening. Some process engineers from wafer fabs I cooperate with even require microsections of multilayer boards to check the copper thickness uniformity of through-hole walls, because a micro crack invisible to the naked eye, in an environment with slight vibration like a power plant, may turn into an intermittent fault half a year later. This kind of board-level, bottom-layer reliability validation is far more substantial than running a few full-load tests during system joint debugging. How should I put it: a reliable PCB in a DCS system is silently the one that bears the load on site, and to make it withstand a ten-year cycle, choosing a supplier by price alone certainly will not do, and you have to treat the board’s process details and validation methods as part of the system function.

Lamination, Hot Swap and Isolation: The Hardware Truth

In my years doing DCS system hardware, the rework I least want to face is a PCB fabrication failure. However beautiful the drawing, if the board factory’s lamination process is not up to it, what arrives simply cannot withstand the torment of that on-site cycle of hot and cold and constant large current. Especially for a board like a Distributed Control System PCB that has to manage safety loops and also run high-speed communication, high-density routing and high-voltage isolation zones squeezed together, the requirement on the multilayer board manufacturer is in fact hidden in every stackup structure. I changed several multilayer pcb suppliers and only at the end found a knack: do not look only at the unit price on the quotation but at how they control the shrinkage ratio when laminating dissimilar dielectrics, which is the lifeline that decides the final dimensional accuracy of your isolation slots.

Hot swap has been simplified by too many people to long and short pins. You can of course choose a connector with precharge contacts and pair it with a hot-swap controller, but the bounce of the ground plane at power-up locks up the logic gates of a board that has not considered copper pour thickness. I met a DCS PCB on which all logic was correct, and as soon as an analog module was inserted and removed live once, the adjacent communication card dropped offline for three seconds. In the end it was not a chip problem but that the power layer used conventional copper thickness, and the surge flowing in from the heavy load end directly raised the ground potential. In the next prototyping, I went directly to a Heavy copper PCB supplier to raise inner-layer power copper to 4 ounces, with the ground plane widened accordingly, and the problem vanished completely. In a real industrial environment, hot swap during normal operation tests the brute-force bearing capacity of copper, not something signal integrity simulation can calculate clearly.

The pits in isolation design are more interesting. Quite a few people feel that placing optocouplers and digital isolators and drawing an isolation band with silkscreen is done. When you send the board for withstand voltage testing and the arcing sound of edge-distance discharge sounds right under the hipot tester, you know that so-called isolation is guaranteed by the PCB’s physical distance and not by the component datasheet. For this I specifically fixed on one multilayer pcb manufacturer, whose lamination process could ensure that the insulating layer produced no micro cracks at slot edges, so that creepage distance could be written into the type test report by measured value. Use the wrong board factory, and even the copper foil on both sides of an isolation device can grow ion migration channels invisible to the naked eye, and after one or two years on site you get intermittent drops in insulation resistance, with fault records full of false alarms that are baffling and untraceable.

Heat Removal on Current Paths

One more easily underestimated point is how to dissipate heat on the current-carrying path. A DCS PCB often has to aggregate system power distribution, and it is not at all unusual for a 24V bus to run several dozen amps. The current-carrying calculation tools for conventional PCBs give ideal uniform-temperature data, but once the board is installed vertically in a card cage, hot air goes upward and the temperature rise gradient is much steeper than you think. At this time finding a Heavy copper PCB manufacturer willing to cooperate on custom copper paste via plugging and thick copper etching saves much worry. They can make your high-current layer 6 ounces or even thicker and, together with local pad HASL thickening, cut IR voltage drop by a large chunk, and with whole-board temperature rise spread evenly, the precision op-amps beside them will not keep drifting.

Honestly, in the post-mortems of several projects, when everyone sat around discussing schemes, few would specifically mention board-level manufacturing of the PCB, as if it were the most inconspicuous link in the procurement process. But in my eyes, whether hot swap or isolation, their final landing point is those few layers of glass fiber and copper foil. However high the parameters are set, choose the wrong multilayer board manufacturer, and it can turn your system with a redundant architecture into a textbook of single-point failure. I think anyone doing control hardware is worth spending some time visiting the board factory in person to see their lamination workshop and flying probe test line, to see how they handle the positive-negative copper distribution of thick copper boards, which is more useful than tuning drawings ten times in the office. Reliable control always starts from the glassy state of the substrate and the bite of copper teeth.

The post Distributed Control System PCB: Redundancy, Hot-Swap and Heavy Copper Reliability in DCS Modules appeared first on SprintpcbGroup.

Adaptive Cruise Control PCB: How Power Integrity, Hybrid Lamination, and Fab Choice Shape 77GHz Radar Performance

sprintpcbgroup — Mon, 09 Nov 2026 07:01:00 +0000

An adaptive cruise control PCB is a system problem, not a materials problem. Across 77GHz radar projects, the biggest losses came from supply noise, hybrid PTFE and FR-4 lamination, dielectric thickness drift, copper foil roughness, HDI via fill quality and thermal design, rather than from the headline Dk and Df numbers on a datasheet. Choosing a high frequency, HDI, high speed or heavy copper manufacturer means checking real 77GHz measurements, cross-section data and process discipline, and asking whether one supplier can combine these capabilities in a single stackup. The sections below collect those lessons from first-hand work, and a short FAQ follows at the end.

Power Is the Annoying Part: Heavy Copper, HDI Fill and Digital Pressure on a Radar Board

People often ask me what the most annoying part of making an adaptive cruise control PCB is, and my answer is never the GHz-level RF matching. It is the power supply. Think about it: a palm-sized board carries three or four radar transceiver channels, and at the rising edge of each chirp every channel pulls current, with a steepness that is no joke. A few years ago I led a 77GHz long-range radar project, and to verify power integrity I deliberately had a heavy copper PCB supplier prototype the board with the inner copper stacked up to 4oz. When I got the board and measured it, the VCO phase noise had improved by nearly 4dB compared with the standard 1oz version. That changed my habits completely. Now, whenever I meet a radar board, the first thing that pops into my head is not which high frequency material to use but whether there is a reliable heavy copper PCB manufacturer who can cooperate on a local thick copper structure.

Do not be fooled by the way millimeter-wave radar is hyped as if it were only a matter of board material and antenna. In reality, the process capability of the HDI PCB manufacturer is the true dividing line. On traces in the 76 to 81GHz band, line width is easily squeezed below 0.1 millimeter, and if the ground via spacing is a bit large and the signal return path detours, the insertion loss mysteriously grows by a few tenths of a dB. Last year we switched to a new HDI PCB supplier who used an improved laser copper-fill process and pushed the void rate at the bottom of the microvias below 2%. With that one small change, in the four-layer RF-to-digital transition area, the entire resonant peak moved to the right by nearly 3GHz. You cannot imagine the joy, like a fish bone that had been stuck in your throat finally being pulled out.

On the choice of frequency band, many people are used to setting 76-77GHz against 77-81GHz, one seeing farther and the other seeing more clearly. What I think is actually ignored is the high-speed pressure on the digital side. Raw radar data throughput keeps growing, and the high-speed differential buses around the processors in the middle easily exceed 10Gbps, forming one BGA fan-out area that gives you claustrophobia. At this point, it is far from enough to look only at whether a high frequency PCB supplier has low-loss materials; you also need to see how experienced it is as a high speed PCB manufacturer. I met one board factory with excellent copper foil roughness control, but its differential-pair length-matching serpentine strategy was extremely mechanical, and the boards it produced had eye diagrams that closed completely at 12.5Gbps. In the end we found the fiber-weave effect in the high-speed area had not been handled. Since then, whenever I choose a supplier, I insist that they provide a joint simulation report for RF and high-speed digital.

Actually there is another topic hidden under the iceberg, which is asymmetric hybrid lamination of materials. When choosing a high frequency PCB manufacturer, many people hold the Rogers or Taconic datasheet and compare Dk and Df line by line, but rarely care how the board will warp in reflow when this Teflon base material is laminated together with ordinary FR-4 in the digital area. Once on the production line, the middle of an entire radar module bulged after the oven, and more than thirty BGA solder joints were cold. Tracing it back, the Z-axis expansion coefficient of the hybrid stackup differed by an order of magnitude. I then specially found a heavy copper PCB supplier who had made a large number of automotive radar hybrid boards to re-prototype, and they adjusted the mirror-symmetric structure in the core arrangement and added stress-relief slots next to the thick copper areas, and only then was the problem contained. Textbooks do not write these things in much detail; they are learned only by taking falls.

Quite a few domestic high frequency PCB suppliers who claim they can do this actually fluctuate widely in mass production level. I dealt with one for two years. Their first batch of 77GHz antenna feed lines had beautiful loss, but in the third batch with the same part number, the resonant frequency was off by nearly 200MHz, and the cause turned out to be that the temperature profile of the lamination process had been quietly changed. Millimeter-wave radar PCBs are just this sensitive, with a process window so narrow that you must get along with your HDI PCB supplier like an engineering partner and not simply as party A and party B. The day they bring even the thickness of the solder mask ink into RF consistency control is the day you can really sleep soundly.

One last thing I want to add: do not easily believe the argument that “materials are king.” I have seen cases that used very ordinary grade materials but achieved extremely low loss through the ingenuity of an HDI PCB manufacturer. I have also seen cases that stacked the most expensive board material, but because the heavy copper PCB manufacturer failed to handle the inner-layer isolated copper islands, the power resonance point fell exactly inside the chirp bandwidth, and the false alarm rate of the radar inexplicably soared. The difficulty of the adaptive cruise control PCB is that it is a system problem.

Finding a Fab That Can Do Both High Frequency and HDI

The biggest pit I have stepped in with 77 GHz radar boards was not how to turn the traces but how to find a reliable factory. Last year I was tuning an adaptive cruise control PCB that used a highly integrated single-chip scheme. The board area was saved, but a four-layer board simply could not run, and HDI became necessary. That was when I realized how hard it is to find an HDI PCB manufacturer that can handle both high frequency and HDI processes. Most factories are either good only at digital HDI like mobile phone mainboards or only know how to make pure high frequency boards, and asking them to combine the two capabilities is enough to scare them off.

Later I simply changed my approach. Instead of staring at the small circle of HDI PCB suppliers, I went looking for high frequency PCB factories with a genuine RF background. There was a high Frequency PCB manufacturer making communication base stations, which had never carried an automotive radar label before, but its control of 77 GHz microstrip loss was far more reliable than many manufacturers who billed themselves as “dedicated to automotive radar.” When talking to them, I asked only two things: could they make a substrate stackup with Df stable within 0.002, and were they willing to take small-batch prototyping. The boss was down-to-earth. He told me they mainly used Panasonic M6 or Rogers 4350B hybrid lamination, and would change my power layer to thick copper foil, equivalent to building a small portion of heavy copper PCB structure into it. We had never thought of this idea before: it turns out there is no need to find a Heavy copper PCB supplier to make a separate power board, because large current and RF signals can both be served within the same stackup.

This matter made me re-understand so-called “solution selection.” Many people agonize over whether to follow the NXP or the TI route for the chip, but on the PCB side, what really decides the life or death of a project is whether the high speed PCB manufacturer you choose has measured your material selection in the GHz band. Later, at the factory of that high frequency PCB supplier, we used their own probe station to measure the S-parameters of the microstrip lines, and the measured loss was a little lower than the simulation, and only then did I feel at ease. An adaptive cruise PCB is not like a digital board; it cannot degrade into a connector part that merely “brings the signal out.” At 77 GHz every centimeter of trace is a filter, and leaving out one prepreg sheet in the stackup may drop antenna gain by more than 1dB. So now when I talk with peers, I suggest they move part of their energy away from the solution selection diagrams and first go to the factory to see a real dielectric loss test once. That is more useful than reading ten datasheets.

Splitting the Radar Board: Why Antenna, Power and Digital Sections Need Different Fabs

Engineers who work on millimeter-wave radar have probably all experienced it: when choosing board material, theoretical derivation and the board you actually receive are often separated by the “processing limit of the HDI supplier.” We once made a 77GHz Adaptive Cruise Control PCB. At the start everyone was idealistic and thought that if we replaced all RF layers with a certain high frequency material, the simulation would show good VSWR and a beautiful pattern. When the board came back, there was delamination, warpage, and even the phase consistency of the antenna array had drifted. Only later did I understand that the temperature coefficient of dielectric constant and the Z-axis expansion rate of high frequency board materials, in multilayer hybrid lamination, are simply not covered by the few numbers on the datasheet.

During that period I asked almost every well-known high frequency PCB manufacturer and supplier on the market, and found that many factories that claim they can do high frequency hybrid lamination reveal their true colors in microstrip loss at the 77GHz band. Once, the salesperson of a heavy copper PCB supplier insisted that their thick copper process could also make RF layers, on the grounds that thick copper for the power layer dissipates heat well. When the prototype came back, the burrs on the copper foil on both sides of the antenna feed line had destroyed the impedance continuity. That made me realize that the power layer and the RF layer need two completely different manufacturing logics, and one manufacturer cannot serve both with the same production line.

Later we adjusted our strategy and split the radar PCB into two parts to manage. For the part that carries 77GHz RF, such as signals and antennas, we specifically looked for an HDI PCB manufacturer good at high-density interconnect to be responsible for fine lines and laser blind vias. For the power and digital processing parts, we looked for a high speed PCB manufacturer and a thick copper manufacturer, treating them as pure digital boards and power boards. There is a very counterintuitive point here: not all HDI PCB suppliers understand high frequency. Some factories have no problem controlling line width to two or three mil, but when asked how the surface plating finish at 77GHz affects phase, the other side falls silent. The one we finally worked with was truly willing to measure the dielectric constant temperature coefficient with us again and again and adjust lamination parameters, and only then was antenna gain stabilized.

After this round, my biggest takeaway is that the resolution of a radar board depends not only on bandwidth and not only on algorithms. If the copper foil roughness of the PCB or the voids at the lamination interface are a few microns more, then however good the range resolution is, the signal-to-noise ratio collapses all the same. Now when anyone asks me whether to choose a high frequency PCB supplier or a heavy copper PCB manufacturer, I say: first look at what your antenna layer and digital layer actually require, solve the Dk matching and copper thickness transition at the hybrid lamination interface, and only then talk about how many GHz of specification. After all, in the mass production stage, stability is more solid than anything.

Dk, Df and Copper Roughness: What 76 to 81GHz Actually Demands

A project some time ago completely changed my view on millimeter-wave radar board material selection. In the past I always felt that since the chip could cover the entire 76 to 81 GHz band, the PCB design could simply follow the chip, and at worst we could tune the matching at the end. When measured, it turned out to be nothing like that. On the adaptive cruise control PCB we used, the antenna section performed acceptably near 76 GHz, but above 79 GHz the gain dropped so far that I did not dare to look, and the sidelobes also popped out. Later, when I talked with an engineer from a high frequency PCB supplier, they pointed out directly that the Df of the batch of base material we used varied too much near 80 GHz, and that once the feed line was slightly longer, the loss could no longer be described linearly.

This incident made me realize that you should not expect to find an HDI PCB manufacturer, make the board any-layer interconnect, and thereby solve the various troubles at high frequency. Routing density and via stubs do improve, but in essence the performance of a millimeter-wave antenna depends first on the base material and the roughness of the copper foil. Later we switched to a manufacturer focused on high frequency PCBs, who recommended reverse-treated low-roughness copper and performed impedance flatness testing between 76 and 81 GHz. Although the cost went up, the improvement in antenna efficiency was visible to the naked eye.

Another underestimated area is heat dissipation and power supply. Many people think the power consumption of long-range radar is just that much and not worth worrying about. In fact, in a high-temperature automotive environment, the temperature rise in the power amplifier area seriously affects phase noise. Later I switched the power layer to a thick copper PCB scheme, and found a heavy copper PCB supplier to add copper thickness under the power amplifier to 3 ounces. You heard it right: mixing a high frequency RF board with a heavy copper process does bring alignment and etching challenges to the PCB factory, but the heavy copper PCB manufacturer I worked with was experienced and used a step-by-step lamination approach to solve the thermal expansion coefficient mismatch, and in the end the thermal uniformity of the whole board improved considerably.

On the choice of band, the 76 GHz point actually feels like a “sweet spot” in long-range ACC. It has less crowded interference, and a slightly lower frequency means impedance is easier to control at the same line width. Later we optimized the antenna matching network specifically for 76 to 77 GHz, and the result was nearly 2 dB better than the original wideband design. Of course, the premise is that your radar only does forward long range and does not need to cover high-resolution scenes such as blind spot monitoring and parking. Otherwise you still have to honestly spread the bandwidth out to 81 GHz.

So, back to selection: when I set the PCB scheme with my team now, I do not look only at the line width and spacing capability of the HDI PCB supplier. I first confirm the Df and copper foil type of the base material. If power is high, I introduce the thick copper process at the same time, and then find a high frequency PCB manufacturer capable of hybrid lamination to do the integration. Even for the same supplier, the output of its high speed PCB production line and its RF PCB production line may be at completely different levels at 77 GHz. These pits cannot be appreciated without running mass production several times.

Single-Chip Radar Versus Chipset: What Integration Does to the Board

After several years of working on adaptive cruise control radar boards (Adaptive Cruise Control PCB), I increasingly feel that the so-called “single-chip one-stop service” is a lazy choice, and a dangerously lazy one. Vendors boast that a single chip covers the transceiver channels and digital processing, so there is less RF routing and power is easier, which sounds especially worry-free. But anyone who has drawn a few versions of a 77GHz board knows that it is exactly this highly integrated kind of chip that bottles all the contradictions inside the package, and once channel isolation is insufficient or local oscillator leakage appears, you have no room for remedy. I handled a project in which a single chip was forced to carry four transmit and four receive channels, and calibrating channel phase consistency took nearly two months of struggle, and in the end we tore it up and started over, moved the RF front end out, and switched to a less “compact” chipset scheme. The board area grew considerably, but the RF traces could finally be length-matched and ground-guarded comfortably, with isolation worked out one by one. To keep the loss of the antenna array and front-end microstrip down, I asked almost every fab around that could do high frequency hybrid lamination, and finally settled on a high frequency pcb manufacturer that could meet my dielectric constant tolerance requirement. The base material absolutely could not be ordinary FR-4; it had to be a dedicated low-loss material.

This board had another pit, which was power. Radar transient current pulses are quite fierce, especially when multiple channels capture at the same time, and the inner copper of ordinary 1 ounce thickness simply cannot cope, with voltage drop and hot spots arriving together. Later I simply found a Heavy copper PCB supplier and took the power layer straight to 3 ounces, using copper-fill for the laser blind vias and redesigning the whole stackup. Once the heat dissipation path was open, the temperature drift of those millimeter-wave chips in the RF front end became much easier to control. Many people shake their heads at the mention of heavy copper, saying the cost is too high and processing too slow, but for a forward radar you are not doing high-volume consumer electronics, and board reliability matters far more than saving a few dozen yuan.

I am used to verifying the digital part separately, so I found a high speed PCB manufacturer I have long worked with to make a high-speed digital daughter card, getting the radar MCU, external memory and interfaces running. After all, the digital domain easily has GHz-level DDR or MIPI traces, which follow a completely different logic from impedance control in the RF domain, and if the two are not split cleanly on one board, the price is a blown noise floor. After the electrical performance on both sides was stable, I found an HDI PCB manufacturer able to do both HDI and any-layer interconnect to laminate the RF board and the digital board together. But this kind of hybrid lamination process is especially sensitive to expansion and contraction control, so when communicating with the HDI PCB supplier early on, you have to squeeze out details like registration tolerance and through-hole back-drill depth like toothpaste, and you cannot get away with saying “follow the conventional process.”

Single chip is not to say it cannot work. For a forward long-range narrow-beam ACC with only two transmit and four receive channels on the board, where space truly cannot fit more, going for a single-chip SoC is also reasonable. But once you aim for imaging radar, with a dozen or so channels, the few transmit ports of a single chip simply cannot hold the waveform. At that point you have to take your eyes off the lazy “high integration” thinking and honestly weigh it: do you want to save board-level area, or do you want a cleaner RF noise floor and channel bandwidth? Once you have thought this through, you will feel confident when you go to haggle over stackup and copper thickness with your high frequency pcb supplier and Heavy copper PCB manufacturer.

Heat, RF Materials and Digital Interfaces on a Compact Radar Module

I recently handled an adaptive cruise radar module where board area was tightly constrained and power consumption was not low, and for a while I spent all my energy looking for suitable PCB manufacturers and relatively little time on chip selection. Many people think that once the chip is decided, drawing the board is a connect-the-dots exercise, but at a band like 77GHz, even the choice of board material can sink the whole project.

Chip solutions now like to lean toward single chip, with RF transceiver, phase-locked loop and ADC all rolled into one part. Board-level RF routing is reduced, but the small package underneath gets as hot as a soldering iron. When we began evaluating, we did not take heat dissipation seriously, and a thermal simulation showed the junction temperature easily surging to 130 or 140 degrees. The thermal pad had only that little area, and without heavy copper conducting heat to the bottom layer, there was no way to hold it down. Later I specifically went to a Heavy copper PCB manufacturer, had them use 2 ounce copper on the inner layers and add dense thermal vias, and only then was the temperature barely held within an acceptable range. This taught me one thing: an adaptive cruise control PCB is not an ordinary digital board; its thermal design and RF performance are almost one body, and if either end is off, trouble follows.

The RF part is even more demanding. Our board area was small, with only that little distance between antenna and transceiver, so that a slight deviation in a microstrip line scattered the pattern beyond viewing. The loss of ordinary FR4 in this band is simply unacceptable, so high frequency material is a must, such as Rogers 4350B or the domestic Wangling board material. The problem is that not every factory labeled high Frequency PCB supplier truly understands this stuff. Some factories have good machines, but their engineers have no concept of 77GHz impedance control, and after the boards are made and measured, the VSWR drifts outrageously. Later we settled on a high frequency factory that specializes in automotive radar boards, and they can even quantify the effect of copper foil roughness for you, and only then was consistency stabilized.

The digital interface was no easier. The high-speed signals from the transceiver have to be pulled to the domain controller via CSI or LVDS, with rates easily reaching Gbps. If the trace length is slightly out of control, the eye diagram fights itself. On this kind of board the RF area and the digital area sit too close, and if isolation is not done well, harmonics from the digital part drill straight into the receive chain. My habit now is to directly look for a high frequency HDI board factory capable of hybrid lamination, using Rogers for the RF layers and medium-loss FR4 for the digital layers, pressed together with prepreg in between. This kind of process cannot be taken on by just any HDI PCB manufacturer; they need laser drilling capability and must also understand the precision of interlayer registration, otherwise the stub effect of RF vias will drop your signal into a pit. For the HDI PCB supplier I found, just confirming the stackup with their engineers took nearly half a month of back and forth.

Thinking back, what is most feared in making this kind of radar module is not any single technical difficulty but the shortage of reliable manufacturers in the whole industry who can string these things together. Some factories make high-speed digital boards well but show timidity on RF, and some RF factories cannot handle heavy copper thermal dissipation. If you force yourself to assemble several suppliers, the result is that the boundaries of responsibility cannot be sorted out. When I choose suppliers now, whether they call themselves a high speed PCB manufacturer or a high frequency board factory, I first check whether they have mass-produced automotive radar cases, and second whether the production line can complete HDI, hybrid lamination and heavy copper in one process. This kind of board is truly not something to hand to whoever is cheapest, because it hangs directly on safety functions, and a little fluctuation cannot be solved by rework.

One feeling runs especially deep: the higher the single-chip integration, the greater the pressure on the PCB. In the era of discrete components, the RF chain could be rescued by tuning the matching. Once integrated, you can hardly change anything, and all the margin must be covered by the precision of the PCB. Making an adaptive cruise control PCB is more like a small systems engineering project: board material, process, heat dissipation and signal integrity constrain one another, and no single link can be optimized in isolation. Each time I meet a new project now, I wear out the manufacturer’s process capability handbook in advance, run a few rounds of feasibility discussion with their process head, and only then come back to touch the schematic, so as not to be trapped in a prototype-failure-prototype death loop afterward.

Selection Is More Than Chip Parameters: Hybrid Lamination, Heavy Copper and a Changing Domestic Supply Base

After these years in automotive millimeter-wave radar, I increasingly feel the word “selection” has been over-simplified. Many people think selection is looking at chip parameters, comparing the number of transmit channels, receive channels and ADC bits, then picking the best value. But in actual product work, especially on a forward adaptive cruise radar, what truly chokes you are things that seemed unremarkable at first: how to choose the PCB board material, which HDI PCB manufacturer can supply steadily, and whether high frequency and high speed traces can coexist on the same board without killing each other.

In one project of ours, the RF front end used an SoC that packed the MCU and DSP together, in order to save board space and reduce high-speed signal routing. This kind of chip is outrageously demanding on the PCB. The RF part needs low loss and a stable dielectric constant, while the digital part has DDR and high-speed serial interfaces, with all the power consumption crowded under a package the size of a fingernail. The first sample from our high frequency PCB manufacturer had decent impedance consistency on the antenna side, but ground bounce and power noise in the digital part had not been handled at all, and the result was a pile of glitches mixed into the intermediate frequency signal and a false alarm rate in target detection that went through the roof. This incident made me thoroughly understand one principle: the higher the chip integration, the higher the demand on the overall capability of the PCB manufacturer, and it is nowhere near as simple as marking a stackup structure on a drawing and sending out the Gerber.

Afterwards we began to re-examine our supplier list internally. Factories that could only make standard multilayer boards, or that claimed to do high frequency but had only run a few RF boards and then called themselves a high frequency PCB supplier, were passed on directly. What we needed was a factory that could knead HDI and heavy copper into the same board with controlled yield. Because for heat dissipation, radar boards often need thick copper on the inner layer directly under the chip, while to fan out the antenna feed lines and other digital lines, they need HDI with two or even three laser drilling passes. Many HDI pcb suppliers shake their heads immediately when you ask for local copper thickness above 3 ounces plus blind and buried vias, saying the process risk is too great. In the end we found a factory with both heavy copper PCB manufacturer and hdi pcb manufacturer capabilities, and when we communicated, their engineers directly pointed out several points prone to trouble, for example that side etching in thick copper areas would cause dimensional deviation of fine-pitch BGA pads, and that after laser drilling the adhesion of the hole-wall copper on the thick copper layer could be an issue. That level of exchange is ten times more reassuring than a supplier who can only reply “we can do it.”

On our new platform we are now gradually separating the RF front end and the MCU again. It is not because an integrated SoC is bad, but because we want to take back the flexibility for RF noise control and digital performance upgrades. With two chips, the RF part is powered by an independent low-noise LDO, and the digital part is handed to an automotive-grade MCU, with a lightweight digital interface between them. This split really did raise board complexity, and high-speed signals again had to cross the split, forcing us to redesign the PCB stackup. But the benefit is that the heat of the RF front end is isolated, so heavy copper only needs to serve that small zone, while the digital part can be done with conventional HDI, and it is also easier to swap the MCU model to meet the computing power needs of different vehicle models. In the end, the evolution of a radar module is not decided by a single chip; it is a pushing-hands contest among the chip, PCB manufacturing and vehicle packaging. Whether the high speed PCB manufacturer you choose can understand the impedance matrix you need, and can give you stable batches in board warpage and pad flatness, directly affects whether the antenna pattern you calibrate can be reproduced.

Sometimes in chats within the industry you hear a view that domestic PCB factories still fall a bit short on RF boards. I admit that was true a few years ago, but the situation has really changed now. Some factories, after supplying automotive customers for several years, have a very deep understanding of IPC-6012 class 3 and thermal management, and they even proactively suggest which underfill material to use at the bottom of the RF chip to balance the CTE difference. To me, a good HDI pcb supplier does not just take orders and make boards; it looks at thermal simulation and stress distribution with you, and even before prototyping reminds you how much reflection a certain trace corner will introduce. Once this kind of relationship is established, you do not want to switch easily.

In the end, whether it is the Adaptive Cruise Control PCB used for ACC or the small boards of corner radar, the core of choosing a manufacturer is not how many certifications they hang on their website, but whether their engineering team can argue with you about the stackup scheme and can frankly tell you that their laser drilling precision gives only a 95 percent yield at this BGA pitch.

Chipset Architecture: Radar Transceiver, MCU and Power Management on One Small Board

Some time ago I took on a project to redraw the board for an automotive adaptive cruise control system, the Adaptive Cruise Control PCB as it is commonly called. Frankly, at first I thought it was just a packaging and process upgrade of the original scheme, but once I actually started, I found it was far from something a switch of high-speed board material could settle. Many things are hard to judge intuitively unless you have tuned a board with a soldering iron and quarreled with a board factory over impedance.

The most tormenting part of the whole system was not the radar transceiver chip itself but the chipset architecture binding it together with the MCU and power management. This chipset is not the plug-and-go kind you imagine: the three chips have completely different requirements for supply ripple and ground return paths. The RF part needs low-noise analog power, the digital core of the MCU is extremely sensitive to transient response, and if the several outputs of the power management chip are not handled well, ground bounce goes straight into the intermediate frequency signal and stirs the chirp waveform until it cannot be looked at.

At first I tried to be lazy and had a high frequency PCB manufacturer who claimed to make mixed RF and digital boards prototype directly. A batch of boards came back with interlayer registration tolerance so large that the microstrip impedance had run off by 8 ohms, and with a 77GHz signal on it the VSWR was pitiful. Later I switched to a small factory that specialized in millimeter-wave boards. Strictly speaking they were not a high frequency PCB supplier, and they were instead more willing to grind through the stackup with me again and again. This made me realize that in a millimeter-wave scenario, whether you look for an HDI PCB manufacturer or an HDI PCB supplier is not about the label but about whether they are willing to keep trial-and-error company with you. Many large factories accept only standard stackups, and if you insist on a hybrid lamination of Rogers RO3003 with FR-4, they refuse the order outright.

The factory I later used is not large, but they have been a high speed PCB manufacturer for a long time and have a particular set of skills in handling LVDS length-matched serpentine lines. The high-speed ADC interface between the RF front end and the MCU runs over MIPI, which requires strict control of intra-pair skew without introducing extra jitter. They directly put two layers of embedded capacitance material into my stackup, pressing down power integrity so I did not need to pile a bunch of decoupling capacitors on the surface layer. This kind of process is hard to discuss with a traditional Heavy copper PCB supplier, because the line width and spacing of heavy copper boards and the fine routing required for millimeter wave are two different logics. Unless your board has both a motor drive channel of a hundred or two hundred amps and a 77GHz antenna feed line, you need a cross-over player who understands both a high frequency PCB supplier and a Heavy copper PCB manufacturer, but such factories are as rare as phoenix feathers.

There is also an unavoidable topic, the mindset of RF debugging. We tend to trust big factories, but the debug port of the MCU in my hands brings out a 400k periodic pulse every time it enters simulation mode, directly interfering with the power rail whose decoupling for the RF front end was insufficient. I had not anticipated this problem at all during layout, and it was suppressed later by adding a ferrite bead with a jumper wire. If, like me, you later want to use a more compact hdi pcb to shrink the size, this kind of pulse coupling will only get worse, because once line width and spacing get smaller, crosstalk control can no longer be solved by small vias. Many hdi pcb suppliers recommend any-layer interconnect, saying the routing density is high, but the continuity of the reference plane of the RF signal is easily broken under this process, and every time you drill a microvia the return path detours once and the insertion loss rises steeply.

My understanding now is that in making an Adaptive Cruise Control PCB, you must not be fooled by the “complete matching” of a chipset scheme. Putting that three-chip combination on one small board actually forces you to face the physical contradictions at board level: RF wants lightweight, low Dk drift; power wants large current and low impedance; digital wants low jitter and length matching. With all three pressed into the same stackup, if the compromise is not handled well, what runs may be only a hot brick. A more practical line of thought is that, rather than agonizing over which high frequency PCB manufacturer has the most certifications, look at their statistical capability in handling layer shift and whether their dielectric thickness tolerance is narrow enough. Because at 77GHz, even a Dk deviation of 0.02 produces a phase shift error that worsens your angular resolution, and none of this is written directly on a datasheet.

Df Is Not the Whole Story: Warpage, Low-Flow Prepreg and the Real Loss of a Routed Trace

Designing a radar board for automotive adaptive cruise, after a long time you understand one thing: the numbers in the material parameter table and the board that actually gets made and run are separated by a very long distance. When we selected materials for the Adaptive Cruise Control PCB, which works at 77 GHz, everyone knew insertion loss had to come down, with Df wished toward zero. But when you go talk with board factories, many high frequency PCB manufacturers frown as soon as they hear multilayer hybrid lamination or partial hybrid lamination.

It was only later that I came to my senses: chasing Df from 0.002 to 0.001 might, in the link budget calculation, save two or three tenths in efficiency, but once the board warps and the antenna feed line impedance shifts, that loss advantage is entirely eaten. So now I tell the team: do not stare at that ceramic-filled PTFE first, and first settle things with the HDI PCB supplier. Why? An ACC radar module has little space, layers are laminated one on another, and the job of laser blind via copper fill is something few HDI PCB manufacturers can do nicely on a hybrid structure. Especially since we also tried heavy copper for the power section, finding a heavy copper PCB supplier to make a thick copper core board and then laminating it with the high frequency layers, the warpage control was simply deadly.

If the rhythm of baking, lamination and cold press is even slightly wrong, the board bends like a potato chip after depaneling. Later we readjusted the stackup and borrowed the low-flow prepreg from the process of a high speed PCB manufacturer. We sacrificed a tiny bit of dielectric uniformity in the high frequency layers in exchange for overall flatness, and the measured 77 GHz insertion loss turned out more stable than with the original pure high frequency board. Strange, isn’t it?

So I no longer very much believe the claim that “a certain base material is the sole benchmark.” You measure 0.15 dB/cm on RO3003, but in actual routing, with two bends and two layer-change vias, the loss is not in the same order of magnitude at all. On the heavy copper PCB manufacturer side, which carries power, once copper thickness goes up, etching accuracy gets rough, and the Dk change brought by line width fluctuation is far more direct than the effect of Df at several places after the decimal point. In the end, in the GHz band, the details all lie in the manufacturing window, not in the few datasheets of the material.

Rethinking the Stackup: Pure Dielectric Paths, Copper as a Stress Balancer and What HDI Really Means for ACC

Not long ago a friend working on automotive radar came to ask me. They had just taken on an ACC radar project and had been stuck on the stackup scheme alone for half a month. They thought they could simply throw the RF layers to the antenna and make do with FR-4 for the digital part, but in simulation the return loss simply could not be suppressed and the 77GHz section jittered beyond viewing. After hearing him out I knew this was not a question of which board material to choose at all; he had never treated the “stackup” as a systems engineering problem from the start.

Discussion of the Adaptive Cruise Control PCB nowadays is easily led astray to the Dk value of high frequency board materials. Many people start by comparing how much the Dk of RO3003 differs from FR-4, as if putting the numbers into a table comparison would produce a good board. But after dealing with several high Frequency PCB manufacturers over the years, I find that what is truly deadly is not the static value of Dk but the “behavior” of Dk in a hybrid structure. For example, when you make an RF trace cross a material boundary once, the impedance reflection from the abrupt change in Dk is far more violent than you imagine, especially in the millimeter-wave band. I have never seen a datasheet that tells you how to digest the problem of such a transition zone with the stackup. It cannot be solved by one parameter; you have to think of ways starting from the continuity of the trace reference layer.

My own habit is to do it in reverse. Instead of patching the stackup, I first decide which layers must be “pure,” and then push the hybrid lamination interface as far away from sensitive signals as possible. I once helped someone revise a case in which they had sandwiched the high frequency layer in the middle for symmetry, hoping to suppress warpage, and the result was that RF vias met two dielectrics of different Dk when crossing layers, and the crosstalk alarmed immediately in simulation. Later I forced the antenna layer and the transceiver into the same dielectric so the RF path never had to cross the FR-4 boundary, even at the cost of losing a bit of stackup symmetry. What about warpage? I went to the heavy copper PCB supplier and had them use the thick copper layer as a natural stress-balancing plate. With 3 ounce copper laid on the ground layer and the prepreg adjusted, the measured bow of the finished board was about 0.6%, even better than their original symmetric structure. This actually breaks the inertia that “hybrid lamination must be symmetric.” As long as you are clear about the copper thickness and the direction of the glass cloth, warpage does not leave you without other paths.

Many people in the industry keep HDI on their lips and think that high density necessarily means finding a factory that bills itself as an hdi pcb manufacturer. But a board like the ACC radar needs a “density” that is not about how high the blind via stacking layers go; it is whether power distribution and high-speed signal isolation can be had at the same time. When I worked with an HDI PCB supplier, they always liked to push any-layer interconnect, and I said: however good your laser holes are, if the ground layer current carrying provided by my heavy copper PCB manufacturer cannot keep up, then once power noise above 11GHz arrives, whatever low-Dk material you use is for nothing. So now when I choose suppliers, I do not look at who advertises how complete they are; I look at whether they can fuse the impedance control experience of a high speed PCB manufacturer with the thick copper process of a heavy copper PCB supplier. This is actually quite hard, because high frequency houses are often unfamiliar with thick copper, and thick copper houses complain that high frequency materials are hard to laminate. But an ACC radar forces you to pinch these two ends together.

When it comes to material selection, I feel many people are limited by the word FR. As soon as they hear FR-4 they think cheap stuff, and insist on piling on high frequency material. In fact, for the medium and low frequency digital layers and power layers, the stability of FR-4 is more reliable than many so-called medium-loss materials, and its bonding strength with copper is a great advantage when aligning heavy copper. The advice I later gave that friend was to reverse his coarse classification: instead of cutting it into “high frequency layers versus low speed layers,” rebuild the stackup by “continuity of signal return and current carrying capacity.” Then when you look at the Adaptive Cruise Control PCB, it is no longer a matter of how many layers of RO and how many layers of FR, but of how the whole board lets RF current, digital noise and power fluctuation each take their own road, while Dk matching is just a result that appears naturally in the end.

Now whenever I meet a similar case, my first sentence is: do not draw the stackup diagram first. First figure out on which physical path your high-speed signals really have to run, and then decide which manufacturer to realize it. Because even if you design the stackup perfectly, if the high frequency PCB supplier you cooperate with has no handle on the lamination shrinkage rate of hybrid lamination, the actual Dk of the board that comes out deviates from theory and the antenna pattern is still deformed.

Layer Count Is a Compromise: Heavy Copper, 4D Imaging Radar and Practical Selection Logic

Many people think that for an ACC radar board, the more stackup layers the better, as if separating RF, digital and power would settle everything. After running several projects, I found it is far from that obvious. Behind the choice of layer count, what you should think through first is how much transient current your SoC actually consumes, how fierce the isolation between RF channels needs to be, and whether the HDI PCB manufacturer you can find is willing to fiddle with some non-mainstream stackups together with you.

I dealt with a Heavy copper PCB supplier in Shenzhen for two years, and their thick copper process was quite seasoned. Once I needed a four-layer board to carry a peak current close to 15A, and traditional 1oz copper simply could not hold; with a slightly longer trace the voltage drop shot up to an unacceptable level. Later I directly used 3oz thick copper on the inner layers and 2oz on the outer layers, and the impedance loop of the power was made particularly small, so that even with only two power planes the ripple was pressed lower than on some six-layer boards. The EIRP of that Adaptive Cruise Control PCB in the 76-77 GHz band, as finally measured, did not drop out because of the low layer count; on the contrary, because the ground copper was thick enough, the reference continuity of the RF loop was quite good. This made me realize that what you lack is not layers but bold copper weight planning and a cooperative Heavy copper PCB manufacturer.

Of course, not every scenario can be carried by thick copper. Later, when I built a prototype of another 4D imaging radar, the RF channels doubled to 8 transmit and 8 receive, and it also used a high-end SoC with a PCIe high-speed bus, and then the physical limit of a four-layer board was right there. The high-speed signals of the digital part needed a complete return path, adjacent ground layer coupling alone was no longer enough, and the crosstalk almost made me doubt my life. During that time I screened nearly every well-known domestic high Frequency PCB manufacturer and high speed PCB manufacturer, and finally chose a factory that could do low-profile copper foil and supported hybrid material lamination. With a ten-layer board, the 77-79 GHz RF layers, digital layers and power layers were completely stripped apart, with a solid reference plane behind every GHz-class sensitive trace. Without that foundation, however accurately you tune the SoC frequency, board-level loss will gnaw away the resolution.

Later I slowly formed a very practical judgment logic: first look at how many independent channels your RF architecture really needs, then look at integration. If a single-chip SoC locks digital and RF together and the surrounding power supply is not complicated, I often have the HDI PCB supplier directly do four to six layers of one-step laser blind and buried vias, combined with local thick copper, and the result is often cleaner than expected. If the system itself must be split into an RF front end plus an external MCU chipset scheme, then honestly raise the layer count, find a good high Frequency PCB supplier to order dedicated hydrocarbon substrates, and at the same time make sure high-speed signals get a complete shielding ground. Layer count is never an absolute value; it is a compromise that you work out together with the manufacturer. Once you have identified the real pain point, the physical form of the board will emerge by itself.

Surface Roughness, Hybrid Structures, Heat and Functional Safety at 77GHz

In my first two years of working on millimeter-wave radar boards, I always thought that as long as the base material Df was chosen right, the rest could be left to the board factory. Only when the project reached the 77GHz band of adaptive cruise did I find how thin that idea was. Once I sent a stackup to a factory that claimed to make high frequency boards, and the other side replied directly with “this copper foil cannot be made,” and only then did I realize that the Adaptive Cruise Control PCB is not the logic of copying FR-4 experience at all. After that I spent a lot of time re-understanding one thing: the nominal Df at 10GHz can only serve as an admission ticket in a 77GHz environment, and what is really deadly is copper foil roughness and processing tolerance. When dealing with a high frequency PCB supplier, I would rather have them first provide the measured insertion loss curve of a microstrip line than comfort myself with a figure like 0.001 on a datasheet, because as you go above the GHz level, the conductor loss caused by surface roughness eats several dB, and many materials with beautiful nominal Df have never been verified on plated circuits at this band.

Another thing that made me take a big detour was the hybrid structure. I used to think “the high frequency layer takes care of the antenna and the rest can just be stacked with ordinary FR-4,” and then a sample cracked straight to the inner layers after thermal cycling. Later I talked with a friend at a high frequency PCB manufacturer, and he told me directly that the thermal expansion coefficients of my two material types were too far apart, the stress in the x-y direction and the z direction had simply not been released, and hybrid lamination is not simply pressing different board materials together; you must tune the whole stackup as a symmetric whole. This also made me start to understand why some hdi pcb manufacturers, when the frequency span exceeds two bands, insist on low-profile copper and high-Tg bonding sheets, even if the cost is higher. When choosing an hdi pcb supplier I also do not just ask how many orders of HDI they can do; I directly ask whether they have done 77GHz multilayer hybrid cases and whether they have compensation data for layer shift control. These details cannot be settled by drawing an impedance line and filling in a line width.

The power consumption and heat dissipation part was the pit I stepped in latest but with the biggest lesson. The thermal simulation looked ample, but the actual sample dropped frequency when run at high duty cycle, and teardown later found the heavy copper distribution on the inner layers was insufficient, with heat blocked under the power amplifier network. I started to look for a dedicated heavy copper PCB manufacturer to cooperate, raising the copper in the power amplifier area from the conventional 1oz to 3oz and drilling dense thermal vias between layers, and only then was the junction temperature pressed down. Factories that are heavy copper PCB suppliers are usually very familiar with current capability but insensitive to the parasitic effects of high frequency signals, so I required them to handle both thick copper and the high-speed layers in the same lamination, which circles back to the old problem of hybrid lamination. It was also because of this that I realized the high-speed digital link part cannot be left to an ordinary board factory to do casually. I would rather spend a little more money to find a high speed PCB manufacturer that truly understands differential routing length-matching control and solve LVDS and RF routing together as two pain points of the same system, rather than each managing its own.

As for the functional safety requirement, it is actually invisible to the board factory. When an ASIL level lands at board level, it does not end with changing to a high-reliability ink; it affects the redundancy design of your entire power delivery network and the failure mode analysis of via chains. In projects at ASIL C and above, I confirm with the board factory from the start the elongation of the copper foil, the uniformity of hole-copper thickness and the thermal stress test profile, because these are directly tied to potential open-circuit failures. Many times you throw the reference design from the chip vendor straight to board fabrication, and overlook that the same design in the hands of different hdi pcb manufacturers can differ greatly in residual copper rate and dielectric thickness tolerance, and the resulting frequency offset and pattern distortion you can only take the blame for yourself. So, circling back, a 77GHz design is not about which numbered board material you chose but about whether you can find a supplier willing to test together with you and share the processing risk. Otherwise even the most beautiful simulation is only numbers on a screen.

Chip Choice Is Only the Beginning: HDI Tolerance, Single-Chip Trade-Offs and Heavy Copper Power

After so many years in millimeter-wave radar, my deepest feeling is that choosing the chip is only the start, and almost all the pits afterward are on the board. Especially for the 77GHz radar heads used in adaptive cruise control, a chip worth a hundred or two hundred yuan paired with a third-rate substrate produces something that cannot even hold a target steady. Conversely, once you put a bit more thought into the PCB, the performance curve converges at once.

On one project we took the easy route and found a nearby HDI PCB supplier, thinking the laser blind vias and fine lines they advertised could meet the accuracy of the antenna feed lines. When the first batch of boards came back, the antenna VSWR was outrageously scattered, and with the same chip soldered on, detection range could differ by 30%. Only after we tore them down and measured did we find that the dielectric thickness deviation of the high-density interconnect layers exceeded the simulation tolerance, and in a band like 76 to 81GHz, a tiny dimensional shift throws impedance matching into complete disarray. Later we switched to a manufacturer that had long made high frequency PCBs, and with their improved hydrocarbon base material and strict thickness control process, the gain of boards in the same batch was remarkably uniform. Since then I have understood that whether a so-called HDI PCB manufacturer can do high frequency work depends entirely on their understanding of GHz signal behavior, and is not settled by high machine accuracy.

Many system people become infatuated with the simplicity of single-chip schemes. One package contains the RF front end, ADC and DSP, and board-level routing really is reduced by a large chunk, looking neat. This is a plus when going through ADAS functional safety certification, but there is a hidden cost: the entire digital processing path is locked in. We once evaluated a single-chip radar SoC with an integrated C66x DSP. It was quick to get started with, but when we wanted to embed our own trained neural-network perception model, we found the scheduling mechanism of its built-in hardware accelerator fought with our algorithm structure, and compiler optimization could not go all the way either. In the end we returned to the chipset scheme, using a pure RF transceiver plus an independent processor. Board-level complexity doubled, but algorithm freedom was completely released, and several later firmware iterations were no longer hampered by hardware. The extra high-speed digital traces are really not that scary either: find a reliable high speed PCB manufacturer, let them control the skew and reflections of the differential pairs, and they can easily hold a length tolerance of under 10 mil.

The power part is also an easily ignored corner. The phase-locked loop of a 77GHz transceiver is extremely sensitive to supply noise, and the current peaks of the transmit channels are very steep. I began with a board of conventional copper thickness, and however well the digital and analog ground splits were done, ground bounce remained under large current pulses. Later I cooperated once with a heavy copper PCB manufacturer, raised the inner copper to 4oz and combined it with embedded capacitance material, and the noise floor dropped immediately. You can hardly imagine that heavy copper PCB suppliers can actually make high frequency hybrid structures too, only that thermal symmetry must be planned in advance, otherwise the warpage after soldering will cause cold BGA joints. Our approach was to place the heavy copper layer in the middle, with high frequency copper-clad laminates of equal thickness above and below to form a sandwich structure, and after reflow the flatness was still within the IPC class 2 standard.

Someone always asks me whether to choose a high frequency PCB manufacturer or a high speed PCB manufacturer, and the boundary has blurred in the millimeter-wave band. A microstrip line carrying RF and a differential line carrying high-speed serial data are essentially both wrestling with skin depth and dielectric loss. A supplier who can stably provide a substrate with a dielectric constant of 2.94 and a loss tangent of 0.0017, and who has experience controlling copper foil surface roughness, is often more important than staring at the chip brand. The adaptive cruise control PCB must ultimately be able to steadily track the tail lights of a car 150 meters away, a small target that appears at a 4GHz sweep bandwidth as a range resolution cell of less than 4 centimeters, and any mismatch on the board-level link is enough to drown it in clutter.

Lessons Beyond the Datasheet: Nominal Dk, Mixed Lamination Costs and What to Measure at 77GHz

After doing this for a long time, I gradually became less superstitious about selection guides. Textbooks always tell you that for the 77GHz Adaptive Cruise Control PCB, the base material must be a model with Dk strict to two decimal places, as if a 0.04 difference in loss could take you to the sky. But anyone who has actually pulled boards and measured them knows that the nominal Dk given by board suppliers is simply not accurate at 77GHz, and even different positions on the same roll of material show drift. What you should really contend over is not the pretty tanδ on the datasheet but whether your high Frequency PCB supplier can stably control the glass cloth distribution and the tooth profile of the rolled copper. An hdi pcb manufacturer who honestly makes ROGERS and TLY series without mixing in filler is ten times better than you flipping through every theoretical formula.

Quite a few people mention hybrid stackups as if pressing a high frequency sheet together with FR-4 were the golden key to cost control. I feel the opposite: for a front radar that needs long-range detection, I would rather sacrifice some BOM cost to ensure dielectric consistency across the whole board. I made symmetric sandwich structures several times, and warpage and impedance problems were not solved any more cleverly; after reflow they instead exposed Dk jumps in isolated copper areas and resin accumulation zones. Later I simply settled with a factory that is a high Frequency PCB manufacturer, using Rogers 4350B throughout with low-profile copper and a four-layer board routed honestly, and the measured 77GHz insertion loss turned out about 0.3dB lower than those fancy hybrid schemes.

Do not underestimate that 0.3dB. In a system link, it is a difference of a few percentage points in detection range. Especially for some new platforms that use single-chip SoCs for adaptive cruise, the board-level space is pitifully small, and you have to fit multiple power rails and also handle ASIL-level isolation, so relying on an hdi pcb supplier for microvia stacking and any-layer interconnect becomes the norm. On such boards, the hole-wall roughness and fill fullness of laser blind vias directly raise return loss at GHz frequencies. Many high speed PCB manufacturers can do prototypes but lack real test experience in the millimeter-wave band, sweeping a vector network analyzer only up to 20GHz before handing you the report, so the 77GHz performance is pure guesswork. My approach is to work deeply only with factories that can provide measured 77GHz TDR and S-parameter curves, and before signing, to take out a previous Adaptive Cruise Control PCB sample board and run chip-to-chip crosstalk on it, and only when the data looks good do I commit to automotive-grade mass production.

Heavy copper PCB is another easily ignored pit. The transmit end of a long-range radar has considerable power demand, and with the vehicle supply margin that must be reserved, the power section often needs thick copper. Some Heavy copper PCB suppliers are only good at power substrates, and once they are hybrid-laminated with a high frequency area and the annealing process is changed, the grain structure of the copper changes, and the surface roughness effect above 1GHz is magnified. I tried moving the power module separately onto a board from a Heavy copper PCB manufacturer, and mating it with the RF core board through a flexible board, which instead bypassed the thermal failure problem that sends Dk crazy. The cost is an additional connector, but the phase noise at 77GHz and the detection consistency of the whole unit clearly steadied, and in road tests there were far fewer false braking and false alarm events.

Many people ask how to choose the base material and structure. My view is just one sentence: do not push down from the chip’s peak specifications; push up from how much interference, crosstalk and temperature drift your sensor can actually tolerate once installed on the vehicle. By the time an Adaptive Cruise Control PCB is done, all the mysticism lands on the microstructure of the copper, the glass fiber wetting and the uniformity of pattern plating. If the high speed PCB manufacturer you cooperate with does not even watch the cpk of the etch factor on the production line, then even the most expensive board material is wasted. These dirty, fine production details are worth far more than the nominal Dk value in the laboratory.

Frequently Asked Questions

Why is power supply noise such a serious problem on an adaptive cruise control PCB?

A radar board carries several transceiver channels that pull steep current at the rising edge of every chirp, and VCO and PLL circuits at 77GHz are very sensitive to supply ripple. In one project, moving the inner layers to 4oz copper from a heavy copper PCB supplier improved VCO phase noise by nearly 4dB compared with a 1oz version. Splitting the power domains, using low-noise LDOs and adding embedded capacitance or local thick copper are often more effective than tuning RF matching.

Is a low Df board material enough to guarantee good 77GHz performance?

No. At 77GHz, copper foil roughness, dielectric thickness uniformity, hybrid lamination warpage and via quality can erase a small Df advantage. In one case, a simpler low-profile copper stackup in Rogers 4350B showed about 0.3dB less insertion loss than a more elaborate hybrid scheme. Ask the supplier for measured 77GHz insertion loss, TDR and S-parameter curves instead of relying on the 10GHz values on a datasheet.

How should I choose between a high frequency, HDI, high speed and heavy copper PCB manufacturer for a radar board?

Look for the combination of capabilities your design needs inside a single stackup. A radar board often needs high frequency material for the antenna, HDI for dense fan-out, high-speed digital control for interfaces such as LVDS or MIPI, and local heavy copper for power and heat. Ask for automotive radar mass production cases, interlayer registration data, cross-sections of laser via fill and hybrid lamination warpage results. Some designs are easier to manage by splitting the RF and power or digital sections across different boards or fabs.

When should I use a single-chip radar SoC and when should I use a chipset?

A single-chip SoC suits a compact forward long-range radar with few channels and limited space. For imaging radar with many channels, or when you need your own algorithm freedom and cleaner RF noise, a chipset with a separate RF transceiver and processor is often better, even though board complexity rises. Higher integration puts more pressure on the PCB, because heat, isolation and signal integrity margins must be covered by the board itself.

The post Adaptive Cruise Control PCB: How Power Integrity, Hybrid Lamination, and Fab Choice Shape 77GHz Radar Performance appeared first on SprintpcbGroup.

Home Health Monitor PCB: Patient Isolation, Creepage and Mixed-Signal Layout for ECG, SpO2 and Wireless Boards

sprintpcbgroup — Mon, 09 Nov 2026 07:00:00 +0000

Why a Hospital Design Cannot Simply Be Shrunk for the Home

After a long time in home health monitor design, I have noticed a fairly common phenomenon: many people think that if they shrink the hospital monitor solution and swap in cheaper components, they can move it straight into the home. This line of thinking is especially dangerous for PCB design. A hospital has a dedicated grounding system and a controlled electromagnetic environment, but at home? The Wi-Fi signal is full, the neighbor is heating breakfast in the microwave, and the electric bike downstairs is charging. All this messy interference is hitting your board. So when I choose a Home Health Monitor PCB solution, I never look at reference designs trimmed straight down from hospital equipment. That is purely looking for trouble.

A Cheap Four-Layer Board and a Ruined ECG

Once we prototyped a home multi-parameter monitor project. In the beginning we felt a four-layer board was enough and found a multilayer pcb manufacturer with a very low quote. The first revision came back with the ECG signal in a complete mess, the respiration wave basically drowned in 50Hz mains interference, and the SpO2 channel jumping badly from time to time. At first we thought the front-end circuit had not been tuned well and fiddled for more than half a month. Later, a check with a near-field probe revealed the culprit: on the board, the digital ground return paths ran all over the place, and the analog ground reference plane was split into pieces. That manufacturer simply did not understand the layout care needed when weak signals coexist. They were used to making consumer electronics boards and had no concept at all of isolation creepage distance and leakage current paths.

Isolation Barriers as Rigid Physical Requirements

This forced me to be extra picky when looking for suppliers later. In a home environment, you must design the monitor by CF-type or BF-type applied part, and the physical size of the isolation barrier is a rigid requirement; it cannot be passed by casually leaving 1 millimeter of spacing. A multilayer pcb supplier that has truly made medical boards will ask right away about your applied part classification, working voltage and transient overvoltage category, rather than caring only about line width and spacing. I now lean toward factories with HDI process capability. For an analog front end in a BGA package with 0.3mm or 0.4mm pitch, it simply cannot be fanned out without HDI routing, and you should not expect the traditional through-hole process to handle it. But not all hdi pcb manufacturers can take medical orders. If their lamination process and laser drilling precision are not good enough, microstrip impedance deviation becomes large and the suppression of high-frequency noise collapses entirely.

RF and Microvolt ECG on the Same Board

This is especially so for home monitor boards with a wireless transmission module, where RF signals of several hundred megahertz and microvolt-level ECG signals exist on the board at the same time. This can no longer be handled by simply pouring copper; you have to run precise electromagnetic field simulation on the stackup structure before it can be fixed. When I choose an hdi pcb supplier, I repeatedly confirm with their engineering using the actual stackup file, and the batch fluctuation of the dielectric constant must be controlled within a very small range, otherwise the matching network of the RF part is made for nothing. Some factories say they can make high-frequency boards, but when you ask them to control the etching accuracy of low-loss materials they give themselves away. I still specifically sought out several high frequency PCB manufacturers focused on medical and RF. When dealing with the 2.4GHz or 5GHz band, their process experience controlled residual copper ratio and copper foil roughness clearly better, and after the boards came back and were measured, the influence of Bluetooth and Wi-Fi spurious interference on the analog front end could be pressed down by several dB.

Spending Where It Counts

Of course I am not saying to blindly pile on high-end processes or blindly look for expensive high frequency pcb suppliers. Home products still have to consider cost, but it has to be spent at the right points. My habit is to spend the money on the core signal acquisition area, such as the right-leg drive circuit loop and the large ECG analog front-end area, requiring strict impedance control and a complete reference plane, and perhaps locally using a first-order stepped laser blind via, while the power and digital processing parts just use a conventional FR-4 multilayer process. This approach can both satisfy the requirements of YY 9706.111-2021 on electrical safety in the home care environment and keep the single-board cost from getting out of control.

The Danger of Consumer-Medical Miniaturization

A trend in these years has given me quite a headache, the so-called “consumerized medical”: a fancy little board that integrates a lot of functions densely, with a beautiful appearance, but when you open it and look at the PCB, the required isolation spacing is insufficient, filter circuits are omitted wherever possible, and in the end when testing, a static hit resets it and a surge burns the port. In the home scenario, the priority of PCB reliability is far higher than miniaturization. Your user may casually put the monitor beside a humidifier, and may use a power adapter without medical certification. These situations all have to be taken in when designing a Home Health Monitor PCB; you cannot expect users to operate as cautiously as nurses.

Visit the Medical Line Before Choosing

So every time a newly entered hardware engineer asks me which board material to choose and how many layers are enough, I say do not stare at the parameter table first. First go to the medical production line of that multilayer pcb manufacturer or hdi pcb supplier in person, see whether they have a dedicated clean workshop for medical boards, and ask how well they understand the 60601 standard.

Isolation, Creepage and Dense Wireless Boards

When Finding a Multilayer Supplier Feels Like Opening Blind Boxes

A while ago a friend who makes home monitors came over to chat. As soon as he sat down he sighed, saying that finding a multilayer PCB supplier now was more and more like opening blind boxes. For the prototyping of one ECG acquisition board he had run through three multilayer PCB manufacturers, and each time the boards came back the isolation voltage test was unsatisfactory. Listening to him, I remembered that the little PCB inside these palm-sized home health monitoring products actually carries quite a few hidden pits.

Why Isolation Is Where Small Monitors Fail

Many people think a home monitor is small, the routing is simple, and an ordinary multilayer board maker will do. But I have seen too many rollovers, and the problem lies in isolation. The large monitors in hospitals have solid ground wires, and the interference from mains power is carried by metal chassis, but for home products it is completely different. There is no earth wire to connect to, and absolute safety must be ensured between the patient and the USB charging cable, so those few millimeters of creepage distance on the board become the life-and-death line. I once had a multilayer PCB supplier run a 4000V withstand voltage test on their conventional FR-4, and after a few boards were hit, there were traces of surface discharge. Later I was forced to look for multilayer PCB manufacturers who truly understood medical isolation. They knew how much the copper foil on each side of an isolation slot should be pulled back on which layer, and whether the solder mask and silkscreen ink would form a conductive path under high voltage. These small matters are things a factory that has never made medical boards simply does not bother about.

Three Jobs in a Matchbox

Another headache is board density. You also have to squeeze in Bluetooth and WiFi modules, receiving extremely weak ECG signals on one side while setting up an antenna to send data out on the other, with LED pulse drive for the blood oxygen probe in between. I say this is “three jobs in a matchbox.” This is when the advantages of an HDI PCB supplier show up: laser drilling can tuck traces into the inner layers, leaving the outer layers with a complete ground plane for high-frequency shielding. One high-frequency PCB manufacturer I dealt with had specifically tested this kind of small wireless module, and controlling impedance and reducing crosstalk was indeed crisp, but before prototyping I would repeatedly confirm the distance from vias to the isolation boundary. Some HDI PCB manufacturers are good at phone boards, their stackup designs fierce, but the moment medical isolation standards are mentioned they show timidity. So I learned my lesson, and every time I change an HDI PCB supplier, the first thing is not to talk price but to have them bring out the stackup cases they have made for medical devices.

Separating High-Frequency and Isolation Suppliers

In the end, for the small PCB of a home monitor, my habit now is certainly to look at high frequency and isolation separately. For the RF part, I find a supplier good at high-frequency PCB to do antenna matching, and for the analog acquisition and isolation part I find a multilayer PCB manufacturer who understands medical safety regulations. Of course, if you can find one factory that does both well, it saves worry, but such teams are hard to meet. I changed four or five in total, and my current stable supplier has been ground for several years on medical equipment boards, picking at creepage distance compulsively, and even marking the position of solder mask bridges for me. The biggest feeling from cooperating with them is that PCBs in this industry really cannot be “usable is enough”; you have to think safety through in advance in every stackup and every millimeter of gap. After all, a home monitor, unlike hospital equipment, has no regular safety inspection. In the user’s hands it may not be turned off for years, and if the board has any flaw in isolation, that is a major matter.

CF and BF Applied Parts and the Supplier Process Window

Splitting Medical Isolation and High-Frequency Interconnect

In the last few years I have handled several home health monitoring device projects, from blood oxygen rings to ECG patches, and my deepest feeling is that board factory selection is more headache-inducing than the schematic. Many people, when they first start making this kind of product, always think of finding a factory that can do multilayer, HDI and high frequency all in one go. But in practice you will find that board factories that can meet medical-grade isolation requirements and board factories that can do high-frequency low-loss interconnect are often not the same group of people.

Process Tolerance Eats Your Safety Margin

When looking for a multilayer board supplier, one pit I stepped in was understanding creepage distance only in theory. For example, seeing in the specification that 2 x MOPP requires 8mm, you feel that leaving an 8mm slot in layout is enough, ignoring that the board factory’s own process capability will eat your safety margin. One multi-floor board manufacturer’s prototype came back with copper edge burrs that stunned me, and the actual electrical clearance had shrunk by nearly 0.2mm, which is simply unacceptable for a CF-type applied part. CF and BF differ not just in leakage current limits; they decide the stackup budget of the whole board and the supplier’s process window. With BF type you can breathe slightly easier, but once ECG acquisition is included, the 10 microamp limit of CF makes every microamp path scrutinized, including the CTI value of FR4, the withstand voltage of the solder mask, and even the thickness uniformity of the prepreg between layers.

Separate Suppliers for High-Frequency and Monitoring Boards

Later I completely separated the suppliers of the high-frequency board and the home monitoring board. For the RF part, such as the Bluetooth and WiFi antenna matching section, I settled on one manufacturer that specifically makes high-frequency PCB, who can control Dk to three decimal places and make differential impedance quite stable. For the Home Health Monitor PCB that carries ECG and PPG signal acquisition underneath, I find another HDI supplier good at medical multilayer boards. There is a very practical consideration here: high-density interconnect board factories control line width accuracy better in laser drilling and via fill plating, because the CF-type isolation channel often needs narrow slots to widen creepage distance, and if the relative position of trace and hole is even slightly off, safety certification cannot pass.

Screening HDI Suppliers for IEC 60601 Experience

Some HDI board factories are not very willing to take small-batch medical board orders, finding the process complex and the responsibility large. At this time you have to spend more effort screening HDI PCB suppliers, for example, to see whether they have IEC 60601 project experience and whether they will proactively suggest adjusting the board thickness to optimize the withstand voltage of the isolation barrier. I remember a project where we designed an isolation band between the ECG front end and the Bluetooth module, and the board factory engineer directly pointed out that the core thickness I selected would degrade insulation resistance after humid-heat preconditioning, and only after switching to a high-CTI material they commonly used was the withstand voltage made stable.

Joint Leakage Accounting Across Suppliers

Another easily neglected point: when you hand the multilayer main board and the wireless module to different factories separately, the contribution of BF or CF type to total leakage current must be accounted for jointly. Even for a daughter board connected through pin headers, as long as it ultimately shares a patient-connected reference ground, all isolation barriers must be designed uniformly at the system level. I once encountered a case where the high-frequency board supplier inadvertently bridged the antenna ground layer to the main board digital ground, and the applied part that was originally BF type was directly upgraded to CF-type requirements, and leakage current exceeded the limit in whole-machine testing. After that, at the prototyping stage we required each supplier to file the stackup, copper thickness and surface finish, and to review them together with the width of the isolation slot reserved in the Gerber.

The PCB as the Physical Carrier of Compliance

In the end, the PCB for home health monitoring is not just about connecting components; it is the physical carrier of the compliance threshold. Whether you look for a multilayer board manufacturer or an HDI supplier, only those who can discuss with you the real impact of CF/BF on board drawing, and who are willing to re-arrange holes for the sake of 8mm of creepage distance, are truly on the right track. Price and delivery time are of course important, but compromising on safety isolation costs far more in turning back later than the difference of a few bucks in unit price.

The Patient Island and Defibrillation Protection

A Rejection That Pulled Me Out of the Textbooks

I remember the first time I communicated my home monitor board requirements to a multilayer pcb manufacturer, the other party directly replied: we generally do not make this kind of ECG board with defibrillation protection, the creepage distance cannot be pulled open, and if something happens we cannot bear the responsibility. These words sounded harsh, but they really pulled me out of the pile of books. Later I figured out that many people understand that barrier across the middle of the board too rigidly, as if you must draw a gulf visible to the naked eye to count as responsible for patient safety. In real production, if you really use a caliper to measure that 8 millimeters of blank band, you often find the signal traces already squeezed until they can barely breathe, the waveform of high-speed pulses either overshoots or sags at the edges, the ECG waveform that is acquired is mixed with a layer of glitches, and you want to throw the oscilloscope.

Compressing Patient Contacts into a Compact Island

My habit now is that when I get the schematic I do not rush to divide domains. First I compress all contacts that touch the human body, the lead wires, the electrode snaps and the right-leg drive feedback point, into a compact “patient island” in my head. Then I look for a factory that is willing to make hdi pcb manufacturer boards and can also handle high-frequency boards, and such factories are actually few in China; most either specialize in multilayer through-hole or only dare to touch the two laser-hole layers under FR4. The board in my hands is less than a palm in overall size, yet has to fit Bluetooth, power isolation and the protection network that can withstand defibrillation pulses. Once the traditional discrete layout of discharge tubes and current-limiting resistors is used, the patient island immediately expands, and the whole board loses its lightness. So I buried part of the absorption loop in the middle of the multilayer board, using a PTFE hybrid sheet material, and relying on the lamination accuracy of the high Frequency PCB supplier, hid the pulse discharge path on the second and third layers, leaving only a few small-package fast recovery diodes on the surface. This saves a large amount of planar space and also avoids tangling over creepage distance on that virtual barrier.

Learning from Supplier Selection Detours

I also took quite a few detours in supplier selection. Early on, in a rush for a prototype, I casually found a supplier that called itself able to make high-density multilayer boards, and as a result when the boards came back there were broken vias and inner-layer cracks everywhere, not to mention the spikes on the impedance lines. Later I switched to a factory that carried both the titles hdi pcb supplier and high Frequency PCB manufacturer. They adjusted the stackup order completely, embedded the power island and the signal shield layer on both sides of the core, and made the ground return around the patient island form a three-dimensional spherical shell, and only then was the ECG input noise pressed from the millivolt level to a baseline noise floor invisible to the naked eye. At this point, looking again at those cold spacing numbers in the medical standard, they feel more like a safety net than a path you must kneel to walk.

A Stepped Protection Chain Tuned by Surge Experiments

On defibrillation, many people think a home environment will not meet a defibrillator, but who can guarantee that an elderly person never has ECG electrodes stuck on the chest while sitting on an electric blanket? I have seen an extreme case where a device was accidentally touched by an electric mosquito swatter, several thousand volts of pulse coupled in instantly, and the channel front end was burned black. Later I made the input stage protection stepped: first coarse blocking through a gas discharge tube, then slowing the rising edge with ferrite beads and high-frequency stray capacitance buried within the multilayer board, and finally a pair of back-to-back TVS at the chip pins, with the total chain length under six millimeters. This part of the design did not apply the standard “keep-out band” width at all; it was tuned hard by repeated surge experiments with a high-voltage probe. The stackup of that test board came from a multilayer pcb supplier willing to modify the stack with me four times; they did not use an off-the-shelf package but recalculated core thickness and prepreg resin content based on my actual breakdown threshold.

The Barrier as a System, Not a Gap

I feel that the most deadly thing in making a Home Health Monitor PCB is not that you do not understand the regulations, but that you treat the regulations as a rigid drawing. The true meaning of a barrier is not a copper-foil break visible to the naked eye, but a whole set of systemic cooperation across physics, electricity and heat. You must understand the microvolt-level delicacy of the ECG signal, and also the volatile temper of a defibrillation pulse arriving at several thousand volts, and then, on a board of a dozen or so layers, make these two worlds live in peace. Many times, rather than painfully counting every bit of spacing, it is better to put more thought into the centripetal nature of component placement and the compactness of loops. When the board is small and the traces are short, parasitic inductance comes down, and the local magnetic field produced by pulse current can actually be suppressed more cleanly. This is much smarter than mechanically cutting a pale isolation band on the board, and it also saves the budget of adding layers later to make up for signal trace losses.

ECG Front End, Right-Leg Drive and HDI

Hardware People Know the Real Bottleneck

Hardware people all know that designing a usable home health monitoring board has never been as simple as buying a pile of top-grade chips and piecing them together. In the projects I have touched over the years, the real chokepoint is often not those few ADCs or op amps but that you cannot find a reliable multilayer board supplier to land it. Especially when the board involves ECG signal acquisition and you also have to squeeze the analog front end into a palm-sized shell, things get interesting.

Right-Leg Drive Is a Board-Level Problem

A while ago I was tinkering with a home multi-parameter monitor board, and the ECG part alone was enough to drink a pot of trouble. Many people start by staring at the AFE datasheet, tuning filters and calculating gain, but ignore the invisible pits on the board itself. Take the right-leg drive: this thing is not simply connecting an op amp in inverted output. It is tied to the reference ground and shielding strategy of the whole board. If you are using just an ordinary multilayer board manufacturer, with poor interlayer dielectric uniformity, the common-mode rejection collapses right in front of you. After all, the feedback loop of the right-leg drive passes through the human body, electrodes and lead wires before returning to the board end, and any impedance imbalance brings mains interference in. Later I switched to a factory that could do HDI, shrank outer-layer traces to 3 mil, and set aside a whole inner layer of ground for the analog part, and only then did I stabilize CMRR above 100dB.

Why HDI Is Not Wasted on Medical Electronics

Speaking of HDI, many people feel it is something only phone motherboards use and a bit of a waste for medical electronics. I do not see it that way. Home health devices now have ever higher stacking density and still have to fit 4G or Bluetooth modules, and if the high-frequency part is handled even slightly badly it crosstalks into the weak signal chain. At this time it is especially important to find a manufacturer that has both HDI process capability and high-frequency board manufacturing. I remember one experiment with a board with Bluetooth using an ordinary FR4 eight-layer board, and as a result, once Bluetooth power came up, the ECG waveform had regular small glitches, definitely not power supply noise. Later a near-field probe sweep showed the RF signal coupled directly into the AFE input traces. After switching to HDI, traces were buried in the inner layers with complete copper outside, and the world was quiet. Also, a good HDI supplier will generally help you optimize the stackup and isolate RF ground from analog ground, avoiding the nonsense of so-called “common ground impedance.”

How Factory Engineering Capability Affects Long-Term ECG Stability

There is one more point easily underestimated: the influence of the board factory’s engineering capability on long-term ECG stability. I often tell peers that when looking for a multilayer board supplier you must ask clearly how far their impedance control can go, because the input impedance of the ECG front end is extremely high, usually at the GΩ level, and a change of even a few tens of pF in stray capacitance on a trace can make all your effort fail. A multilayer board manufacturer that truly knows the trade will tell you the batch fluctuation range of their board material Dk value and how to compensate after high-temperature and high-humidity aging. This kind of detail is something low-end factories cannot be bothered to explain. I have a batch of boards in hand whose baseline drift grew larger and larger after two years of use. After switching to a supplier with medical product experience, with the same stackup design, the drift amount fell directly by an order of magnitude.

Involve the HDI Supplier Early

Finally, let me say something from the heart: engineers who make home health monitoring boards should stop believing that one AFE settles everything. Spend a part of your energy on board-level implementation, especially if you plan to go to volume: from involving the HDI PCB manufacturer early in the design to watching the process window of the multilayer board supplier at all times in production, these trivial chores are the key to whether the product can survive. However beautifully the circuit is simulated, if the board factory gives you dielectric thickness off by a few microns and the right-leg drive feedback signal is half a beat slow, the heart rate variability you measure is a joke. This industry has no magic, only the patience to quarrel with every parasitic parameter.

Dry Electrodes, Board Leakage and Wireless Density

Why the Small Board Decides the Waveform

In the last two years I have tinkered with several versions of small home ECG monitoring gadgets, from the initial circuit built on a breadboard to now actually sending boards to prototyping and mass production, and the pits I have stepped in are enough to bury a person. Many people say that the most important thing in building health monitoring devices is the sensor and the algorithm, but later I understood that the palm-sized circuit board is where the real skill shows. The design of the Home Health Monitor PCB basically decides whether the waveform you finally read out is a clean, crisp ECG signal or an occasional heartbeat popping up among a pile of glitches.

High-Impedance Electrodes and Ionic Contamination

When I first started looking for a board factory, I never imagined there would be so many troubles later. I casually drew a four-layer board, sent it to a quick-turn shop, and when the boards came back and I soldered on the components and tested, the noise floor was so large that it could not be used at all. After a long investigation, the problem turned out to be in the stretch of trace where the electrode enters. Home electrodes are different from the hospital kind with conductive paste; many are dry electrodes or capacitively coupled patches, with scary contact impedance, megohms and above not being unusual. At this time, if the PCB trace is even a little longer, or the board’s surface insulation is not good, the signal has already leaked away outside before it even enters the AFE chip. Later I shortened the input traces to the extreme and had the board house strictly control ionic contamination, and after several reworks it barely passed. In that phase I began to understand that an ordinary multilayer prototyping factory and a multilayer board supplier that can truly make medical device grade circuit boards are not the same thing at all.

Moving to HDI and Isolating the Right-Leg Drive Output

For later versions, because Bluetooth and storage were added, board density jumped up. At first I foolishly still used the traditional through-hole process, and as a result the routing could not be fanned out at all and I had to turn to HDI. Only then did I come into contact with a real HDI circuit board manufacturer, using laser blind vias and buried vias to build up the stack, and the routing went smoothly. But problems came with it. The interlayer dielectric of HDI is thin, and the Bluetooth RF crosstalk from the digital part started to run toward the analog front end, especially the output trace of the right-leg drive, which is particularly sensitive. The power supply of the drive circuit here must be clean and completely separated from the main system power, otherwise digital noise will be poured back into the human body along the ground loop. I put it on a small isolation island of its own, with an independent isolated power module, and series a sufficient current-limiting resistor at the output to ensure that even if the chip breaks down internally it would not push large current into the person. The trace path also has to avoid all high-speed signals and cannot run parallel to the Bluetooth antenna, and for that stretch I even directly dug out part of the ground copper to avoid forming an antenna effect.

Dielectric Tolerance for Antenna Efficiency and Battery Life

The RF part also gave me a headache for quite a while. Bluetooth antenna impedance matching needs extremely strict dielectric thickness and copper foil accuracy, and at this time, if the multilayer board supplier does not control the dielectric constant well, the signal strength of that batch of boards comes out entirely by luck. Later I specified that a high-frequency PCB manufacturer make the stackup for the RF side specifically, requiring a dielectric constant tolerance of plus or minus 0.04, and also confirming copper foil roughness back with the factory. You may think a home ECG patch does not need such fussiness, but when you find that a slight loss in antenna efficiency forces the device to raise power in transmit state, the battery fails to last two days, and users will certainly leave a pile of bad reviews. The materials provided by a high-frequency PCB supplier can be selected by myself, for example using a low-loss hydrocarbon ceramic substrate. Although it is damn expensive, the connection stability it brings is very much worth it in daily use.

Balancing Scales: Precision Analog and RF Care in One Space

Thinking back now, doing a Home Health Monitor PCB really tests your ability to balance at different scales. On one hand you must serve the few input lines of the ECG front end with the attitude of building a precision analog board, with every electrode trace coming in being like a high-impedance probe, where the slightest carelessness couples interference. On the other hand you must control the Bluetooth part with the meticulousness of making a phone RF board, and you must squeeze both into the same physical space with only a few layers of thin prepreg between them. At this moment you can feel the difference in the level of HDI circuit board suppliers: a good manufacturer will give you process advice, telling you where a blind via will not become a weak point for later moisture ingress, while a poor factory only processes according to the drawing, and the board warpage that comes out is invisible to you, but tiny warpage causes cold solder joints when soldering the BGA-packaged AFE. The integrated front end I used, the Texas Instruments AFE4960, is a DSBGA package with a ball pitch of only 0.4mm, and if the PCB flatness is slightly worse it cannot be used. Moreover, such small-package devices generate concentrated heat, so the thermal design of the whole board cannot look only at the power consumption number but must simulate the effect of body temperature coupling in the actual wearing environment. The consideration of these details is far more important than choosing a “low-power chip.”

The electrode connector is also an easy place to roll over.

An AFE Pinout Dense Enough to Need a Magnifier

Not long ago, while building a portable health monitoring gadget for myself, the most headache-inducing thing on the circuit board was that fingernail-sized AFE chip. The pins were so dense that I needed a magnifier to see them. At first I wanted to save trouble and found a conventional multilayer board supplier for prototyping, but they directly refused the order, saying that fanning out traces under a 2.6mm package needed laser blind vias at least, and conventional mechanical drilling simply could not fit. That was when I realized the routing density of this Home Health Monitor PCB was already at the HDI level. I re-found an hdi pcb manufacturer specializing in high-density interconnect. Their suggestion was to use a 3-order any-layer, with line width pressed down to 3mil and one more trace threaded between pads, so that only then were the power supply and analog front end reference under the AFE barely sorted out. The split of power and ground was also tricky. If the analog domain and the digital domain are mixed, the ECG acquisition baseline drifts to the sky, so the stackup had to be pulled to more than 6 layers, and this factory could also make high-frequency boards, so the antenna section of the Bluetooth module hanging on the SPI port was handed to their internal high-frequency impedance control, saving me from sending it out again to another high Frequency PCB supplier. In fact, finding a multilayer pcb manufacturer that can handle both HDI and hybrid high-frequency lamination is quite critical, otherwise if the board is made by two factories, impedance matching data gets passed back and forth with finger-pointing, and the waveform rings in a mess during joint debugging. Also, the several pairs of differential traces around the AFE I hand-matched in length for a whole afternoon, winding serpentine until my eyes blurred, but there was no choice, because the loop of the right-leg drive, if handled carelessly, couples back into the input and turns into a kind of low-frequency hum. Looking back now, even that tiny 222 microamp quiescent consumption forced me to calculate one by one the battery life of the coin cell under high duty-cycle sampling, and only then did I find that choosing a more professional hdi pcb supplier is not just about solving processing capability: the copper thickness carrying current on the board and the thermal resistance of vias quietly affect the supply ripple of the AFE in pulse mode. The whole process forced me to push PCB design from front-end selection all the way to back-end manufacturing, like delivering this little board through its whole birth.

LED Pulse Noise and Why the Factory’s Noise Understanding Matters

A while ago, when choosing a PCB factory for a home blood oxygen monitoring project, I stepped into quite a few pits, and looking back some things really should have been sorted out earlier. At first I always stared at the parameters of the board itself, such as layer count and line width and spacing, and found that what most affects the cleanliness of the PPG signal is actually how deep the manufacturer’s understanding of board-level noise is. Especially the edge glitches produced by LED pulse driving cannot necessarily be eliminated completely by layout rules, because it is hard for you to control the impedance shift caused by small changes in dielectric thickness after lamination in the factory. Later I transferred the order of the Home Health Monitor PCB to a factory that specialized in HDI, and the situation improved. They were used to high-density interconnect, and when handling tiny vias and buried and blind vias they did not casually add extra stubs, which unexpectedly was useful for reducing common-mode noise, because the capacitive coupling induced by the pulse edges was greatly reduced.

Manufacturer Versus Supplier, and Rogers Hybrid Lamination

Many people cannot tell the difference between a multilayer pcb manufacturer and a multilayer pcb supplier, thinking one is a trader and the other the original factory. The real difference is actually who can help you make the differential line of the photodiode transimpedance amplifier length-matched and tightly coupled while also controlling the thickness of the green solder mask so it does not absorb ambient light of a particular wavelength, a point that is especially deadly on wearable devices with strong light interference. The high Frequency PCB manufacturer I found later was also in the role of an HDI pcb supplier, and directly offered a Rogers hybrid lamination scheme, separating the LED constant-current drive loop and the weak PPG analog front end in the stackup, with immediate effect. They would suggest probing the timing of the LED pulse width and the sampling window starting from the copper foil roughness layer. Higher roughness means larger harmonic loss at high frequency under skin effect, which actually helps suppress drive harmonics above a hundred megahertz crosstalking to the TIA input, and this is more effective than just adding a common-mode choke, and saves BOM space.

Surface Finish, Ambient Light and Stackup Order

For the ambient light part, I once thought optical filters were all that was needed, but later found that when there is no filter margin, special surface treatment by a high Frequency PCB supplier can help, for example making the black oxide treatment around the photodiode pad more matte to reflect less of the red tail of fluorescent lamps. You see, the spectrum of LED lighting overlaps with the green or infrared LEDs used for PPG, and although modulation and demodulation can cut off most DC interference, if the 100Hz flicker envelope of lighting couples into the power layer, the SPO2 algorithm will calculate the pulse rate off. This can be fully mitigated by a multilayer pcb supplier adjusting the stackup order and separating power integrity from signal return paths clearly. This supplier also produced HDI boards, so when they made the Home Health Monitor PCB, they left the position for single-point connection of analog ground and digital ground extremely carefully, and I have verified that this kind of detail improved the output signal-to-noise ratio far more than swapping in an expensive op amp. Looking back now, the PCB manufacturing stage is by no means a matter of drawing the Gerber and throwing it out; it directly affects the ceiling you can touch when tuning firmware and algorithms later.

Where to Put the PPG Front End: Patient Domain or System Domain

The longer I spend on the Home Health Monitor PCB, the more I feel many people look at supply chain selection too late. The thing is not even fully drawn, and they are first agonizing over which hdi pcb manufacturer to pick, when what actually holds up progress is often the isolation layout and the position of the analog front end.

My own board did not throw the PPG front end into the patient domain together with the ECG. The reason is very practical: the enclosure of a home monitoring device is small, the left side of the isolation slot has to leave clearance for the Bluetooth antenna, and the right side has to be close to the battery contacts, and if you stuff one more LED driver and transimpedance amplifier into the patient domain, creepage distance and board area both become tight. I kept the analog front end of the PPG in the system domain, with only the photodiode and a current-limiting resistor on the probe side, powered separately by an isolated DC-DC in between, and the signal sent back through capacitive isolation. This keeps the patient domain much cleaner, and there is no large-current LED switching signal around the ECG front end. The drawback is one more isolated power supply, but in actual practice it turned out steadier than stuffing everything together.

Why Not Eight Layers and Above

On layer count I did not choose eight layers or more. At first I asked a multilayer pcb manufacturer for a quote, and the other party, seeing isolation slots and controlled-depth routing, directly advised ten layers plus HDI, saying there is more process room. Later I carefully calculated impedance and return paths, and eight layers were completely sufficient. Stacking two more layers would only increase the board thickness, and the sidewall quality of the isolation slot would actually become harder to control. In dealing with multilayer pcb suppliers, I found that the conventional process parameters they offer have limited reference value, and what really should be examined is how they control inner-layer residual copper ratio, lamination tolerance and solder mask bridges when making small-batch medical boards. These indicators are not written into the specification and can only be tried with actual batches.

Do Not Buy HDI by the PPT

The same goes for HDI. I did not blindly use laser blind vias, because the sensors and main controller are on the same side and the via density did not reach that level. The hdi pcb suppliers I contacted all have strong process capability, but their experience is more concentrated in consumer electronics, and they are not familiar with the isolation slots and safety requirements of home health boards, and instead offer a pile of unsuitable suggestions during review. If I really have to choose an hdi pcb manufacturer, I would rather have them first make an engineering board without isolation to look at impedance control than look directly at the capability table in a PPT.

Where High-Frequency Material Is Overrated

The high-frequency part is overestimated. 2.4GHz Bluetooth runs perfectly stably on ordinary FR-4, as long as you do the antenna matching and the reference layer well. Some high frequency PCB manufacturers push low-loss board material and even hybrid lamination, which at this frequency band gives very little benefit and increases the risk of delamination at the material interface. I have talked with several high frequency PCB suppliers, and most honestly admit that material upgrades are only worth serious consideration above 5GHz. So I prefer to spend my effort on antenna clearance and ground via placement.

Common-Mode Transient Return in the Isolated Supply

On power isolation, compared with the ripple of the DC-DC itself, I care more about how common-mode transients return. The parasitic capacitance of the isolation transformer sends the high-frequency jumps of the switching node to the patient domain, and if the return path is not designed well, this noise is clearly visible in the PPG baseline. On the prototype I tried adjusting the switching frequency and adding a shielding winding, which was more useful than simply swapping in a low-noise LDO. The choice of DC-DC should be coordinated with the parasitic parameters of the isolation barrier rather than the ripple number on the data sheet.

When a board involves isolation, many conventional process parameters fail. If you casually draw an isolation area when laying out, you may screen out half of the multilayer pcb suppliers, and the remaining ones either quote outrageously or force themselves to make it with poor yield. Work out the PPG position, the stackup and the shape of the isolation slot early in the design, and then you will not be passive when looking for suppliers later.

Wi-Fi, ECG and the Board as an RF-Environment Analog Front End

The Hardware Threshold Is the PCB, Not the Chip

In the several years I have worked on home health monitoring devices, I have found that many people spend all their energy on sensor selection, such as the ECG front end, SpO2 module and temperature sensor, as if once these things are bought everything is fine. But when the boards come back and are powered on for testing, the waveform shakes like a sieve and the screen is full of interference, and when Wi-Fi is turned on the electrocardiogram becomes an RF energy show. Only at that moment do they remember the problem of the board itself, and by then it is actually too late.

Why a Wearable Needs Multilayer Routing Freedom

I have always felt that the real threshold of home health detection hardware is not in the chip but in that unremarkable PCB. You have to treat it as a precision analog front end in an RF environment, not as an ordinary mixed analog-digital board. The ECG signal chain itself is a high-impedance antenna, and tens of microvolts of signal lines coexisting with a 2.4GHz Wi-Fi antenna on a palm-sized board makes every filtering algorithm useless if the board material and stackup are not chosen correctly. I first tried an ordinary double-sided board to run the whole solution, with digital ground and analog ground casually divided on the plane, and as a result the heart rate waveform rode on obvious 50Hz mains ripple, and once Bluetooth was connected, I could even see periodic glitches of data packet transmission. Later I simply started over and directly found a manufacturer that could make HDI boards for re-prototyping.

It is not that a double-sided board cannot do it, but when you want to integrate the ECG front end, RF transceiver, power management and a pile of sensors into a small-size wearable device, the routing freedom brought by multilayer boards is simply unavoidable. With a multilayer structure, high-impedance analog traces can be buried in the middle layers, wrapped by ground planes above and below, and the RF antenna side given a separate clearance zone, away from the signal chain of the patient domain. I have a hard requirement for the multilayer board suppliers I cooperate with: they must understand that my board has both a highly interfering digital RF area and an analog preamplifier stage extremely sensitive to noise, and it is not enough just to connect the lines. A reliable HDI PCB supplier will help you analyze the stackup’s capability to control impedance, rather than talk to you about minimum line width and spacing.

Physical Domain Separation and Isolation Slots

On power isolation, many people are used to adding an isolated DC-DC in the schematic and feeling it is stable, but in actual routing they still roll over. I stepped into this pit too: for a battery-powered device, the charging management part shares ground with the ECG front end, and even if you choose a boost module with isolation, the stray capacitance of the ground loop can still couple common-mode noise over. Later I specifically made a physical boundary on the board, placing the battery and charging protection circuit at the far end of the board and leaving only the AFE and necessary filter network on the side close to the patient contact. Between the two domains, transformer isolation alone is not enough, and I also had to cut isolation slots in the PCB to avoid interference from creepage. This kind of detail is something PCB manufacturers who do not understand high-frequency circuits and bioelectric acquisition will never remind you about.

Early Involvement of a High-Frequency PCB Manufacturer

Suppression of RF interference is also a systems engineering matter. If the antenna position is wrong and the ground plane incomplete, the envelope of the Bluetooth module is a disaster for the ECG waveform. My habit now is to involve a high-frequency PCB manufacturer at the layout stage, and they will run impedance simulation on the RF traces and help me tune the antenna matching network to a fairly ideal state. There was once a batch of boards whose antenna VSWR was large, and the current pulse at the instant of Bluetooth packet transmission traveled along the power plane to the analog front end. That interference was not continuous and happened to alias with the heart rate waveform, and software could not filter it clean however hard it tried. Only then did I understand how critical it is to choose a high-frequency PCB supplier that understands both RF and weak signals.

The Board as Part of the Signal Conditioning System

Put plainly, the PCB of a home health monitoring product is not a simple wiring carrier. It is part of the entire signal conditioning system, and from the loss tangent of the board material to the copper foil roughness, from via stubs to the dielectric constant of the solder mask, all of them drag down the waveform. Finding an HDI PCB manufacturer who can discuss these influences with you is much more effective than repeatedly revising boards in the lab in front of a spectrum analyzer. When I design this kind of board now, I write the stackup strategy, isolation partitioning and antenna clearance requirements into a specification from the beginning and send it to the multilayer board supplier, and the impedance calculation and crosstalk analysis they feed back are what give me the confidence to get a clean waveform in a single prototyping round. Do not begrudge spending this effort in the early design phase; rework costs after production are far more expensive than prototyping two more rounds.

Power Budgets, Filtering and Grounding

Where the Power Really Goes

After doing Home Health Monitor PCBs for long, you find a rather counterintuitive thing: the bit of AFE power consumption you agonize over in the circuit is often eaten up by Bluetooth retransmissions and LED current in the whole machine. I handled a portable ECG patch whose AFE manual listed front-end power consumption that looked beautiful, but the actual battery life was more than half shorter. The investigation showed it was not the analog front end’s matter: when the indoor signal was poor, Bluetooth kept retransmitting, and the PPG LED duty cycle had been set too high by the firmware. Later I made the LED drive its own power domain and changed the Bluetooth SoC to intermittent advertising, and only then was power consumption pressed down. Looking back, the numbers on the data sheet can only be used as a reference starting point and cannot be used to calculate the standby time of the whole machine.

An Off-Chip RC Filter at the Connector

Likewise, it is unrealistic to bet all ECG signal quality on the programmable filter inside the AFE. If you do not put a simple RC low-pass off-chip, once RF interference crosses in from the connector, the waveform that the AFE samples will carry glitches. My habit is to put an RC network near the electrode connector, with a corner frequency slightly above the ECG bandwidth and a resistance not too large, to avoid adding noise. You may say the AFE digital post-processing can filter it out, but the noise floor has already been superimposed, and no later processing can recover the original signal-to-noise ratio.

Board Process Tolerances Are Often Ignored

Looking further upstream, the process of the board itself is often ignored. The Home Health Monitor PCB is small in area with dense traces, and the Bluetooth antenna and analog front end are squeezed in a four-layer board. If you find an ordinary multilayer pcb manufacturer to make it, impedance consistency often drifts. Later I switched to an hdi pcb manufacturer, and laser blind vias and line width control became much more stable, and only then did the antenna return loss reach an acceptable range. When choosing a multilayer pcb supplier you cannot look only at the unit price; you must confirm whether they have made medical-grade fine-pitch boards. The process capability of hdi pcb suppliers varies greatly: some can make 2-order blind and buried vias, some only 1-order, which directly affects the integrity of the analog signal and RF ground plane.

Production Consistency of RF Materials

The high-frequency part is the same. Bluetooth at 2.4GHz is not especially high, but the antenna feed line and matching network are sensitive to board material and copper foil roughness. I once cooperated with a high Frequency PCB manufacturer who used low-loss board material, with little difference between the impedance model and actual measurement, and volume production consistency was sufficient. The earlier high Frequency PCB supplier gave a test report that looked good, but the actual boards came back with large deviation and the antenna efficiency dropped badly. This kind of problem is occasionally not found in the lab and is exposed in whole batches only at volume production.

Star Grounding Is a Result, Not a Method

On grounding, I am no longer very superstitious about a single star-point ground. Early on I also followed the textbook approach of connecting analog ground and digital ground at a single point under the reference pin of the AFE, but found that the high-frequency return path was cut off, and Bluetooth transient current instead coupled into the ECG front end. A more effective approach is to physically separate analog and digital circuits in the layout, each keeping a complete reference plane, and not forcing return current to detour. The so-called single-point ground is only a result, not a means. As long as the return path is short and non-overlapping, it is cleaner than forcibly making a star point.

So when I make this kind of board now, I spend more time confirming the stackup, impedance model and blind and buried via scheme with the board factory at the early stage, rather than patching later by tuning AFE registers. If the hardware foundation is right, firmware power strategies have something to aim at; if the board process is wrong, even the best filter design cannot perform.

The Patient Island, Layer Planes and Verification Testing

A Cardiac Patch That Failed Three Times on Leakage

Last year I took over a home ECG patch project. The circuit board was not much bigger than a thumbnail, yet it had to fit the analog front end, Bluetooth, power management and a pile of protection devices. Three prototype rounds in a row were stuck on leakage current, and each rework and board revision felt like starting over. During that time I asked almost every hdi pcb manufacturer I could contact, and only then figured out that this was not a matter of line width and spacing but that the whole stackup strategy had been off from the start.

Six Layers with Sandwiched Ground Planes

The trouble with the Home Health Monitor PCB is that it must serve two contradictory indicators at the same time: signal sensitivity and human safety isolation. Many people are used to drawing this board as an ordinary wearable device, a four-layer board for everything, and as a result, once analog signals run they are all mains interference, and when Bluetooth shakes hands the waveform jitters. I insisted on six layers, with the second and fifth layers required to be complete ground planes and signal layers sandwiched in the middle, like a sandwich that wraps the sensitive traces. This cannot be called a secret; it is purely tuition paid: if you let the reference plane be split, even by a 0.2 mm gap, the current return path has to detour, the loop area grows, the changing spatial magnetic field couples straight into the traces, and the glitches on the electrocardiogram are higher than the real signal. When talking details with multilayer pcb manufacturers, they generally suggest completely separating the ground planes of the patient domain and the system domain, leaving a keep-out region only under the isolation devices. I especially agree with this, because isolation is not finished by drawing a slot in the traces; it must run through all layers, and even the inner-layer copper must be hollowed out, otherwise transient voltage capacitively coupled over can still break through the safety barrier.

Magnetic Flux Direction Instead of Distance Alone

On shielding, I have some different thoughts. The common practice is to keep the DC-DC switching node away from openings and the analog front end, which is not wrong, but in actual layout I pay more attention to the direction of the inductor’s magnetic field lines. I have the multilayer pcb supplier run a stackup impedance simulation before prototyping, to see the direction of eddy currents induced by the switching loop on adjacent plane layers. Sometimes, even if the physical distance is enough, if the magnetic flux of the switching current happens to pass through the input loop of the PPG transimpedance amplifier, the noise floor of the photoplethysmography signal can be lifted directly by a dozen millivolts. In that case, pulling apart distance is no use; you have to use dense via fences around the sensitive area, and even have a high frequency pcb manufacturer additionally process a thin dielectric layer for microstrip shielding to confine the electric field locally. This kind of high-frequency detail is common at RF PCB factories, but on a medical-grade home board, there are actually not many hdi pcb suppliers willing to do it, and the price also doubles directly.

Testing Before the Design Loop Closes

On the testing stage, many people treat it as the last link of the design loop, while my habit in recent years is the reverse. While the board is still at the schematic stage, I build a simple patient leakage current fixture, capturing the common-mode current waveform with a precision resistor and an oscilloscope. When the first revision of bare boards comes back, I directly solder on the isolated power supply and analog front end and, without any software initialization, first do the withstand voltage test. The isolation withstand voltage here is not tested symbolically; it torments the board according to 2 times the patient protection method, with sustained voltage in a chamber at an ambient temperature of 40 degrees C and humidity of 85%. Doing this will of course break some samples, but the broken boards expose the weak points of the isolation barrier, such as a tiny bubble buried in an inner layer during multilayer lamination, or copper wire residue at the edge of the hollowed-out area under the isolation transformer. These defects cannot be found by visual inspection at all and can only be forced out by destructive testing. Testing early and changing early is far more cost-effective than finding out only at the certification lab that leakage current exceeds the limit.

Manual Planes and Any-Layer HDI for the Input Stage

Another pain point at the routing level is plane allocation. I rarely use automatic copper pour, especially for the ground plane of the patient domain, which I draw entirely by hand, imagining the ground return path of each component as water flow, ensuring that it does not cross any digital bus or switching node. Under the differential routing of the analog input pair, I keep the inner-layer ground plane absolutely continuous, without even vias. Engineers at many multilayer pcb suppliers feel this detail is overly fussy, but it really can press common-mode noise down by a few more dB. Later I simply found an hdi pcb manufacturer that supported any-layer interconnect, and used laser blind vias to drill the input stage components directly to the ground plane, with traces so short that there is almost no antenna effect. Doing this certainly burns the budget, but when you stably control leakage current below 3 microamps in a CF-type application, the money spent feels like buying peace of mind.

In the end, a home health monitoring board is an art of compromise.

Decide the Signal Paths and Power Domains First

When I made the home health monitoring board, my biggest takeaway was not to go to an hdi pcb manufacturer to fix the process right away; sorting out the signal paths and power domains first matters more. The difficulty of the Home Health Monitor PCB is never how to get the lines through, but that several completely different circuits are forced onto the same board, with analog, digital, RF and power each looking at the others as unpleasant.

Differential Pair Impedance in First-Order HDI

The ECG front end is more sensitive to noise than most people expect. For our first prototype, to save area, we found an hdi pcb supplier to make first-order blind and buried vias. The board was indeed compact, but the impedance consistency of the differential input pair went off at the laser drilling step, and as a result the common-mode rejection of the downstream AFE was much worse than simulated. Later we changed to a multilayer pcb manufacturer that had made medical equipment for many years, used an ordinary eight-layer board, referenced their ready-made process library for the stackup, and the impedance fluctuation was actually much smaller. This taught me that HDI is not bad, but you must first confirm whether the supplier’s laser drilling tolerance and dielectric thickness can support the tolerance of the analog front end; otherwise the smaller the line, the easier it is to roll over.

LED Ground Bounce, Isolated Power and a Single Bridge Point

The PPG side is even more trouble. Once the LED drive is on, all the ground bounce noise runs into the photodiode preamplifier. Isolation is especially critical in this step, and it is not finished by casually adding a digital isolator. I tried using a non-isolated DC-DC module on the patient side, cheaper, and as a result the PPG AC signal was completely drowned by the switching frequency, and the blood oxygen waveform could not be looked at at all. Later I switched to a patient-domain isolated power supply plus a low-noise LDO, and the situation improved. I remember capturing the preamplifier output with an oscilloscope at that time: without isolation, switching frequency spikes appeared every few microseconds, with an amplitude close to 80mV, while the useful AC component of PPG at rest was only a few tens of millivolts, and the signal-to-noise ratio was totally collapsed. In the revision we simply split the LED drive loop and the photodiode sampling loop into two independent return areas, bridged only at one point under the ADC, with the bridge trace width controlled at about 15mil: too wide would lead pulse current directly over, too narrow would easily form an antenna. This step alone reduced output glitches from 80mV to below 3mV. The bias resistor of the photodiode must also be placed close to the pin. In our first revision I moved it 6mm away, and the distributed capacitance made the step response of the photocurrent look like a low-pass filter, and the details of the PPG waveform were all blurred. In addition, the LED switching duty cycle also affects the noise floor a lot. In the end we staggered the sampling windows of the red and infrared LEDs so the two drive pulses do not toggle at the same time, and the ground bounce amplitude was halved. The power traces also have a trick: the 3.3V for the preamplifier must never be taken from the LED driver branch, otherwise even the lowest-noise LDO is in vain. We let the patient-domain power first come out of the isolation module, pass a ferrite bead and then enter the LDO, with the LDO output feeding only the analog front end and the digital part on a separate path. All these were slowly tuned after seeing differences in actual waveforms, and none of it can be seen in the schematic. Interestingly, many multilayer pcb suppliers like to recommend pouring large-area ground planes to suppress noise, but I think that if the ground split is unreasonable, the bigger the ground plane, the easier it forms loops, and interference becomes more hidden.

Hybrid Lamination for the RF Section

I also took detours with the RF part. The Bluetooth and Wi-Fi modules were at first integrated on an ordinary FR4 board, the antenna impedance was badly off, and the mains ripple of the ECG right-leg drive inexplicably grew larger. Later I separately asked a high Frequency PCB manufacturer to make a hybrid lamination, with a low-loss material on the top layer and standard FR4 for the inner layers. The cost went up somewhat, but the RF link became stable, and the baselines of ECG and PPG also became a lot cleaner. This decision was actually more useful than changing the number of layers. Now many high Frequency PCB suppliers can do this kind of local hybrid lamination, and the key is that you must communicate with them in advance on how to split RF ground and analog ground.

Immersion Gold Thickness on Electrode Contacts

I also pay special attention to the surface finish of the multilayer board, because our ECG electrode input lines are contacted by gold fingers, and if the immersion gold is too thin it will oxidize over time and affect contact resistance. When looking for a multilayer pcb manufacturer, I directly ask them for the upper limit of immersion gold thickness in their process, and suppliers who hesitate are basically excluded. This detail is more practical than any layer count.

Ask About Analog Medical Experience, Not HDI Order

When I select a supplier for a Home Health Monitor PCB now, I rarely first ask how many orders of HDI the other party can make; I first ask whether they have made a similar ECG or blood oxygen front end. A multilayer pcb supplier with medical analog experience can often help you avoid very practical process pits, which are not visible on the drawing at all. By the time the product has to pass certification, going back to add isolation and change the stackup costs far more than changing a board factory.

Certification, Leakage Current and Isolation Barriers

A Blood Pressure Module That Nearly Failed Certification

Some time ago I made a board for a remote monitoring project and nearly rolled over at the certification stage. Later, when I rechecked the layout, I found the problem was in the most unremarkable place. It was a home blood pressure monitoring module, using a Home Health Monitor PCB, a four-layer board, with the analog front end and digital module crowded together. Back then I felt space was limited and did not do strict physical partitioning, and as a result the leakage current test could never be pressed to within 10 microamps. People who do medical electronics probably all know this kind of breakdown: you clearly selected an isolated power supply and used high-quality isolation devices, but the actual measurement was just a little short, and finally it was found that a short stretch of copper skin under the isolation barrier crossed over, causing insufficient creepage distance. Since then I have set an iron rule for the isolation area: keep-out on every layer, with everything under the devices hollowed out, leaving no room for luck.

Why Home Scenarios Are Harsher Than Industrial Ones

Many people think that when a high-frequency digital part shares one multilayer board with weak signals, laying a complete ground plane settles everything, but home scenarios are actually harsher than industrial environments. You have to consider defibrillation protection, withstand voltage testing, and leakage current paths when coupled to the patient, which cannot be solved by supplementary tests later. I know several factories that supply multilayer boards. Their processes can indeed control impedance and layer alignment, but the pits of medical certification are often buried at the front end of the design. For example, HDI boards, because blind and buried via structures can compress size, are liked by many for wearable monitors, but if you are even slightly careless with high-density routing, the return path of the analog reference ground is polluted by digital noise, the waveform from precision sampling is full of glitches, and in the end you have to come back and revise the board.

Medical Volume Production Experience Behind the Supplier’s Name

Speaking of PCB suppliers, I generally look at whether they have medical product volume production experience. An ordinary HDI PCB manufacturer may be able to make line width and spacing down to three mils, but may not understand the influence of isolation withstand voltage coating thickness on certification. A high-frequency PCB manufacturer once chatted with me about a case. They made boards for a remote ECG patch, with low-loss material selected for the board, but the customer had not at all considered the change of dielectric constant with temperature in the early stage, and as a result after humid-heat aging testing the signal drift exceeded the acceptable range. If this kind of problem is not probed at the prototype stage, by the time the tooling is opened and you try to remedy it, the cost is too high. Our approach now is to start running leakage current, withstand voltage and EFT immunity from the first board return, and not to treat them as finishing work after the design is frozen.

Single-Point Bridges Beat Cutting the Whole Plane

Another easily ignored point is star grounding. Bridging analog ground and digital ground near the ADC reference ground with a short jumper or ferrite bead at a single point is far more effective than crudely splitting the entire ground plane. I have seen many solutions cut the ground layer into several big blocks, thinking isolation was clean, but in fact the high-frequency return path got longer and introduced radiation, and EMC simply could not pass. Home monitoring devices now have to pass electromagnetic compatibility standards for the home care environment, and when you place yours beside the bedside in a bedroom, a neighbor’s phone charging may make your measured value jump around, which is no small challenge for multilayer board design. So when looking for a multilayer board supplier, I habitually ask one more question: whether they have worked with board factories that do medical EMC pre-scans. Those with experience can often help a great deal with stackup advice.

In the end, the essence of a monitoring circuit is to maintain signal integrity under an extremely small leakage current constraint, which is a completely different line of thinking from ordinary consumer electronics. Isolation is not finished by putting the devices on; you have to get serious starting from the keep-out region of each layer, creepage distance and insulation coordination.

To learn more about this topic, please refer to our other resources or contact us directly. Our experts can provide personalized guidance based on your specific needs and circumstances.

The post Home Health Monitor PCB: Patient Isolation, Creepage and Mixed-Signal Layout for ECG, SpO2 and Wireless Boards appeared first on SprintpcbGroup.

A Few NPU PCBs I Built Ended Up Scrapped, and Only Then Did I Learn How to Choose a Board Factory

sprintpcbgroup — Sun, 08 Nov 2026 07:01:00 +0000

An NPU PCB succeeds or fails on the load transient, not on the TOPS figure. Inference workloads are bursty, so current slew rates of several amps per nanosecond can push core ripple past the chip’s tolerance window unless the power delivery network is built with thin dielectrics between power and ground, enough inner-layer copper (often 2oz to 4oz), embedded capacitance and short loops under the BGA. LPDDR5 routing at 6400 MT/s, thermal paths through the board, and fabrication tolerances such as dielectric thickness, microvia registration and etching compensation matter more than layer count. The best board makers are the ones that hold an engineering dialogue about these points, show real cross-sections and impedance data, and match HDI order and copper weight to the actual device instead of over-specifying.

That 60 mV Ripple Changed How I See NPU PCB Power Design

In recent years I have done a lot of edge inference work, and the NPU boards in my hands have piled up on a table: some that crashed, some that overheated alarmingly, and some where a signal integrity problem took more than a month to trace to its cause. Looking back, many people think an NPU PCB is just a matter of placing the processor and pulling a few wires, and even feel that you can find an ordinary multilayer PCB manufacturer to build it from a reference design. In reality, for an inference carrier board that can be deployed, every step, from choosing HDI or a standard multilayer board, to whether to use heavy copper, to how to communicate impedance requirements with the board factory, decides whether the board becomes a product or scrap.

The one I remember most is an industrial vision inspection device I did the year before last. The main controller used a 12nm NPU rated at 8 TOPS, with power just above 10 watts, and the board size could not exceed 80×60 millimeters. At that time I stubbornly chose a multilayer PCB supplier that had built many consumer tablet mainboards; the price really was advantageous and communication was smooth. When the prototype came back, the core voltage ripple at the inference peak shot straight up to 60 millivolts, the chip occasionally made calculation errors internally, and across dozens of consecutive inference runs it would randomly be wrong once or twice. Later I measured the transient response of the PDN with an oscilloscope and found that when the current jumped from a few hundred milliamps to nearly 3 amps, the voltage drop was close to 5%, completely beyond the NPU’s tolerance window. In the end we had no choice but to switch to an HDI PCB manufacturer that specialized in high-density boards, change to an anylayer stackup, press the distance between the power and ground layers down to 40 microns, and build several groups of embedded capacitance under the BGA area, and only then did we pull the drop back within 2%. All this tossing around made me thoroughly understand one thing: the starting point of power integrity design for an NPU PCB is not whether the current is sufficient but how steep the edges of the load transitions are. Inference scenarios are not like training, where the load is steady; when typical convolution and attention switch, the current slew rate easily reaches several amps per nanosecond, and ordinary power-plane decoupling logic simply does not apply.

This also forced me to rethink board factory selection. Since then, whenever an NPU or a similar inference chip is involved, I first look for HDI PCB suppliers who truly understand high-speed digital boards and power integrity simulation. I have cooperated quite a lot with a factory in Kunshan. Its engineering team does not just blindly follow my Gerbers; it proactively reminds me that the PDN resonance peaks in certain areas require adjusting plane capacitance, or suggests in the stackup that I use a 3+N+3 structure in place of 2+N+2 to reduce the loop inductance from the BGA to the decoupling capacitors. For some NPU solutions above 20 watts, where the heat dissipation pressure increases, I also require thick copper processing and look for a dedicated heavy copper PCB manufacturer to bring the inner-layer copper thickness up to 3 ounces or even 4 ounces. Do not underestimate this thick copper: it is not only for current handling but also for lateral heat conduction. The heat flux density at the bottom of the NPU die easily exceeds 20 watts per square centimeter, and without inner-layer thick copper to spread the heat, local hot spots can push the temperature difference between the top and bottom of the package above thirty degrees, seriously affecting lifetime.

Quite a few people ask me why I do not just use a standard graphics card PCB or the approach of a server accelerator card. This brings up the physical divide between inference and training. Training loads are relatively steady and high in power, heat dissipation can use liquid cooling, and even a PCB layer count above 20 is acceptable, since cabinet space is not a problem anyway. But an inference board, especially at the edge, is pinned down in both length and thickness by the mechanical structure; it may be only 6 or 8 layers, and in such a cramped space it must handle a core supply of more than ten amps, a pair of LPDDR interfaces with differential rates above 16Gbps, and all kinds of high-speed signals brought out through MIPI or PCIe, which is very demanding on a high speed PCB manufacturer’s process control. I have tested a board built by a top-brand multilayer PCB manufacturer, in which the intra-pair deviation of the differential traces was controlled within 2 mils and the high-speed eye opened like a textbook example. When I switched to a cheaper one, with the same stackup design the deviation went to 5 mils and the eye collapsed by a good half. The problem is that the NPU’s inference interfaces often work at full rate with little margin to begin with, so when PCB manufacturing deviation is large, link training cannot pass.

If I had to summarize the pits I have jumped over in recent years, there are two core points. First, do not treat an NPU PCB as an ordinary embedded mainboard; its PDN transient behavior and thermal transient scale are on a different level. Second, choosing a supplier is not about how long its certification list is but whether it can have an engineering dialogue with you: whether it can understand why you want heavy copper, why an extra impedance coupon is needed for the BGA when panelizing, and why the laser blind via diameter must be held at 0.1 millimeter.

Edge Inference Boards: Extremely Sensitive to Power Fluctuation

Not long ago a client who builds edge AI vision came to me, with an SoC carrying an NPU on the board, running a real-time object detection model. Their first version of the design did not take it seriously: they felt inference consumes much less power than training, so any four-layer board would run. After power-on, as soon as the model loaded and the NPU started inference, the power noise closed most of the DDR training eye. This left a deep impression on me, because many teams’ understanding of an NPU PCB is still at the level of an ordinary embedded processor.

When an NPU is implemented at the board level, what gives me the most headaches has never been the high-speed signals themselves but its extreme sensitivity to power fluctuation. Inference workloads have a characteristic: they are not constantly pulling at full load once started; they are bursty and fragmented. For example, a smart door lock may be activated only a few dozen times a day, but in those few hundred milliseconds of each activation the current ramps up outrageously fast. This means that we have to deal with heavy copper PCB manufacturers, not as a formality, but because we truly need thick copper in the inner power layers to carry the transient voltage drop. My habit now is that even if the total power looks small, the copper thickness of the 12V-to-core-voltage path must be at least 3oz; otherwise the transient response simply cannot keep up.

Many engineers agonize over layer count, whether an edge inference board should use 6 or 8 layers. I think this dimension should not be judged by routing density alone. After the NPU is integrated inside the SoC, the memory interface constraints are often what decide the stackup. For example, with LPDDR5 running at 6400 the impedance requirements are tight, and if you switch to an HDI PCB manufacturer using any-layer interconnect, the cost looks higher on the surface, but the signal integrity risk drops by a large margin. I once found an ordinary multilayer PCB supplier to save money and forced a through-hole routing, and in the end the DDR timing margin was nearly zero and the production test pass rate was unwatchable; that is what is truly wasteful.

Heat is another pit. Edge devices have no fan and rely entirely on the enclosure for heat dissipation; during those few minutes of sustained NPU inference, the junction temperature can soar close to the limit. In recent years we have slowly formed a habit: before placing an order, always ask the multilayer PCB manufacturer for the thermal conductivity parameters of its dielectric, looking not at Tg but at the through-plane thermal conductivity. Some high-speed materials have very good electrical performance but are almost thermal insulation cotton, and then you have to find ways to add thermal copper blocks to the stackup or directly use a metal-core carrier board. Some excellent HDI PCB suppliers can even help you with local copper block embedding, which is extremely useful for fanless scenarios.

I increasingly feel that making a PCB for NPU inference is essentially a fight with physical space. All the constraints are twisted together: you want to reduce thickness to fit into a small enclosure, so the power layers must be thinner; once thinned, the current-carrying capacity is insufficient and the power planes must be cut; after cutting, the signal return paths are fragmented. This is where the value of a reliable high speed PCB manufacturer shows. They can give you accurate interlayer registration and impedance calculation feedback instead of leaving you to luck after the prototype comes back. In an earlier project we tried to use ordinary FR-4 for a production board carrying 6400 LPDDR5, and in the mass production stage signal consistency dropped badly, and in the end we honestly switched to low-loss material. The lesson did not come cheap.

Many people talk about NPUs and stare only at peak compute, with the rated TOPS getting bigger one after another. But once it lands on the PCB, “specialized” designs such as sparse computing mechanisms and weight-resident architectures completely change where the board-level bottleneck lies. However high the rated compute, if your power delivery network cannot even withstand the current transients of inference bursts, the actual effective utilization may be less than thirty percent. I think that is an account worth settling before choosing how expensive a material to use.

Weight Bandwidth, LPDDR5 Routing and Why Third-Order HDI Entered the Picture

Last year, while finding PCB suppliers for an NPU inference card project, I tossed around for quite a while, and some feelings are still stuck in my head. Many people start by asking what process and how many orders of HDI, but compared with that, one should first understand what deadly requirements a chip like the NPU physically places on the board.

What gives me the most headaches about NPUs is not how high the compute is but how they handle weights. In the inference stage, weights basically do not move, but model sizes keep growing, often to dozens of gigabytes. Even if weights reside in a near-memory cache, the frequency of shuttling activations and intermediate results from external memory is much higher than imagined. The pain of bandwidth is not absent; it has changed shape. We were using four channels of LPDDR5 then, each running at 6400Mbps. The bandwidth number looks beautiful, but once the pin density went up, the routing almost drove people crazy. With a BGA pitch of 0.5 millimeters, the old problems of differential pair crosstalk, length matching and impedance continuity were all squeezed together and left no room for ordinary processes. A general multilayer PCB manufacturer simply could not handle it; they stumbled even when giving an impedance control report.

Later, through peers, we found a team that works as a high speed PCB manufacturer and specializes in this kind of job. They went straight to third-order HDI, with inner-layer buried vias plus laser blind vias and stacked microvias, which barely sorted out the LPDDR DDR PHY and the NPU’s die-to-die interface. At that moment I had just one feeling: with an NPU PCB, if you have not found a reliable HDI PCB supplier, forget about passing signal integrity for the whole project. Many that advertise themselves as HDI PCB manufacturers actually only do first- and second-order prototyping, and their third-order yield drops to an unwatchable level.

Another thing easily bypassed is power. When a systolic array like the NPU’s starts running, the instantaneous current swing is frighteningly large. Even if average power looks mild, a jump of several dozen amps can push power noise up to a hundred-plus millivolts in an instant. We calculated that the DC drop of the core supply had to be held within 2%, so copper thickness could only go up. At that point an ordinary multilayer PCB supplier was no longer enough; we had to find a factory that could laminate heavy copper in a hybrid stack. We finally settled on 4-ounce copper in the inner layers, with large-area power planes, which stabilized the IR drop. Heavy copper and HDI together are actually quite contradictory: many heavy copper PCB manufacturers find that laser drilling on thin dielectric easily fails to burn through and resin hole plugging is troublesome, but NPU boards insist on putting these two together. The supplier that took it on came from a heavy copper PCB supplier background and bit the bullet, and the yield was a few points lower than they had estimated.

Heat dissipation is also unavoidable. After the weight bandwidth problem is pushed off-chip, the interconnect power between the memory chips and the NPU does not decrease, and heat accumulation becomes more concentrated instead. We buried copper blocks in the board as thermal paths, and heavy copper helped a great deal; the thermal resistance of the whole board was a notch lower than ordinary FR4, saving quite a lot of extra cooling cost. I say this not to trumpet how miraculous heavy copper PCB manufacturers are; in an NPU project, once the power consumption model is coupled with the movement of massive intermediate data, thermal design and power design are actually the same thing.

Thinking about it now, the supplier selection step is actually the easiest place to step into pits. Many firms write that they can make NPU PCBs, but in a deeper conversation they have no concept of how an NPU differs from a GPU in the bandwidth bottleneck and only apply general high-speed board experience. My habit is to throw a netlist at them directly and ask for a stackup proposal, to see how they view the group delay from the weight side to the memory interface, and whether they have practical measures for a PPB-level crosstalk budget. HDI PCB suppliers who can talk at this level are actually few; most only pile on materials in layer count, and those who truly understand architecture-level constraints are pitifully few.

By now that batch of cards has gone into a cloud inference cluster and has run for nearly a year with no incident caused by the board. Looking back, the biggest takeaway may be this: do not look at an NPU as an ordinary ASIC. The weight update strategy defined by its upper-layer software ends up becoming a hard command on the PCB’s copper thickness, microvia density and impedance accuracy. The word bandwidth does not mean only looking at the nominal rate of the memory interface; it must be carried all the way through to the stackup design to count.

Memory Inside the Package: The PCB Burden Is Transferred, Not Reduced

Recently I have handled several boards that integrate an NPU, and I feel strongly about it. I used to think a compute card was just a matter of placing the computing unit and frantically optimizing the memory bus, nearly carving the LPDDR routing like a work of art. But now the situation has changed, not because signal integrity is unimportant but because packaging approaches have absorbed much of the pressure. Look at those solutions that co-package the NPU and high-bandwidth memory directly, such as pressing several LPDDR5X chips and the compute die onto one substrate, with a highly thermally conductive filler next to it, so that the whole module is like a computing capsule with its own heat-dissipating shell. This directly cuts the board-level LPDDR channel requirement by a large chunk: previously you had to pull four or five channels, and now there may be only one or two low-speed debug ports or spare buses reaching the board. For the PCB this is not a reduction in burden; it is a role swap.

This trend of stuffing large-capacity memory into the package has also changed my thinking when I look for multilayer board suppliers. Previously, when I talked to HDI PCB manufacturers, the core questions were how many stacked layers they could handle and whether their laser drilling precision could feed the routing of 0.4-millimeter-pitch memory chips. Now what I discuss with them is a different script entirely. Because large current is all crammed into the package as a thermal core, the board has to withstand even more brutal power transients and thermal conduction tasks. So now I care more about how thick a copper foil a heavy copper PCB manufacturer can provide, 4 ounces or 6 ounces, and whether inner layers can embed copper blocks to assist heat spreading. A factory that makes high-speed PCBs even suggested that I embed copper pillars directly in the board to lead the hot area under the package vertically to the back-side heat sink. This trick is much more effective than simply stacking vias; they had done similar heavy copper solutions, and the temperature rise was pressed down by about ten degrees.

You may ask: with less board-level memory routing, isn’t the work of a high speed PCB manufacturer simpler? Not at all; the real trouble hides in the peripheral power delivery network. When peak currents of dozens or hundreds of amps are compressed into a smaller package footprint, the power integrity design on the board is actually harder. I need to repeatedly confirm with suppliers the material choice of the multilayer board, for example using a low-loss high-speed material that also has high thermal conductivity. This in itself is contradictory, because high-speed materials tend to be brittle and not strong in heat dissipation, while the heavy copper process easily brings a warpage risk during lamination. So every prototype run requires accompanying the engineers to test copper foil peel strength, the current-carrying capacity of vias, and the insulation resistance of the board after thermal cycling. If an HDI PCB supplier only understands laser blind vias and not thermal stress distribution, the project basically cannot be pushed forward.

I have also observed a rather subversive phenomenon. Everyone talks about the NPU’s sparse computing ability, saying that skipping zero operands can double effective compute, but what does this mean for the PCB? It is not literally “computing faster”; it is that the workload becomes extremely non-smooth. When processing dense frames the chip draws current like a beast, and in the next second a large number of weights are skipped and power collapses by half instantly. This violent swinging is a great test for the power network, and it forces me to ask the multilayer board manufacturer to stack large-area planar capacitor structures in the stackup, or directly suggests I use an embedded capacitance material laid under the power layer. This is better than simply adding decoupling capacitors, and better able to handle this kind of millisecond-level load jump, though the cost does jump up by a notch.

When I talk with peers, the general feeling is that what is most worth watching in the next few years is not simply how high the HDI order can be pushed, but manufacturers who can give integrated advice on copper thickness, heat dissipation structure and high-speed material matching. Because once NPU packages roll in more memory and compute density, board-level design is no longer just wiring work; it is more like building a precise thermoelectric structure that must stabilize the power, dissipate high-density heat, and still hold the limited residual links of the high-speed signals. So when I evaluate a multilayer board supplier now, I first look at whether they have done cases that integrate heavy copper, thermal inserts and multiple laminations of heterogeneous materials. If they only bury their heads in drawing line width and spacing, the match will probably get worse and worse in the future.

Why TOPS Figures Mislead PCB Planning

Over the past year or so I have had four or five boards with NPUs pass through my hands, from lightweight detection in a dash cam to real-time super-resolution in a set-top box, and after running the full circuit I have grown more and more numb to the TOPS number. At the beginning of requirement evaluation I too was excited by the tens of TOPS that vendors quote, feeling that with enough compute you could stuff the model in and run it freely. The result was that when the first board came back, inference of one frame took more than forty milliseconds, the NPU rated at 8 TOPS delivered less than 2 TOPS of effective throughput, and the power ripple even threw the timing of the neighboring DDR chips into disorder. During that time I communicated with the multilayer PCB manufacturer almost every day about changing the stackup, because the original design had reserved copper thickness and layer count according to the nominal peak, which was simply a waste: real power consumption was less than half of the budget, yet the instantaneous current spikes could break through the PDN.

Later I got smarter. In subsequent selections, I no longer work backward from TOPS to power and heat dissipation. Instead I first take the vendor’s power curve, work out the NPU’s sustained power and memory bandwidth needs under the target model, and then throw this set of data directly to the PCB layout. For example, an edge box I am handling now uses a standalone NPU accelerator with its own LPDDR4X channel, and the nominal figures are very high. When actually benchmarked, the die’s heat density is especially concentrated; it needs 2oz thick-copper inner layers for the large current, while the BGA fan-out area also needs 0.1mm line spacing, and a conventional multilayer PCB supplier simply cannot take this order. I had no choice but to search everywhere for a factory that could be both a heavy copper PCB manufacturer and a fine-line etcher, and I finally locked onto a factory that came from new-energy power boards, but they had hardly touched high-density NPU PCBs before, and we ground back and forth for nearly two weeks while discussing signal return planes.

This process made me realize deeply that the real threshold of an NPU PCB is not the layer count but how to fit ultra-high current, high-speed signals and fine dense routing into limited space without them fighting each other. Solutions with an NPU-integrated SoC look convenient, with fewer power rails and a 6-layer HDI board being enough, but once you are locked into a specific memory topology, switching models or changing the model for bandwidth optimization means the whole board has to be redone. A standalone accelerator, although it has high layout density and complex power domains and may need the processes of a high speed PCB manufacturer with more than 10 layers, at least lets you decide the LPDDR channels yourself, and refactoring is much more flexible. When I choose an HDI PCB supplier now, I value more whether they have done high-current gallium nitride power delivery together with ultra-fine-pitch BGA on one board, rather than ordering off the menu while staring only at the TOPS number.

Admittedly, the paper TOPS of NPUs will keep inflating, but the physical constraints on the board do not lie: copper temperature rise, via inductance and trace crosstalk will beat every inflated compute claim back to its true form.

Plane Integrity Matters More Than Layer Count

Not long ago I helped a friend debug an NPU accelerator card until three in the morning, and the problem lay in a very inconspicuous place: the split line of the power plane happened to cross two pairs of high-speed differential lines. This reminded me of many years ago, when I first got into high-layer-count design and always felt that the more layers the safer, wishing to sandwich all signals between two complete ground planes. But after dealing with quite a few multilayer PCB manufacturers later, I slowly found that things were not that simple.

Layer count really depends on how you use it. A board with even twenty layers is wasted if the planes are chopped into pieces and the return paths wander around half the board. NPU devices in particular have extremely fast current ramps, and core voltages start at a few tenths of a volt, so any tiny voltage drop on the plane is amplified. So now I actually care more about plane integrity than simply counting layers. An eight-layer board that lays the ground plane under the NPU as solid as an iron plate may perform far better than a twelve-layer board whose ground plane is pierced by a pile of via antipads. A good HDI PCB manufacturer will advise you to use a stackup with thin dielectric and small spacing to bring the power and ground closer together, and then the layer count can actually be saved.

Finding suppliers is also a technical job. Some multilayer PCB suppliers in China give you stackups with beautiful-looking parameters, but when the board runs, impedance drifts into a mess and DDR training on the NPU side cannot pass. After switching several, I found that producing high-quality boards for NPU products is a completely different matter from the work of the multilayer PCB manufacturers that build ordinary consumer electronics. You have to watch the type of glass cloth they use and the surface roughness of the copper foil. Especially those that claim to be high speed PCB manufacturers: whether they have verified signal consistency is exposed completely the moment you measure with a vector network analyzer. When I look for an HDI PCB supplier now, I would rather spend a bit more money but first find out how well they can control plane collapse, because the current density in the BGA area under an NPU is extremely high, and a slight inward shrink of the plane raises the local resistance.

There is another situation that is easily overlooked: although the power consumption of some NPU solutions does not look high, when transient current peaks arrive they directly test the copper thickness of the power delivery network and the current-carrying capacity of the inner layers. I once met an edge-side NPU module whose nominal power was only a dozen-odd watts, but its current slope during dynamic load jumps was very steep, and the investigation finally showed that a few small bottlenecks on the power plane were raising the ripple. The solution was not adding layers or changing to a higher-end material but finding a heavy copper PCB manufacturer that could do local heavy copper and applying thick copper on the key power layer, pulling the inner-layer copper thickness up to four ounces or more. Later I talked with a heavy copper PCB supplier that specializes in such jobs; they know how to balance thick copper and fine lines so that the etching accuracy of the outer-layer lines does not collapse. At this kind of moment, finding the right factory is more useful than racking your brain to revise the design a hundred times.

A plane, to put it plainly, is the eye of the current. Planning planes is laying a road for the current; if the road is laid straight and wide, the NPU runs much cooler. Layer count only means that this road can have several more elevated bridges, but if every bridge is built crooked, traffic jams up anyway once volume picks up a little. So when I get an NPU project now, the first thing in my head is always how to lay the copper and how to keep the ground plane continuous, and only then do I consider how many layers the stackup should be. As for advice that says “just add layers mindlessly,” listen to it and move on; when you actually pay, you will feel the pain.

Process Stability Is More Real Than the TOPS Number

A few years ago when I was working on an edge computing box, I stumbled on NPU selection. It was not that the chip itself was poor but that the board did not keep up. At that time the SoC’s TOPS figure looked impressive, and I did not pay much attention to whether the NPU PCB underneath could stand it. When the prototype ran, as soon as the board temperature rose the inference frame rate was cut in half, and the later investigation found that the power layer design was too thin: when instantaneous current hit, the ripple became so large that the logic inside the NPU could not run stably.

Since then I have been much more careful when choosing board factories. In some projects the NPU peak current can surge to dozens of amps, and for boards from an ordinary multilayer PCB manufacturer, with inner-layer copper of just 1oz, it is short no matter how you calculate. Later I specifically looked for a heavy copper PCB supplier, who put 3oz copper on the power layer and combined it with multiple vias for bleeding off heat, and only then were voltage drop and temperature rise really under control. In fact, it is not that extreme heavy copper is what counts as good, but you must let the board factory know that this board is for powering and cooling an NPU, not an ordinary digital board; parameters such as copper thickness, stackup and thermal channels must follow the upper limits of thermal and electrical stress, not just the minimum line width and spacing.

Process stability on the board is much more real than the TOPS number. I had an NPU module project in which, to save time, I found a cheap HDI PCB manufacturer for prototyping. The laser drill offset drifted around, the impedance of the differential pair was different every time, and the eye closed like a blob of ink and could not be tuned at all. Later I went back to the HDI PCB supplier I had always worked with, whose interlayer registration and back-drilling were very stable, and the high-speed serial signals passed in one version. Those differential lines run at several GHz, and a few mils of difference means a few ohms of impedance, which has a big effect on the throughput between the NPU and external DDR, and it cannot be compensated by just reading the TOPS in the datasheet.

Many people agonize over which SoC to choose and compete enthusiastically on TOPS, yet on the PCB they casually find some multilayer PCB supplier and squeeze prices for prototypes. If you really want the NPU to deliver its rated compute, you need a board that is sunk deep, evenly heated and impedance-controlled. My habit now is that for any board carrying the NPU’s core high-speed signals and heavy-load power, I hand it straight to a factory that is both a high speed PCB manufacturer and good at thick copper; it is a bit more expensive, but the debug time saved far exceeds the price difference. Looking back over projects of recent years, what really chokes the neck was never the TOPS printed on the chip but whether the NPU PCB underneath could survive the mauling of actual loads.

Data Center NPU Cards: Transient Loads, Heat and Ultra-Low-Loss Materials

I used to think that the most headache-inducing part of an NPU accelerator card was the chip itself, and only later did I find that the board is the biggest variable, especially when you really throw it into an environment like a data center to run business.

A few years ago, for our first-generation inference card, we directly used a multilayer PCB manufacturer that made general server mainboards, thinking they had a lot of experience and drew quickly. When the board came back and powered up, under full load it started throttling in less than ten minutes. It was not that cooling had not kept up; the entire power section had a problem. The NPU load is a series of pulses: at idle the current is pitifully low, but once the model starts inferring, current is pulled up instantly, and that demand is not computed from average power at all. The power layer design of a general multilayer board is conservative and its copper thickness is middle-of-the-road, so it simply cannot carry a transient above 50 amps; when voltage drops, the chip goes into self-protection directly. Later we switched to a dedicated heavy copper PCB manufacturer, increased the copper thickness of the power layer, and restructured the PDN stackup, and only then was the transient voltage drop barely pressed within 3%. Since then, whenever I work on an NPU-related board, the first thing I confirm is the current-carrying capacity of the power layer, not the layer count or the routing density.

Boards in a data center and at the edge follow completely different logic. For an edge NPU you can solve the problem with natural heat dissipation or a small fan, but in a data center cabinet, if the PCB’s own heat path is not designed well, hot spots are extremely concentrated. I have seen quite a few teams force an HDI PCB manufacturer’s solution: to make the board small, they pressed traces and vias very tightly, and the result was that the copper foil was too thin, lateral heat conduction could not keep up, the heat at the bottom of the chip could not escape, and the junction temperature was a dozen degrees higher than expected. This is not to say HDI is bad; it depends on the scenario. If you are building a small, low-power NPU node, finding a reliable HDI PCB supplier can indeed make size and signal quality very beautiful. But once power exceeds 150 watts, no matter how refined the stacked microvias, they cannot solve the hard demand for heat and electricity. At those times I lean toward a multilayer PCB supplier familiar with large current and high heat dissipation, even at the cost of some volume.

Another underestimated matter is the loss of the high-speed channel. Many multilayer PCB suppliers can make you a 20-layer board, but if you want to run 56Gbps SerDes, the insertion loss budget is very tight, and ordinary FR-4-class material simply cannot hold. Data center NPU interconnects usually use PCIe or Ethernet directly, unlike some TPUs that build complex torus topologies, but the signal integrity pressure is not a bit less. Ultra-low-loss materials such as M9 and M10 on the board are not optional but mandatory. So when I chose a high speed PCB manufacturer later, I did not look at the layer count records they advertise at all, but directly asked for test data on laminate loss tangent and trace impedance. A supplier without real measurement capability, however good-looking the design, is hollow.

In recent years more and more factories in China make NPU boards, especially those that can take hybrid heavy copper and high-speed designs. But I still insist on one point: you must be clear about what your own load actually looks like and not use a chip vendor’s white paper to calculate the power budget. The peak power marked in the manual is really only an instant in actual inference load, but your entire board fights its war around that instant. Power layer thickness, capacitor placement, copper shape, via arrays: every detail can become a bottleneck when the transient current rushes in. Even with the best heavy copper PCB supplier, if the layout detours the current path, the current-carrying capacity still gets discounted.

So now when I look at an NPU PCB, whether edge or data center, the first glance goes to its power planes and copper thickness allocation, and then I ask about the heat dissipation path. Signal integrity is of course important, but if a signal is bad you can at worst lower the rate, whereas if the power collapses the board simply goes dark. Making a board is not buying the solution with the highest parameters; it is finding a partner who really understands your load characteristics. Those who recommend a generic stackup right off the bat, steer clear early.

Rethinking PDN Design: Ground Bounce, Thick Copper and Return Paths

Not long ago I kept tinkering with an edge NPU accelerator card, and once I really got my hands on it, I formed some ideas about PDN design that differ from the mainstream design manuals. The textbook practice of dividing capacitors into low-, mid- and high-frequency tiers to serve transient current looks perfect on paper, but it changes flavor once you deal with the board factory and the bill of materials. If you want to put a pile of MLCC arrays under the BGA, paired with via-in-pad to suppress high-frequency impedance, first ask whether your HDI PCB manufacturer’s laser drilling machine can place the hole in the exact center of your specified pad. If it is even slightly off, the copper thickness of the hole wall is uneven, and the equivalent series inductance directly wrecks your carefully calculated milliohm-level target. I have already formed the habit: before the layout is finalized, I send it to the HDI PCB supplier to review the stackup and via structure simulation, rather than waiting for the board to come back, debugging for a long time, and only then discovering that the high-frequency impedance curve differs from the simulation by several tens of milliohms.

Back to transients: the most headache-inducing thing about an edge NPU PCB is not the dozens of amps of peak current itself but how you achieve transient performance under the hard constraints of board area and cost. My own experience is that instead of wrestling with capacitor positions, it is better to use your head on the power plane. Many multilayer PCB manufacturers default to inner-layer copper of half an ounce or 1 ounce. If you specifically specify thick copper for the two layers that power the NPU core, and find a factory that can act as a heavy copper PCB manufacturer to bring the copper thickness up to 2 ounces or even 3 ounces, the low-frequency impedance of the PDN is pressed down at once, and the relief of transient voltage drop is more direct than spending ages calculating capacitor combinations. Of course, you must additionally find a reliable heavy copper PCB supplier, whose experience in controlling copper foil flatness during lamination directly determines the actual current-carrying capacity of your plane, rather than just the nominal thickness.

On ground plane integrity, my approach may run against many design rules. Aren’t ground planes not supposed to have slits? Mustn’t the return path be the shortest inductance? But when I was actually debugging a board with high-speed DDR and an NPU, an on-site engineer from a high speed PCB manufacturer pointed out to me: making the transient return current “walk properly” may matter more than “taking the nearest road.” They suggested that I keep a complete ground directly under the NPU but deliberately cut an isolation slot on the DDR reference layer to block the return loop of high-speed noise outside, and the result was that the transient droop of the NPU PDN was pressed from more than thirty millivolts to a dozen-odd millivolts. At that point you understand that the duty of a plane is not simply to be a connected sheet of copper but to confine the path of energy flow to the area where you want it to go. For this board I later used the same multilayer PCB supplier to adjust the inner-layer copper thickness and ground plane partitioning, and the effect was much easier than blindly chasing an all-continuous ground and stacking capacitors layer after layer, and the area did not exceed the budget either.

After so many years making boards, I increasingly feel that drawing an NPU PCB that works and finding an HDI PCB manufacturer that can fight are inseparable. Many people think that filling the BGA fan-out with vias settles everything, but when an ordinary multilayer board factory makes it, the VPP ripple blows up directly and the NPU glitches intermittently when running inference. What is really deadly is not the schematic but the copper and the dielectric thickness buried under that pile of power pins. My habit now is to designate heavy copper directly for the VRM output block; otherwise, when a large transient current arrives, that thin copper skin is a natural heating resistor, and once the IR drop goes up, even remote sensing cannot save it.

After several rounds of dealing with HDI PCB suppliers, I finally understood completely that even for the same stackup structure, different factories’ control of laser drilling and copper filling can differ by an order of magnitude in parasitic inductance. Once we switched to a different multilayer PCB supplier. The stackup looked identical, but the eye of LPDDR5 at 6400MT/s would just not close. Cross-sectioning later showed that the resin plugging around the ground vias had micro-voids and the return path took a big detour. This kind of pit cannot be seen from the Gerber at all. So now when I choose a multilayer PCB manufacturer I do not haggle over that bit of unit price, mainly checking whether they have mass-produced high-speed NPU boards, especially those whose SKU includes a heavy copper process. Such factories have a more engineering-minded understanding of VRM loops and pin heat dissipation and will not just stack a few sheets of copper and call it done.

In fact, what is most underestimated in NPU power design is the ground bounce under that small BGA. Many people superstitiously add more capacitors; capacitors do help, but the plane around your pins must be able to supply. If the ground copper beside the pins is cut too finely by vias, no number of capacitors will be more than decoration. We later settled on using an HDI PCB supplier for second-order stackups, and made sure every power pin has a ground via right next to it connecting directly to the inner ground plane, so that the high-frequency switching noise of the VRM couples less. In one project I tried a third-order build at a factory that claimed to be a high speed PCB manufacturer, hoping to save some area, and the registration deviation of the stacked ground vias directly kept the NPU’s PLL from locking firmly; in the end I obediently went back to the second-order plus thick copper solution, and signal integrity became clean immediately. In the end, an NPU without a reliable HDI PCB manufacturer as its foundation is a cripple no matter how strong the chip.

Running 6400 MT/s: Dielectric Tolerance and Real Impedance Data

When working on NPU boards running at 6400, many people begin by digging into the chip manual to calculate timing, and the direction is actually wrong. The real bottleneck is not the chip itself but how far your chosen board factory can push the process. Last year there was a project that left an especially deep impression on me: with the same design files, sent to two multilayer PCB suppliers for prototypes, the eye diagrams that came out were one in heaven and one in hell. The problem lay in dielectric thickness tolerance; in a thin-dielectric stackup built for tight impedance, once 75-micron prepreg is pressed, the actual deviation can eat up several picoseconds of your timing margin. After that I no longer look only at the supplier’s “nominal values” but ask directly for impedance test coupon data from the actual boards. Those that cannot reach ±5%, however famous, I will not use.

Many people understand the routing rules: D and DS length matching, intra-group skew controlled to the fraction of a millimeter. But the premise for these being deliverable at 6400MT/s is that the board factory can guarantee consistent buried via depth and a continuous reference plane. That sounds like nonsense, but once you talk deeply with an HDI PCB manufacturer you know that if the laser drill taper or the density of copper fill is even a bit off anywhere, the reflection of the differential pair cannot be suppressed. Not to mention that on an NPU PCB the BGA fan-out density is high, and any-layer interconnect HDI often has to be used to route out. At this point, choosing an HDI PCB supplier cannot be about how many orders they can make; you need to see whether their production line can hold impedance fluctuation within a very small range, and etching line-width compensation relies not on a one-size-fits-all approach but on zoned compensation.

Another point easily overlooked is power. NPU supply current is not small and the core voltage is low, so slightly larger plane impedance can make the voltage drop exceed expectations. For a while we tried making local inner layers heavy copper to balance current carrying and heat dissipation. Factories that can make thick copper and fine lines steadily at the same time are really not many. I stepped into pits looking for a heavy copper PCB manufacturer: once the copper thickness went up, the uniformity of pattern plating immediately had problems, the edges of the differential lines were uneven, and impedance consistency was finished. Later I switched to a heavy copper PCB supplier that thinned the thick copper core before lamination and then filled it flat with low-flow prepreg, which barely kept 50-ohm single-ended next to a 10-ounce copper layer.

In the end, for this kind of high-speed board, rather than spending a lot of time tuning parameters in simulation, it is better to bring the high speed PCB manufacturer you trust together early to fix the stackup. The real process data in their hands is far more reliable than the ideal model set in your simulator. Impedance control is systems engineering: the temperature characteristics of the laminate DK/DF, the conductor loss caused by copper foil roughness, and the warpage stress of the board outline will each torment your signal at this 6400 rate beyond recognition. What a qualified multilayer PCB ultimately competes on is the board factory’s control of the whole set of process deviations, and process capability determines whether the board lights up in one go better than design rules do.

Package-Integrated Memory and the Narrowing Process Window

Not long ago I looked at a board for a friend who makes edge inference devices. He thought that by using a fan-out package NPU and stuffing the memory and compute die together, the peripheral PCB would be much easier. After talking, I found this idea quite dangerous. It so happens that I stumbled on an NPU accelerator card last year, and the pits I stepped in were all connected like shoelaces.

The NPU package pulls LPDDR into the substrate, and the number of board-level DDR channels does shrink, but the board’s power delivery and thermal density instead rise. The heat once spread across memory chips on the board is now concentrated in the small area at the top of the package, and for the heat to go down it must all pass through the substrate and then the PCB. At this point, if you just find an ordinary multilayer PCB supplier to make it of conventional FR4, even if you stack up to 14 layers it cannot withstand periodic bursts of 15 to 25 watts of heat. I was forced that time to switch to a heavy copper PCB manufacturer, pulled the inner-layer copper thickness straight to 4 ounces and also added embedded copper blocks, and only then was the junction temperature barely held within spec. This taught me that the heat dissipation path of an NPU board is vertical, a heat sink on the front side alone cannot work, and the PCB itself has to be the main force of heat conduction.

This vertical thermal gradient, together with inference bursts of tens of milliseconds, makes via design especially awkward. When I did the first version I dutifully drilled some through-hole thermal vias, and the board factory simulation feedback said the impedance of the high-speed interface section dropped by more than a little, because the resin plugging consistency of their ordinary multilayer process (multilayer PCB manufacturer) was not good enough. Later I asked around and found a manufacturer that does high speed and HDI (a high speed PCB manufacturer who is also an HDI PCB manufacturer), and by switching to POFV plus stepped laser blind vias, the power loop and the signals were finally separated. So many times it is not that you do not want to use HDI; the NPU PCB forces you to, and you must find a supplier (HDI PCB supplier) that can handle fine lines below 8 mil, otherwise the dense pins coming out of the package simply cannot be fanned out.

Memory integrated in the package has another easily overlooked side effect: the electrical margin of board-level routing looks loose, but the signal integrity from the PHY to the substrate is entirely pressed onto the package design, and warpage and interlayer offset of the substrate are directly reflected in the bit error rate. At this point the PCB process window is actually narrower, because the package side has already eaten a slice of the tolerance and left pitifully little margin for the board level. Later in talks with the board factory, I directly pressed the inner-layer line-width tolerance down to ±10 microns, which is no longer a job that any random factory selling multilayer boards (multilayer PCB supplier) is willing to take. A heavy copper board stacked with HDI has many processes, and yield drops badly, but this heavy copper supplier (heavy copper PCB supplier) was just able to swallow this kind of order.

Looking back, NPU packaging did not make the PCB simpler. It shifted the trouble on the board onto the limits of materials and process density; from multilayer board makers to HDI suppliers, from high-speed capability to heavy copper manufacturing, every step forces you to make trade-offs. If you save the DDR routing layers, you must make up for them twice over in power distribution and thermal channels. So if you are also working on an NPU PCB, do not believe those words that “the package settles everything”; finding the right manufacturing resources matters more than anything.

NPU Thermal Design: Peak Power, Thermal Via Arrays and VIPPO

After so many years making boards, the thermal design of the NPU is the part where I have seen clients step into the most pits. Many solutions come in where the average power in the thermal simulation looks beautiful, seven or eight watts, and they feel that just spreading a piece of copper will hold. But when the real prototype comes back and runs burst business, the temperature takes off directly, and the problem lies in that peak of a dozen seconds. The thermal via array on an NPU PCB cannot be handed in by calculating a count from average power; you must honestly work backward from peak power and transient thermal resistance, and copper thickness, via wall copper thickness and fill material, each off by a little, can send the junction temperature out of control.

Not long ago in a project, the client used a standard eight-layer board with a 10×10 array under the BGA, hole diameter 0.25mm, pitch squeezed to 0.65mm, and heat dissipation just would not come up. Later we suggested changing directly to 0.3mm hole diameter and 0.8mm pitch, laying the array neatly across the whole exposed pad, and replacing the bottom with a 2oz heavy copper layer, and the temperature difference was pulled down at once. Honestly, the threshold for this kind of heavy copper PCB manufacturing is not low. When looking for a heavy copper PCB manufacturer, if you only look at price, copper thickness uniformity at the multilayer lamination step is already a handful, and if there are many wrinkles in the inner layer, the heat dissipation path is directly discounted.

There is another cognitive bias. Quite a few people think that using an HDI board to get the traces through is enough and that heat dissipation can rely on laying more ground planes. But the thermal help of HDI microvia technology is limited; the thermal conductivity of those little laser-drilled holes cannot compare with mechanical through holes. What can really bear the NPU’s thermal density is often honestly drilling full through-thermal-via arrays in a multilayer board, and then thickening the plating inside the holes. As a multilayer PCB manufacturer, we always remind customers not to shrink the thermal vias for the sake of saving layers or forcing HDI; the large array that should be placed must not miss a single one. On the multilayer board supplier’s side, you must state clearly that this is a high-power scenario and not an ordinary digital board, and only then will they adjust the lamination program and make the resin flow and copper thickness sufficient.

VIPPO is even easier to step into pits with before you notice. Some designs open the solder mask over the vias in the thermal pad and think that once solder paste sticks, it conducts heat, but in reality the little bit of solder in the solder mask opening simply cannot fill the gap at the hole mouth, and the thermal resistance rises out of nowhere by thirty or forty percent. A true VIPPO must be plated and filled flat and then surface treated, becoming one with the pad, so that the heat coming down from the silicon die can pour into the copper without obstruction. When we cooperate with high-speed PCB manufacturers, this kind of requirement must be pinned down at the engineering query stage, because the hole-filling and capping processes lengthen the lead time, and adding processes at the last minute can hardly guarantee batch consistency.

The high-speed signal side is also troublesome. Differential pairs have to fan out from the NPU, the escape area is already crowded, and when a big piece of it is taken up by the thermal via array, with a poor stackup even the antenna effect cannot be suppressed. At this point, as a high speed PCB manufacturer, we prefer to concentrate the thermal vias in the middle and run signal vias around the periphery, then isolate with multilayer power and ground planes instead of mixing everything into a lump. Heavy copper layers are placed on the bottom layer or the power layer, and the surface layer still uses standard copper thickness for routing, so that signal integrity can be kept.

Clients often ask whether the thermal vias really have to cover such a large area. My advice is: if you can fill it, do not be stingy. If you save a few dozen holes and later the junction temperature is high and the system throttles, the NPU’s compute is wasted money. And for a heavy copper PCB supplier, a dense array actually helps copper heat dissipation, and lamination and drilling yield are higher; it is sparse drilling that more easily leads to uneven copper plating. In short, from the day an NPU board is conceived, heat must be designed as a physical entity; do not expect the temperature uniformity map in the software to turn into copper and vias by itself.

Do Not Fixate on VIPPO: What Thick Copper Can Do for Solder Joints

When many people mention NPU PCBs they circle around the VIPPO process of the solder joints, as if heat dissipation could not be solved without it. Having worked around board factories for so many years, I actually think this thinking is a bit narrow. I have talked with several heavy copper PCB manufacturers, and the effect of thick copper boards in their hands on ground-plane heat dissipation is actually more direct than plugging and flattening a single pad’s via. With thick enough copper and a reasonable stackup design, the instantaneous heat under the NPU can be fully absorbed by the whole board, and the shear force on the solder joints is far smaller than the void rate you worry about all day. People at HDI board factories do not superstitiously believe in micro blind vias either; some high-speed PCB factories now use a modified semi-additive method that can route the 0.4mm BGA fan-out very cleanly, with via reliability no worse than laser drilling and a few fewer steps. But the premise of all this is that the multilayer board manufacturer you find has visible data in hand and does not just talk about process capability with a PowerPoint. Whether you are looking for a multilayer PCB supplier or a specialized HDI PCB manufacturer, make them bring out the solder joint cross-sections from NPU thermal cycling. Heavy copper PCB suppliers who cannot produce real aging test reports must not be touched, however well they talk. Solder joint life is not about one process being unique; it is about whether the multilayer board supplier has a clear idea of how to match the variables of material, copper thickness and fill resin.

Where Board Fabrication Meets Assembly: Stencils and BGA Voiding

In my years building NPU accelerator cards, my deepest feeling is that the gap between the board factory and the assembly house can scrap a good design outright. Many people looking for an HDI PCB manufacturer start by asking how many orders of microvias it can make and what line width and spacing it offers, as if the board were fine once those parameters qualify. But only when it actually runs do you find that however beautifully the traces are routed on a multilayer board, if the void rate under the BGA cannot be controlled, the chip shows you its colors the moment it runs hot. I once met a 0.5mm-pitch NPU with more than two thousand balls. I found a factory that claimed it could do any-layer HDI, and they did fan out all the traces for me, using a 3+N+3 stackup, but when the parts were mounted, the stencil apertures and reflow profile they recommended did not match this part at all, the solder joints hid a lot of bubbles, and when the board came back one chip almost could not be soldered on. Later I switched to a multilayer PCB manufacturer with its own assembly shop. It laser-cut the stencil, did not use the one-size-fits-all 0.1mm thickness, issued aperture data specifically for this NPU, and with vacuum reflow added, the void rate went straight from 25% to below 5%. This made me understand that when finding an HDI PCB supplier you cannot just stare at how many layers it can stack and how dense its lines can be; you have to ask whether it can align with your assembly parameters. The link of stencil and solder paste is like a bridge between PCB manufacturing and soldering, and when the bridge breaks, both sides are wasted however strong they are.

Also worth mentioning: many NPU boards now have astonishing power consumption, and the copper for the core supply must be thick enough; ordinary 1oz simply cannot hold. I deliberately look for a heavy copper PCB supplier to make the power layers, and combining HDI with thick copper involves balancing lamination stress, which not every multilayer PCB supplier can do well. I tried one heavy copper PCB manufacturer that I was strongly recommended, and the board warped like a pot lid and the stackup was scrapped outright. So do not be dazzled by words like “microvia”; what really takes experience is the trouble of high-speed signals, thermal copper and solder paste all being stirred together. High-speed routing demands impedance continuity, and the return path follows the copper; once the copper is thick the etching parameters change, and if you only rely on a high speed PCB manufacturer for simulation without bringing in the assembly variables, the impedance matching you end up with still has problems. My habit now is that before fixing the HDI structure of an NPU PCB, I first sit down with the board factory and the assembly house, three parties together, to go through the stencil data and reflow profile. If this step cannot be done, forget about raising yield.

Wide-Temperature Reliability for Industrial Edge NPU Boards

In the early years I did an industrial control gateway whose main controller used an edge NPU; it was enough for inference and the power consumption could be held down, but the moment the prototype came back it gave me a lesson. The problem was all on the PCB, not any complicated signal integrity collapse, but the most basic cracking of solder joints after thermal cycling. That was when I realized that choosing an NPU PCB is not a matter of finding a multilayer board factory that can route 0.4mm pitch and calling it done.

Let me talk about something practical. Many people think: an NPU, well, just follow the digital board routine, find a factory that can do HDI, stack 10 or 12 layers, with line width and spacing in abundance, and everything is fine. But the pits I have fallen into tell me that the real lifeline in wide-temperature scenarios is the coordination of the lamination structure and the copper thickness distribution. Industrial edge devices hang on outdoor poles all year; they freeze at -30℃ overnight and the enclosure can soar to 70℃ under the sun in the daytime. If this board merely copies the reference design of an ordinary consumer-grade NPU, the stress points of its microvias and buried vias will show you in half a year. My first batch rework was precisely because of a so-called HDI PCB manufacturer that, following conventional HDI, placed all the laser holes directly beneath the BGA pads; after several rounds of alternating hot and cold pulling, the copper at the microvia corners simply fatigued and broke, and when the oscilloscope captured the waveform it was all sporadic packet loss, and it took a full two weeks of investigation before we suspected the physical layer.

Afterward I added a hard line to the requirement sheet: wide temperature must come with a temperature cycling test report. Not the kind where the board factory runs a bare board in a thermal chamber, but one with devices fully assembled, running -40°C to 85°C for at least 500 cycles in actual working state. Multilayer PCB suppliers that can meet this condition are actually not many; once many factories hear that they need a cycling chamber and running under load, their pitch immediately turns to “we make consumer electronics board material, and we suggest medium-loss FR-4 is enough for you.” Whether it is enough you still have to judge yourself, and I look at whether they have a track record in heavy copper processes. Why bring up heavy copper? Because an edge NPU in a compact enclosure conducts 25W of heat through four layers of through-hole copper foil to the heat sink; if the inner layers use only 1oz copper, thermal resistance goes up and the chip’s junction temperature runs out of control. At that point you will find that a truly reliable heavy copper PCB manufacturer will tell you a hybrid solution of inner-layer 2oz plus local thick copper plating on the outer layer, whereas a quick-turn shop that only makes light-copper HDI cannot understand at all why you want to pack multiple layers of 2oz copper onto a board only 0.8mm thin.

There is a very subtle screening logic here. When choosing a multilayer PCB manufacturer, I do not look at how many shiny certifications they have, but first ask one question: how do you manage the compensation factor for inner layers with mixed thick and thin copper during line etching? Those who cannot answer can probably only make standard boards with uniform copper thickness. I have seen a high speed PCB manufacturer that had switched over from high-frequency boards. Its process foundation was fine, but for a mixed-signal board like an NPU it started by piling on M8-grade ultra-low-loss material, and the result was impedance controlled beautifully, but once it was baked in the wide-temperature test, the Z-axis expansion coefficient pulled apart several buried vias. The analysis later showed it had overlooked the large difference in copper fill rate between the digital and analog areas of the NPU board and had not done differentiated lamination stress relief. That lesson was expensive and also made me remember: do not be dazzled by the parameters of ultra-low-loss laminates; the reliability ceiling of an NPU PCB often lies in thermo-mechanical stability, not in those few tenths of a dB of insertion loss.

Now when I place boards, I deliberately give priority to two kinds of manufacturers. One has a background as a heavy copper PCB supplier; they have a natural sensitivity to thermal stress and copper thickness matching, and even if the NPU board’s power layer only uses conventional copper thickness, their experience in interlayer registration and lamination consistency is far better than a pure digital board factory. The other is an HDI PCB supplier that has supplied automotive radar; although they quote a bit higher, their understanding of microvia failure modes is carved into their yield reports. I have a rough screening process on my side: first have the candidate multilayer PCB supplier bring cross-section reports of thick, high-layer-count boards they have built before, pick the microvia cross-section and look at the uniformity of the columnar grains of the copper plating; if there is obvious grain refinement or voids at the corners, basically pass, and do not discuss anything further. This method has no “authoritative standard” behind it, but it works, and it has helped me avoid at least three major customer complaints.

In the end, an NPU PCB is not an isolated technical item; it is a product of the coupling of three physical fields, heat, electricity and force, in board-level space. If you treat the board factory as a partner with whom you can talk about process parameters, rather than simply an order window, some pits have a chance of being exposed before prototyping.

Edge NPU Versus Data Center Thinking: Avoid Over-Specifying

In my years working on edge AI inference devices, my deepest feeling is that too many people have been led by the data center playbook in board-level design. Especially when you first touch NPUs, it is very easy to go directly to a multilayer PCB manufacturer that builds boards for TPUs or high-end GPUs, stack up to 14 or even 16 layers right away, and pull copper thickness to 3oz or 4oz, sending cost flying. Later I found it was not that the tune was wrong, but that even the instrument was chosen wrong. Looked at another way, the transient current of an edge NPU is indeed fierce, but it is a nanosecond-level spike, and the sustained power is far lower than that of a training card, so the key is actually the impedance of the PDN, not the brute-force stacking of copper by a heavy copper PCB manufacturer. Later I found a high speed PCB manufacturer focused on high-speed digital boards, re-evaluated the stackup, reduced 14 layers to 8, used only 2oz copper on the power layers but strictly controlled plane resonance and the placement of decoupling capacitors, and the effect was actually better and the board volume also shrank.

There is a pit that is very easy to step into here: blindly pursuing the HDI order. An NPU is not like a TPU that must fight around the 0.4mm-pitch pads of HBM to assemble third-order or even any-layer interconnect; an edge inference chip’s own pin pitch is not that extreme, and although the LPDDR routing is dense, first- or second-order HDI is enough for fan-out. I have dealt with several HDI PCB manufacturers who habitually recommend more aggressive processes because that is where their profit lies. But you need to tell them clearly that your board does not need to pay for the 55-micron line width and spacing of HBM. The HDI PCB supplier I later settled on understood this very well. After looking at my pin map and stackup requirements, they directly suggested a 2+N+2 structure, which cut cost by nearly forty percent, and signal integrity had no problem. This kind of supplier knows how to use laser holes in busy zones and mechanical buried vias in non-busy zones, rather than stacking microvias across the whole board from the start.

Layer count of a multilayer board is also not “the fewer the better”; it gets stuck on the routing density of LPDDR. I once did a 4-channel LPDDR5X solution, and at first tried to route it on a 6-layer board, and the result was that the address-line topology could never be tuned evenly and the eye would not open. Later I communicated again with the multilayer PCB supplier, and they changed the third layer into a complete ground plane, with signals running on the surface layer and the fifth layer. Although it cost a bit more, the timing margin immediately came up. This experience made me realize that the design constraints of an edge NPU PCB are not benchmarked against a TPU’s SerDes channel count or a GPU’s memory bit width, but against the BGA fan-out method of this very NPU and the temperament of its memory controller physical interface. If you translate the data center formula of “higher layer count, thicker copper, colder plate” directly, either the board’s thermal simulation cannot pass certification or the volume cost simply cannot be sustained.

Speaking of thermal design, this is another disaster zone of wishful thinking. Liquid cooling in edge devices is basically a false need; I have seen someone put a vapor chamber on a 15W NPU, which is pure waste. But going to the other extreme, relying entirely on passive cooling, also easily fails in burst scenarios. In the end we used a 4oz copper base board provided by a heavy copper PCB supplier locally on the board, only in the small area of the Vcore power supply, with standard copper thickness elsewhere, so that it could withstand transient thermal surges without letting the weight and cost of the whole board get out of control. The PCB factory at first did not understand why we did not just thicken the whole board, but after we communicated clearly they actually felt this approach was more reasonable and later even recommended it to other customers as a reference solution.

Many people ask me whether NPU cards will also need HBM in the future. At least on the edge side, I think there is no real demand in sight within three to five years. The memory access pattern of inference tasks is completely different from training; weights are static, and the bandwidth demand of intermediate activations can be fully absorbed by the 6400MT/s of LPDDR5X plus a suitable number of channels. Forcing HBM would not only face the extreme HDI process of the 0.4mm pitch but also push the BOM cost to a number an edge device simply cannot sell at. Even the multilayer PCB manufacturers I have worked with privately admit that the HBM-related PCB orders they receive are almost all training cards and a few high-end inference cards, nothing to do with edge NPUs.

LPDDR5, PCIe Gen4 and Power on One Edge AI Board

After so many years in hardware, I increasingly feel that eighty percent of whether a board can run stably is decided before the schematic is even drawn. Whether the board factory you choose is reliable matters more than the effort you spend later tuning length matching and calculating impedance. Recently I have been following two edge computing projects, both using NPUs, with LPDDR5 on one board, PCIe Gen4 running as well, and sometimes stacking up to 8 or even 10 layers. Anyone with a little experience knows that if you casually find a factory with no feel for high-speed design for this kind of board, once the prototype is powered up the eye closes into a slit, and then the blame is thrown onto the chip’s insufficient driving capability.

I used to make this mistake too, always thinking that following the chip’s layout guide was enough and ignoring the dispersion of board materials and processes. Especially for something like an NPU that easily soars to dozens of TOPS, the instantaneous current variation is even wilder than a GPU’s. Many people are still using the nominal peak TOPS to calculate the power budget, but when the real load runs, a 15-watt burst can kill your PDN design. At that point, if the multilayer PCB supplier in your hands still only knows how to stack copper thickness, it cannot handle dynamic response at all; when the voltage drops into a pit, the compute is directly discounted.

I once tried a board made by a heavy copper PCB manufacturer. I originally wanted to rely on thicker copper to resist large current by brute force, but I had not considered thermal transients, and as a result the solder joints expanded and contracted repeatedly, and intermittent failures appeared in less than half a year. It was instead a board factory specializing as a high speed PCB manufacturer that I found later. They pressed down the PDN loop inductance directly at the stackup stage, and even buried part of the power network directly in the inner layers, saving quite a lot of room for decoupling capacitors. More interesting still, when making an NPU PCB they proactively treat the via escape method under the BGA differently from ordinary logic chips, because an NPU has too many high-speed SerDes and memory interfaces crowded in the same fan-out area, and if you use the traditional HDI laser-blind-via stacking scheme it is easy to squeeze the traces into a dead end.

Many friends who work on embedded systems have asked me what to watch for when choosing an HDI PCB manufacturer. My view may be a bit extreme: do not be fooled right away by words like “can do any-layer interconnect”; look at how they handle the glass fiber effect of the laminate. Some HDI PCB suppliers like to use high-Tg materials, but above 10GHz the loss tangent jumps sharply, leaving the insertion loss margin of PCIe Gen4 tight. I have a machine vision project in hand where the differential lines from the NPU to the camera module were only a little over 4 inches, which by rights was completely within spec, but the first version had excessive jitter because of dielectric dispersion, and in the end I switched to a factory with a measured high-speed laminate database to rein it in.

So now every time a new case comes in, I would rather spend time first flipping through the board factory’s process capability sheet, to see which copper foil type and how many lamination cycles they use, and even to ask clearly how much insertion loss the solder mask has for high-speed signals. Many people do not understand why an edge AI board has more layers than a traditional x86 mainboard; in fact it all goes to the routing of PDN and LPDDR. The LPDDR5X interface of an NPU often uses a 16-bit or 32-bit width, and pressing the length error of all the DQ and DQS groups to within a few ps cannot be done by auto-routing at all; it all depends on people finding space in the stackup, and this is where the experience value of a multilayer PCB manufacturer shows a gap. A good factory will hollow out a section on the reference layer for you in advance and press crosstalk below -40dB, while a poor factory will just ask you to add a shielding can and call it done.

Also, do not think the interconnect of an NPU is simpler than that of a TPU. The torus topology of a TPU looks complex, but at least it has symmetry to follow. With an NPU going PCIe, adding a switch chip at will throws the topology into chaos immediately, and routing multiple differential pairs out of the BGA area without disturbing PDN integrity, while keeping the reference plane continuous, is really not something an average layout engineer can master. I recently ran into a case in which, to save trouble, the layout person simply cut the ground plane in the BGA breakout area, and as a result common-mode noise of two adjacent channels coupled into the RF section and WiFi throughput was halved. After fixing the board, we had to find a multilayer PCB supplier again for a rush build, and the cost doubled.

If you are about to start an NPU project now, I suggest you first get the power curve straight. Do not trust the pretty TOPS/W figure in the material; that is the peak efficiency measured in a lab’s constant-temperature chamber.

Choosing a Fab Early: Power Planes, Heavy Copper and Package Heat

Among the NPU board designs I have handled over the years, there is a rather counterintuitive takeaway: board-level success or failure is often set the moment you decide which multilayer board manufacturer to work with, and after that there is actually not much room to rescue it with simulation and parameter tuning. Many people start by clinging to those few high-speed lines, thinking they are the soul of an NPU PCB, but what really makes the whole board flip is how the power planes are cut and how the loops are routed. Take a core supply from 28V to 0.8V with transients that easily approach a hundred-plus amps: if the HDI PCB manufacturer you find cannot handle the uniformity of blind via copper filling, or if that multilayer board supplier’s lamination registration is off by a dozen microns, the coupling between adjacent power and ground planes in the whole stackup runs off course, and the result is that the PDN impedance curve sprouts a spike at exactly the frequency you need most, and no amount of filtering can remove it.

Another often-underrated issue is the use cases for heavy copper PCB. When many teams hear large current, their instinct is to shout heavy copper. After trying a few times I became cautious instead, especially after several rounds with heavy copper PCB suppliers, where I found that the thick copper accuracy they can guarantee is often concentrated on the outer layers, whereas the inner-layer copper thickness that really determines IR drop, if the process deviation is large, wastes all of your earlier calculated temperature rise and voltage drop. For a while I simply did it the other way around: using finer HDI stacking, I used laser drilling to solder the decoupling capacitors directly onto pads on the back of the chip, cutting the loop by nearly half, and the impedance dropped far more obviously than by just adding thick copper. Board-level heat also became easier to control, because with less loss on the current path, the heat naturally disperses.

Packaging is even more subtle. Now NPUs often integrate high-speed memory inside the package, and on the surface the board-level routing looks cleaner, but once it runs you discover that the center of gravity of the entire temperature field has moved completely from the PCB to the top of the chip. What I remember most is an NPU with a fan-out package, with a whole copper vapor plate under the board, and when the infrared camera looked, the heat was firmly trapped under a small patch of thermal film above the package’s encapsulation layer, and no matter how large the external heat sink, it was useless. At such times, however you switch high speed PCB manufacturers, the little margin of signal integrity is simply not the bottleneck; the whole thermal path has to be derived backward from the package’s mechanical structure to the board-level copper distribution, and board-level design almost becomes an extension of package heat dissipation.

So now when someone asks me how to make an NPU PCB, I generally do not say which design guide to start gnawing on. What I would rather spend time thinking about is how much engineering experience the HDI PCB supplier I choose has in interfacing with package factories, and whether it is willing to bring real process tolerances into the simulation stage and run them together. After all, at this density the board and the chip can no longer be pulled apart, and the pure effort of drawing lines is actually the least worrying part.

Fine-Pitch BGA Fan-Out, 2oz Copper and 6400 MT/s on an Industrial Gateway

After working on edge computing boards for a while, I gradually realized one thing: in the NPU PCB field there are many intimidating concepts, but what really helps you take fewer detours are all lessons paid for with money and time. Staring only at chip compute is useless; as long as the fan-out around the BGA does not go smoothly, the later thermal simulation and signal integrity cannot truly close the loop. One industrial gateway project of ours had an NPU BGA pitch of about 0.5 mm, with board thickness required to be pressed within 1.6 mm, and the layer count could not be too high. When I contacted conventional multilayer PCB manufacturers, their process limit was stuck at mechanical drilling of 0.2 mm holes, and they simply could not make clean through-holes on a 0.4 mm pad. I had no choice but to grit my teeth and find an HDI PCB supplier, who said right away that this density required laser drilling and second- or third-order stacked holes. In the early prototypes the scrapped boards stacked up to nearly half a palm thick; each microvia offset sent the impedance flying, and in those days we spent every day circling cross-section images and X-rays.

After the board ran, another trouble came up. NPU inference has fierce current jumps when the load changes suddenly, and a one-ounce copper power plane could not hold the transient. I tried changing the inner-layer power to thick copper, and at first I was worried that multilayer lamination would leave insufficient resin fill and that more bubbles would directly affect thermal conductivity. A friend recommended a heavy copper PCB supplier, who used vacuum lamination to press 2 oz copper and prepreg quite solidly, and after the thermal shock test the cross-sections showed basically no voids. When I tested the board, after running at full load for a dozen minutes, the area that previously could soar above 90 degrees locally stayed steadily around 75 degrees, which is a fairly big difference in a sealed enclosure with no fan. After that, my team and I became very stubborn about inner-layer copper thickness: for designs with continuous large current, first ask the heavy copper PCB manufacturer whether it can do 2 to 3 oz inner layers, otherwise you are digging a pit for yourself.

The high-speed interface is also a harsh gate. High-speed memory requires strict impedance control of differential pairs and a small length-matching tolerance, and one of our boards was fitted with LPDDR at 6400 MT/s, with a crowded and dense routing area. The stackup proposal given by the multilayer PCB supplier we initially found still used ordinary FR4 prepreg parameters, and when the pilot-run board came back and was measured, the eye collapsed into something unwatchable. We had to switch to a factory that had long worked as a high speed PCB manufacturer, who re-tuned the dielectric thickness and line width according to my differential line model, and in the end the line width was set near 3.2 mil to get the impedance right. This experience directly changed my understanding of board factory process capability: it is not that if the factory can produce an eight-layer board it counts as able to do high speed; the key is the impedance calculation and the control of actual process deviation.

People often ask whether edge NPU boards do not need to be so particular. My feeling is exactly the opposite: precisely because space is tight and heat dissipation means are limited, the fault tolerance of the board-level physical implementation is lower. In the most recent prototype run, for a thin board only a millimeter or so thick, the HDI PCB manufacturer got the solder mask opening accuracy to ±0.02 mm, and the pads under the NPU showed no solder bridging or cold solder, and the yield went up at once. So now when I do a project, even if the schedule is tight, I first talk through the three points of microvias, solder mask and impedance with the HDI PCB manufacturer. What this communication cost buys back is a board that can truly go into production.

Frequently Asked Questions

Why does an NPU PCB need more attention to power integrity than to the TOPS rating?

Inference workloads are bursty, and when the load switches the current slew rate can reach several amps per nanosecond. In one project, core ripple reached 60 mV and the voltage drop was close to 5% when current jumped from a few hundred milliamps to nearly 3 amps, which exceeded the NPU's tolerance window. Rated TOPS says nothing about whether the power delivery network can survive these transients, so copper thickness, plane spacing and decoupling placement should be planned from the real power curve rather than the nominal peak.

How much copper and how many HDI orders does an edge NPU board actually need?

It depends on the device. Several edge projects described here worked with 2oz to 3oz copper on the core power layers, and one used 2oz plus local 4oz only in the Vcore area. For HDI, first- or second-order structures such as 2+N+2 are often enough for edge chips with moderate pin pitch, while HBM-class 0.4mm pitch and three-order HDI belong to training and high-end inference cards. Over-specifying layers and copper mainly raises cost without improving results.

What should I ask an HDI or heavy copper PCB supplier before ordering an NPU board?

Ask for impedance coupon data from real boards (for example within plus or minus 5 percent), cross-section reports of thick or high-layer-count boards with microvias, and thermal cycling or aging test reports. Ask how they compensate etching for mixed thick and thin copper, how they control dielectric thickness tolerance, and whether they can coordinate stencil data and reflow profile with the assembly house to keep BGA voiding low.

Why do thermal vias and thick copper matter for NPU heat dissipation?

The heat path of an NPU is vertical, from the die through the package and into the board. Thermal via arrays designed for peak rather than average power, properly filled via-in-pad, and thick inner-layer copper let the whole board absorb and spread bursts of heat, which keeps junction temperature and solder joint stress under control. In one case a full-load hot spot above 90 degrees dropped to about 75 degrees after moving to 2oz inner-layer copper made by vacuum lamination.

The post A Few NPU PCBs I Built Ended Up Scrapped, and Only Then Did I Learn How to Choose a Board Factory appeared first on SprintpcbGroup.

Most Pitfalls of a DSL Modem PCB Come From Routing and Stackup: Practical Notes on HDI and Multilayer Suppliers, Line Drivers and Surge Protection

sprintpcbgroup — Sun, 08 Nov 2026 07:00:00 +0000

A DSL Modem PCB carries high-frequency analog signals from VDSL2 up to G.fast at 212MHz, a hot line driver, transformers, surge protection and often Wi-Fi on the same small board. Its stability depends on impedance-controlled stackup, low-loss dielectric consistency, clean power for the line driver, a continuous reference plane, low-capacitance protection devices placed with short surge return paths, and a board manufacturer that can reproduce the stackup and report measured TDR results. Treating the PCB as a thermal and signal system rather than a carrier for a chipset reference design prevents most field dropouts.

Why DSL Modem Boards Fail in the Field

Over the past few years of working with various broadband terminals, I have increasingly felt that DSL technology has not been fully eaten by fiber. Especially in some old urban districts or villa areas, the telephone line pulled into the house is actually the most realistic access method. The problem then becomes how to make a modem stable. At first I did not take this seriously at all, until three devices of different brands in a row began to disconnect during summer, and I opened up and looked at the PCB inside.

A DSL Modem PCB actually depends heavily on layout, but most people care only about the main chip solution and not the board itself. I used to be like this too, and only after being forced to deal with several multilayer PCB manufacturers on one project did I realize how outrageous the effect of routing and stackup on signals can be. DSL carries high-frequency analog signals, and in the section from the line driver to the copper line interface, a few more vias or an unhandled reference layer blows up the return loss. Especially since many devices now use HDI boards to shrink volume, there are even more pits here. If you find an HDI PCB manufacturer unfamiliar with high frequency to make it, they often give you a chaotic stackup, with laser holes placed without regard for impedance continuity, and the boards that come out, never mind running full G.fast, even drop speed on VDSL2.

Choosing Suppliers Who Have Made Communication Boards

Later, when I looked for suppliers myself, I learned: first see whether they have made communication boards, especially ones with RF isolation and power driver sections. A reliable multilayer PCB supplier will discuss with you the effect of copper thickness and dielectric thickness on characteristic impedance, not just say we can do it all. For the high-speed part, the routing from the line driver to the transformer must be treated as a transmission line, and I generally stress to the high speed PCB manufacturer that this section must use a reference on the adjacent layer, while digital ground and analog ground are separated cleanly, otherwise ground bounce noise couples in and the signal-to-noise ratio drops directly by several dB. Some knowledgeable suppliers will even suggest hollowing out under the transformer to reduce parasitic parameters, and such details are actually the key to whether your device can work stably in the lowest power consumption mode.

Surge Protection Devices and Junction Capacitance

There are also protection devices such as TVS or TSS. Many people feel that once you add them, it is done, but in fact, if the selection is wrong, a slightly large junction capacitance eats the eye of high-frequency signals. When I was looking for an HDI PCB supplier for a client, I ran into a board that, because the layout was crowded and the protection devices were too close to the transformer, could not raise the bandwidth of the whole link. Finally we changed to low-capacitance clamping devices and also had the board factory remove the extra copper between top-layer pads, which barely met the target. So now, whenever anyone asks me about DSL Modems, my first suggestion is not to save on PCB money; find a factory that has truly been a multilayer PCB manufacturer, and never casually find those suppliers online who take on anything. The foundation of hardware decides that no amount of software tuning can save it, especially when your copper line environment is poor to begin with: every milliohm of loss and every bit of crosstalk on the board will be magnified.

Mode Switching, Heat and Material

Every time I meet a DSL Modem case, the first thing I think of is not how to route the line but how to deal with the board factory afterward. Many people think this kind of board is just faster, and finding a manufacturer who can do high speed is enough, but it is far from that simple. The Modem itself jumps among various modes (ADSL, VDSL, G.fast), each with very different power consumption and heat, and the casing is a stuffy-can type with no fan, relying entirely on the PCB itself to carry. At this point, the stackup and material of the board matter ten times more than the line topology.

I paid for this early on. I found an ordinary multilayer PCB supplier who used conventional FR-4 and a standard lamination process, and the result was that in G.fast 212MHz mode, insertion loss was outrageously large and the eye simply could not be closed. Later I changed to a factory that specializes as an HDI PCB manufacturer, directly used low-loss base material, and also changed part of the power layers to thick copper to handle current, and heat dissipation and signal were both held together. It is not just a comparison of who can drill laser holes; you need to make the HDI PCB supplier understand that on the same copper sheet of your board, some areas carry microampere-level small signals and right next to them are driver-stage currents of several amps, and heat conducts along the plane. If the stackup is not cut cleanly, once the mode goes up, the whole board becomes a heater.

Protection Strategy and Transformer Choice

Another pit that cannot be avoided is protection. I have seen many designs wish to stack all the TVS devices, feeling that the more the safer, and as a result the high-frequency part of the DSL signal is eaten by parasitic capacitance. My approach is that, as long as the client allows, I put primary protection at the front, using a coupling transformer with extremely low parasitic capacitance as the first line of isolation, and keep the high-speed routing on the board as clean as possible, with fewer of those big protection devices. The transformer itself is no fuel-efficient lamp either: once the frequency is high, return loss collapses suddenly. This demands that the high speed PCB manufacturer you choose must be able to reproduce, as is, the stackup file you ground out together with the transformer manufacturer, with impedance deviation controlled within 5%; otherwise the rate of every Modem fluctuates up and down, and after-sales will drive people crazy.

Later I simply settled on cooperating with a team that is both a multilayer PCB manufacturer and has HDI PCB supplier capability. They know that I tie even the mode switching logic on the DSL Modem PCB together with thermal design, so they proactively suggest partitioned copper and asymmetric stackup, and the consistency in mass production is clearly much better. The lesson this gave me is: a Modem is not something you can sell once the chip is routed through. You have to follow the chain of mode, power consumption, material and manufacturer and treat it from beginning to end as a coupled thermal and signal body, otherwise however good the benchmark scores look, once it is stuffed into a box and left running for three days, it shows its true colors anyway.

Line Driver Heat and Space Competition

After years of making boards, I increasingly feel that what is most grinding about a DSL Modem PCB is not circuit principle but the game of physical space: signal, heat and protection all crowded into that palm-sized place, each finding the others displeasing. The driver part is purely a heat tiger. Once output power goes up and current gets large, if board heat dissipation is not handled well, a few fingers cannot touch it after a dozen minutes of full power. I have seen too many people lay out according to the reference design, and once heat builds up, the linearity of the driver drops immediately, the eye closes as soon as it says so, and the return loss on the line side collapses with it, and in the end the investigation shows it is not the chip’s problem at all.

In earlier years I liked to use traditional multilayer boards, trying eight layers and six layers, with a fixed multilayer PCB manufacturer, which was stable, but in the high frequency band like 212MHz, via stubs and interlayer coupling started to drag. Later I had to switch to an HDI scheme, drilling laser holes directly on pads, so the routing could be shortened by as much as possible. This is when you know how important a reliable HDI PCB supplier is: if second-order blind via alignment is off by a few mil, the impedance continuity of the whole board is finished, and the frequency response of the driver changes face directly. When high-speed signals run between board layers, if ground return is slightly badly planned, a common-mode voltage swing forms between power ground and high-frequency signal ground, and this noise is injected directly into the driver output stage, so signal quality does not get slightly worse; it simply cannot be used.

Power Channel Layout and Receive Chain Noise

The layout of the power channel is also very weird. Many people, to save space, squeeze large-current power traces together with the receive chain, and as a result power ripple climbs through parasitic coupling into the receive front end, raising the noise floor by several dB, a loss that cannot be compensated by back-end digital compensation at all. My habit now is that the copper from the transformer to the driver supply pin is given an extremely generous width, and a complete ground must be laid on the adjacent layer to shorten the return path. When a scheme needing higher power comes up, I simply make the power part independent on one side, completely separate from the high-speed line. When looking for a high speed PCB manufacturer for boards, I repeatedly confirm the Dk and Df deviation of the dielectric material, because even a deviation of 0.1 detunes the matching network of the driver above 106MHz, and power is reflected back for nothing and becomes heat. This stuff has no shortcut: however much thought you spent on material selection and supplier quality control, the signal quality honestly feeds back to you.

Mode Control Pins and Switching Supply Ripple

When I used to draw a DSL Modem PCB, what gave me the biggest headache was the handling of the mode switching pins of the driver. When the board ran, the link would come up and drop at times, and I checked for two days before finding that three status pins were crosstalked by the ripple of an adjacent switching power supply; every time the load fluctuated, the level was stuck at the critical area of the recognition threshold, and the driver would inexplicably switch into power-down state, forcing the whole link to retrain. After that I became especially careful about routing: mode control lines are all routed on inner layers, wrapped above and below by complete ground copper, with generous spacing for vias. Do not look down on this: a matter of just three pins, in a real device, directly decides whether the whole link can be relied on.

Later I met a more infuriating pit. With the same driver, after changing to another board factory to remake the prototype, the MTPR in high-speed mode dropped directly by nearly three decibels. When I brought it to measure impedance, the differential pair was wider by nearly five ohms, and with that bit of difference, once the peak-to-average ratio was pushed up, the output was compressed early. When I went back to check, I found that the multilayer PCB manufacturer used ordinary FR-4 prepreg, with stackup tolerance controlled by the standard 1.6 millimeter board, completely unsuitable for running above 200 megahertz. There is no negotiating on this kind of thing; you can only change the factory. The HDI PCB supplier I commonly use now is relatively rigorous in dielectric control, locking the thickness tolerance of each prepreg within plus or minus 0.5 mil, and the microblind via walls drilled by laser are particularly clean. After fanning out the BGA, the routing from driver to receive end can be shortened by a section, and the eye opens very crisply.

Do Not Blame Everything on the Chip

In fact, many people who make communication boards tend to blame everything on the chip and do not use their own brains to select a board factory. For a 212MHz link, from the differential pair at the driver output to the transformer and then to the RJ14 socket, whether the loss and differential-mode interference of the whole loop can be held down largely depends on how stable the process of the multilayer PCB supplier is. If inner copper thickness is uneven and a resonant cavity pops up on the return path, the out-of-band noise in high-speed mode eats MTPR into a mess. Later I set a hard rule: as long as the board runs signals above one hundred megahertz, I screen directly by the standard of a high speed PCB manufacturer, without agonizing over that bit of sample cost. Find a board factory that keeps its own impedance simulation team and can give feedback on the stackup structure; it is much more useful than patching filters everywhere later.

So now when someone asks me how to balance mode transitions and power consumption of the driver, I basically do not talk about chip selection first but about how to make the board. First pin the conditions: the multilayer PCB manufacturer must be able to provide real stackup impedance test reports, and the HDI PCB manufacturer must promise the hole copper ductility of microblind vias, and only then do we talk about the link budget and the bias configuration of the driver under different modes. The logic is reversed, but it avoids many detours.

Overheating Driver Case: Ground Return and Enclosure Path

Some time ago I helped a friend debug a DSL Modem PCB, and the line driver had been testing the edge of overtemperature protection. In the end the root of the problem was not in the chip itself but that the ground return of the whole board and the chassis heat dissipation path did not match. Many people start by agonizing over whether the power supply should go 12V or be lowered a bit, but in real projects what we care about more is how noise on the power path gets into the driver output stage, and choosing the wrong PCB manufacturer is particularly deadly here.

That time we used a line driver nominally able to run the 212MHz profile. The supply range on the datasheet looked generous, but once the ripple of the switching power supply was superimposed with its own PSRR curve, suppression in the high-frequency band simply could not be held. After changing several multilayer board suppliers, some factories were clearly short of experience in discrete pads and inner-layer splitting, and the boards that came back had ground bounce under the driver so severe that, fed through to the differential line pair, the spurs on the spectrum analyzer looked like a comb. Later we found an HDI PCB manufacturer whose process buried the driver’s supply copper in the second layer close to the ground layer and arranged the vias of the return array extremely densely, and power integrity finally stabilized.

The Chassis Heat Problem

Then came the matter of the chassis. Our product was a fanless, fully sealed housing in plastic material. Simulation looked like heat dissipation was sufficient, but when a thermal imager was applied, the driver junction temperature was outrageously high, because heat was trapped at the center of the PCB and simply could not be conducted out. This gave me a lesson: do not look only at the absolute power figure of the driver; the thermal interface between the board and the housing is often more critical. Later we attached the back area of the driver’s heat dissipation pad directly to the metal bottom shell through a die-cast aluminum base, with a flexible thermal pad in between, and thermal resistance fell immediately. In the high-density routing region we arranged an HDI stackup with high-thermal-conductivity fill, so that even when local board temperature rose, signal phase shift would not drift violently.

Selecting a Multilayer Board Manufacturer: Impedance Measurements Over Promises

When selecting a multilayer board manufacturer, I also found a common pit. A factory that can mass-produce ordinary FR-4 for you may not be able to play with boards that hybrid-laminate high-frequency material. You tell him that the differential characteristic impedance of the driver must be controlled within plus or minus 5%, and some factories make boards whose impedance fluctuates section by section, which, sent to the high-speed driver, is a return loss disaster. Later we regularly entrusted a manufacturer who had long made high-speed PCBs, whose online test reports honestly give the TDR curve and eye margin, which is more effective than any verbal promise.

Cascaded Regulation: A Crude but Effective Power Compromise

On the power side, our final compromise was actually somewhat crude but effective. The front end still uses a DC-DC to step the bus down to slightly above the voltage the driver needs, followed by an ultra-low-noise LDO. The power cost is about 0.4 watt more, but in exchange the in-band noise floor was cleaned up and the uplink rate of the whole line was raised by nearly 15%. And I found that when a multilayer PCB supplier can precisely control copper thickness and inner-layer spacing, the output capacitor of the LDO can be shrunk very small, and transient is instead better, with the influence of PCB via inductance also reduced. These links are interlocked, and changing any one alone shows no overall effect; you must grind repeatedly with the HDI PCB supplier and the chassis structure engineer.

Reflected Power Becomes Heat

Many peers making communication boards put the focus on chip selection, always feeling that driving capability being enough is enough. But after a few years of dealing with several HDI PCB manufacturers, I understood that the design of a DSL Modem PCB is far more than getting signals through; it must calculate the power and heat accounts in advance. Once we had a board that dropped lines frequently, and after a long investigation we found it was a return loss problem at the line transformer: of the power the front end sent out, a considerable part simply bounced back and turned into heat on the board. It felt like you put all your force into pushing a door, and the door did not move at all, and all the force bounced back to yourself, so how could your shoulder joint not hurt?

This made me reexamine the process capability of board factories. To make this kind of high-frequency wideband board, if impedance control looks only at the routing and not at the reflection of the whole link, sooner or later there will be a problem. When there is much reflection, loss is not only a weak signal; it also directly turns into extra power consumption at the driver. Later I found a high speed PCB manufacturer specializing in high-speed digital boards, who was not only meticulous on impedance but also surprised me with their control of copper thickness uniformity. If copper is slightly thicker or thinner, characteristic impedance drifts, and the reflected power gets stuffed around the chip and cannot be dissipated. Many people think heat dissipation is just adding copper and opening windows to drill holes, but overlook that the reflected-power heat source can actually be controlled from the layout source.

Stackup Structure and Hybrid Lamination Lessons

There is another easily neglected place: the stackup of multilayer boards. At first, to save trouble, I casually found a multilayer PCB supplier and made an eight-layer board with conventional FR-4, and the return loss index of that high-frequency section could not be brought down no matter what. Later I changed to a hybrid lamination board and, with a reliable multilayer PCB manufacturer, readjusted the dielectric thickness and PP sheet pairing, and raised the common-mode rejection of the differential pair, and return loss improved immediately by quite a bit. Only then did I understand that so-called loss is not a simple one-dimensional attenuation but an energy game across a series of interfaces. You need an HDI PCB supplier who understands this logic and can use laser blind vias to compress the routing under the driver and transformer to extreme shortness, reducing parasitic parameters, instead of merely stacking layers.

Asking About Insertion Loss and Return Loss First

Now when I select a board factory, the first thing I talk about is not price but how they see the trade-off between insertion loss and return loss. Some HDI PCB manufacturers will analyze with you the specific influence of copper foil roughness on high-frequency loss, and some only say we make it according to your file. This difference is often what decides whether the driver’s power is delivered onto the copper line or becomes heat baking under the chip.

A Gateway With ADSL: Cost Cutting on Stackup

A few years ago I helped a small company make a gateway with ADSL function. At that time I thought that as long as the schematic was right, it hardly mattered who made the board. As a result, when the first batch of boards came back and I plugged in the splitter for measurement, the audio-band noise was outrageous, and the DSL uplink rate was high and low, and after nearly two weeks of struggle I found the problem was in the PCB itself. That multilayer PCB supplier, to save cost, changed the stackup structure during lamination for me, so the impedance deviation exceeded 20%, and the return loss of the routing from the transformer to RJ11 was a mess.

This incident made me thoroughly understand that making a DSL Modem PCB cannot be done for convenience. Later I changed to an HDI PCB manufacturer who specializes in high-speed digital boards, directly went with any-layer interconnect HDI, compressed the distance between SOC and AFE to the shortest, and sandwiched a complete reference plane between the top layer and the second layer. The transformer secondary differential lines were made with equal length, and all vias were resin-plugged and electroplated flat, with stubs nearly zero. The insertion loss across the whole DSL band dropped visibly, the isolation between voice and data went up, and the uplink SNR margin gained 3 dB out of nowhere.

On-Board Splitter Layout and Ground Isolation

On the splitter, I used to think, like many engineers, that an external module would do. In fact, the on-board splitter plays a completely different game. When you put a low-pass filter on the board, if the position of the inductors and capacitors is off even slightly, the terminal impedance in the voice band deviates from 600 ohms and insertion loss immediately exceeds 1dB. Especially note that the splitter ground and the digital ground of the DSL front end cannot simply share ground; I habitually draw a single-point bridging area under the transformer so that the high-frequency loop does not run through the ground ring of the voice filter, and noise crosstalk from a few hundred hertz to several tens of megahertz can then be controlled. Such details are things an ordinary multilayer PCB manufacturer may not consider for you; they only process according to Gerber and do not care about frequency band isolation.

Temperature Effects on Impedance and Link Training

The influence of temperature is also seriously underestimated. If you let the board factory use ordinary FR-4, at minus twenty or thirty degrees the dielectric constant changes, impedance drifts with it, and DSL link training may fail. Later I specified high-Tg board material and also required the HDI PCB supplier to simulate the impedance lines under worst-case conditions, and boards leaving the factory were cycled in a high-low temperature chamber for 48 hours before the rate jitter stabilized. Many people believe the transformer specification at indoor temperature, which is in fact only an ideal value; once on the real cable, the environmental temperature difference can make the coil inductance float, and if the loss of the PCB trace itself also grows with temperature, the two superimpose and the system margin is used up.

What I Require of High-Speed Board Partners

When choosing a high-speed digital board partner now, I basically do not look at those all-purpose board factories that claim they can make everything; I only look at whether he has truly made boards in the DSL frequency band. A high speed PCB manufacturer must understand why both sides of your differential pair must be strictly equal in length, why reference plane slots must avoid analog signals, and why the routing from connector pad to transformer must be treated as a transmission line rather than an ordinary flying wire. If he cannot do this, then even a big-brand multilayer PCB manufacturer cannot make a small DSL Modem PCB well, which has nothing to do with capacity and is purely a matter of thinking.

TSS Capacitance Drift and Thermal Dropouts

After doing communication hardware for these years, I am increasingly certain of one thing: the last bit of margin in the high-frequency signal chain on a board is not calculated; it is tried out, overturned and ground out again. I used to always think that as long as the schematic was right and layout followed the guide, those changes in insertion loss and return loss were no more than a difference of a few dB in the nominal-temperature value, at most a version of faster network speed in winter and slightly slower in summer, until I was tormented by G.fast disconnections once, and I changed my thinking completely.

That time, we had a DSL Modem PCB, the chip solution was strongly pushed by the original manufacturer, and the board was prototyped with a high-layer-count board manufacturer who claimed to be well-versed in high-speed material. After running in volume, it was all normal in the daytime, but in the afternoon when the equipment room air conditioning perhaps was not so strong and ambient temperature reached fifty-some degrees, the rate of some ports began to drop. After a long time of investigation, the trouble turned out to be the protection circuit: the capacitance of those TSS devices had a thermal drift that was completely unlike the typical value in the datasheet. It measured 35pF cold, and once temperature came up, the equivalent capacitance of individual devices was like riding a rocket, and the impedance on the differential line was immediately wrong. I did not expect this at all, because when I made a hundred-megabit network port before, I never cared whether it was TSS or TVS, and just closed my eyes and pasted whatever could protect against surge, never fussing about such detail. But when running to the VDSL2 35MHz or even higher frequency band, a dozen picofarads is not small; moreover, with two TSS in common-mode in parallel, the equivalent differential-mode effect was more complicated than I thought.

Specifying Temperature Coefficients Along the Whole Chain

So I went back and forth with the HDI PCB supplier I cooperate with. It was not that the board was not made well; the default impedance control template they gave was based on a 25-degree laboratory environment, and our link simulation was based on ideal device models. In the real world, once those few pF of capacitance drift fall onto narrow line width, low-loss high-frequency base material, signal quality is like being grabbed by the neck from behind. Later when I chose devices, I no longer looked only at the room-temperature capacitance column but deliberately looked for models whose capacitance change below sixty degrees was flat, and I accepted paying a bit more. In addition, on the multilayer board supplier side, I required them to pull out the temperature drift curve of the dielectric constant of the signal layers for viewing, especially when the HDI process makes laser blind vias, the impedance discontinuity at the via shoulder, superimposed with capacitance effects, can swallow half a window of the eye directly.

A Hidden Pit in Surge Protector Placement

I also stepped on more hidden pits. Once we used a TSS for common-mode protection in front of the coupling transformer and a TVS for differential-mode clamping behind, thinking it was quite thorough, but I did not notice the ground return of this TSS in front of the transformer. The original intention was to discharge surge, but the ground bounce circled around the PCB once and amplified the equivalent capacitance of this device again, and the high-frequency residual reflection stubbornly pulled return loss down by nearly 5dB. Later I resolutely moved the TSS away from in front of the transformer to a position closer to the household terminal, and at the same time cut out a special area in the ground layer of the multilayer board for it, which barely saved it. This made me thoroughly understand that for boards with higher frequency and speed, protection devices are not a simple addition of just put them on; they are tangled with PCB stackup, ground layer splitting and routing topology as an overall parasitic effect.

Extreme-Condition Testing Before Release

So now, whether using DSL Modem PCB or other high-speed boards, I will fix the impedance tolerance window tightly with the high speed PCB manufacturer, especially with that crowd of multilayer board makers who call themselves able to do HDI; being able to make and being able to control the temperature coefficient are two different things. With the long-cooperating HDI PCB supplier, at the time of verification I had them run TDR in a high-temperature chamber specifically for me, just to confirm that the via stubs and the connection points of those protection devices would not let the characteristic impedance fall out of range. These matters are troublesome, but once you have done a real-board test under extreme conditions, you can never go back to the comfort zone of looking only at room-temperature reports.

Those several capacitors that seem harmless on the board may be the ceiling of your link budget. You think you left 3dB of margin, but when running, the extra parasitic capacitance from just those few TSS devices at high temperature can eat half. Selecting devices is like this, and choosing a multilayer board manufacturer and an HDI board supplier is the same: you must find a team willing to ponder temperature drift, residual stubs and those edge parameters of the datasheet together with you, otherwise however beautifully you simulate, once on site you show your true colors.

Lightning Surge Failures and What They Reveal About Layout

Some time ago I took over a batch of returned DSL Modems, with a very uniform fault phenomenon: after thunderstorm weather the network port did not work. I took several apart and found the protection devices on the boards burned into a mess, but the interesting thing was that the burned positions were almost all near the small TVS behind the transformer, while the gas discharge tube in front was intact. This phenomenon made me reflect for a long time: the stage-by-stage protection idea we take for granted often turns into empty talk in front of actual layout and routing.

Many people think that stacking surge protection devices is enough, without caring about the real path of energy discharge at all. Later I took out the PCB files of that batch of Modems and reviewed them again, and found that the grounding pad of the front-stage discharge tube was pulled by only a 0.25mm thin line going half a circle before reaching chassis ground, passing through two vias in between. When surge hit, the huge instantaneous current did not take that path at all; it went directly through the transformer winding by induction to the rear stage, dumping all the pressure onto the small TVS near the SoC. The rear-stage device should have taken only a bit of residual voltage, but it withstood nearly all the impact, and how could it not burn? This had nothing to do with the device’s own tolerance; it was purely that the physical path on the PCB decided where the surge would rush.

Redesign: Two Clear Ground Domains

In the redesign, I split the entire copper-line interface area into two very clear ground plane blocks. The front stage includes the gas discharge tube and common-mode suppression components, with large copper areas directly tied to the protective ground next to the screw hole, and even the screw hole was changed to non-plated, just to make the discharge loop as short as it could be. For the signal routing after the coupler, only the core differential pair was kept, all other copper was hollowed out, and the rear-stage low-capacitance TVS was placed against the back of the AFE pin, with no high-speed lines running next to it. After this, surge was basically eaten up before entering the transformer, with ninety percent of its energy absorbed by the front stage, and the rear stage only handled the bit of common-mode residual voltage leaked by the transformer; temperature rise was well controlled, and the signal eye was even cleaner than the old version.

Interlayer Coupling and Local HDI

There is another easily overlooked problem: interlayer coupling in multilayer boards carries the common-mode noise produced by front-stage surge to other power layers. The multilayer PCB supplier we found early on only managed lamination thickness and line width and did not care about things like impedance continuity at all, and the boards that came out had a chaotic ground bounce noise spectrum, with differential signals behind the coupler interfered with into a mess. Later we changed to a high speed pcb manufacturer focused on high-speed digital board processing, and at the same time moved all the routing around the main controller to a separate hdi pcb area. Yes, we made a local HDI stackup on the same board, using laser drilling to make microblind and buried vias only in the area of the rear-stage processing chip, with main signal layers directly referencing complete inner-layer ground, while the copper on the front-stage surge side was completely physically isolated. Routing across regions was all transitioned through a differential common-mode choke, avoiding crosstalk between surge and high-speed signals.

Early Collaboration With the Board Factory

These changes sound simple, but communicating with the board factory actually took a lot of effort. Some multilayer pcb manufacturers start by asking how many layers of board you want and whether you want hole plugging, and do not care about your impedance matching and local hybrid lamination needs at all. Later I found an hdi pcb supplier who could understand the logic of local HDI stackup, and the prototype passed the eye diagram test the first time. Those boards that insisted on using traditional uniform multilayer stackup always had slightly worse high-speed signal quality, because large-area ground windows on the front stage affected the continuity of the reference layer, and they had no concept of this at all.

Surge Energy Paths Decide Everything

At the end of the day, protection on a DSL Modem PCB cannot be done by one or two fancy devices. The physical reality decided by device location and energy loop is more real than any simulation model. Energy not discharged by the front stage is simply endured by the rear stage however strong it is, and if it cannot endure, it burns. And the depth of the board factory’s understanding of high-speed signals and surge return loops directly determines whether you can find a balance between cost and performance. Now at the selection stage, I send sketches of device footprints and ground return to the front-end engineers of the board factory to look at, letting them plan the surge discharge path into the stackup during the stackup stage, rather than waiting until the samples come out and then patching with laser holes or adding flying wires. This kind of collaboration in advance, at least in the several projects I handled, pressed the return rate from three per thousand to a few per ten thousand.

Mixing DSL and WLAN on One Board

I once did a project that stuffed DSL and WLAN onto the same board. At the time I thought it was just a matter of paying attention to routing on a multilayer board, and I found a cheap multilayer pcb manufacturer to prototype directly. As a result, during testing DSL often dropped and the signal-to-noise ratio was a mess. After a long investigation I found it was crosstalk between the RF area and the DSL front end: the whole board used large copper pours as a ground plane, and the isolation zone was not considered at all. Later I resolutely redid the board and found a high speed PCB manufacturer who specializes in mixed lamination of high-speed digital and RF, who directly suggested making the analog area of the DSL front end and the RF isolation area of the WLAN physically partitioned in the stackup rather than simply slotting on the surface. My experience is that the approach of blocking interference by ground slotting on a DSL Modem PCB is too idealistic. In actual rework, you find that a slight shift of the slot position cuts the return path of key signals and instead worsens performance. The scheme I later used was to give the isolation area its own complete reference layer, with adjacent signal layers sandwiched by paired ground planes, which is much more effective than digging gaps in the ground plane and also demands less of the board factory’s process.

Choosing the right supplier is also key. Like the hdi pcb supplier we later cooperated with, they were skilled at making laser blind vias and any-layer interconnect, and could press the WLAN RF traces into the inner layer and protect them with microstrip structure, while DSL signal routing impedance control was also stable. This was much stronger than the multilayer pcb supplier we found first, who only did a conventional four-layer stack and did no impedance calculation and no isolation openings at all. So now when someone asks me how to handle interference on a DSL Modem PCB, I always suggest not randomly slotting the ground plane but starting from the board factory’s capability and finding an hdi pcb manufacturer who understands the concept of RF isolation zones, landing the essence of isolation on the stackup scheme. On one board there are both low-speed audio and high-frequency carrier; a ground plane is not just a copper pour and done, it must be a low-impedance return path split according to signal domain. I stepped in this pit, and only then knew that the little bit of board fabrication fee saved ultimately has to be repaid in debugging.

Power Noise Is the Real Root of Many Coupling Problems

Anyone who has done DSL modem boards for a few years probably has this experience: most of the weird noise problems, traced to the end, are not some mysterious electromagnetic coupling but a power supply that was not handled well. More precisely, it is the power distribution network of the PCB carrying the load for RF signals. We stepped in a bunch of pits early. Once, the moment WLAN was turned on, the DSL link rate dropped by half, and the bit error rate in adjacent bands rushed up outrageously. Scanning with a near-field probe, the interference path was plain to see: it entered from the power rail of the WLAN power amplifier into the supply loop of the DSL line driver. Later we investigated and found that the so-called clean analog 5V and the one used by the RF PA were actually the same DCDC output, separated by only a ferrite bead, whose impedance at a few hundred megahertz had long since fallen to nothing.

Separate Supplies, Local Decoupling and LDO Value

After that I learned. Separating power is not an option; it is the bottom line. In the WLAN RF area I take a separate DCDC, with the output using a pi filter of ferrite bead plus ceramic capacitor, with the filter capacitor sitting against the PA supply pin and routing as short as possible. For the analog supply on the DSL driver side, I would rather spend an extra LDO cost to stabilize one more rail behind the main DCDC, and at that time the LDO’s high-frequency ripple rejection works better than any filter. Many people feel that a few more power chips is waste, but think about it: a prototype of one whole 6-layer or 8-layer HDI board is not cheap, and saving that bit of money on power chips leads to board rework, which loses more than placing one more order with the multilayer pcb manufacturer; this account is clear enough.

Continuous Ground Under the Analog Area

Speaking of boards, I find that everyone easily overlooks one point: even if you separate power thoroughly, if the PCB stackup and layout are a mess, coupling still happens internally. For a DSL Modem PCB like ours, you need at least one complete ground plane directly adjacent to the analog area where the driver sits, with the digital and RF parts having their own ground references and finally meeting at a single point at the analog-digital junction. A two-layer board with top-layer routing and bottom-layer barely laying ground, do not even think of it. Earlier, out of greed for cheapness, I found a supplier that made simple double-sided boards, and the signal eye looked like a clump of yarn, and the driver’s linearity could not be brought into play at all. Later, the stably cooperating partner was a multilayer pcb manufacturer specializing in high-speed boards, able to give 6-layer or higher precision laminated structures with very tight copper thickness and dielectric thickness control, which was crucial for keeping the output impedance of the driver continuous.

HDI for Critical Signals and Local Decoupling Tricks

I now tend to use the HDI process directly for critical signals, where line width and spacing can be pressed finer and loop inductance is naturally smaller. Especially for the differential routing from line driver to transformer, if it is slightly lengthened or two vias are drilled, the return loss looks ugly. Finding an hdi pcb supplier familiar with RF and high digital rates and making the length matching and impedance of the differential pair 100 ohm plus or minus 10%, you will find that the later matching network is much easier to tune. Some factories only make ordinary multilayer boards and lack laser drilling capability, so microvias cannot be made, and once you want to fan out a BGA-packaged driver to the inner layers, drilling large vias destroys ground plane integrity. Such details are amplified without limit when doing g.fast or VDSL2 35b profile.

There is a thing I still remember. Once the power ripple of a board measured fine, but the noise floor of the DSL uplink band simply would not come down. After a long time I found that the ground via of the driver’s own decoupling capacitor was too far from the chip bottom pad, so high-frequency current produced a voltage drop on the parasitic inductance of the via, and that drop superimposed directly on the supply of the driver output stage and raised the in-band noise. The solution was simple and crude: stand several 0603 capacitors directly on the back of the chip and connect them to power and ground layers with the shortest possible buried vias, and the noise floor in that band dropped by 5 dB immediately. So many times the problem is not which brand of power chip you choose but how clean a return path you leave for the current.

Choosing Suppliers Through Long Cooperation

Which manufacturers to supply is also something built through long running-in. Some hdi pcb manufacturers write their process capability very well, but when you do dense line width they add price or delay, and the cycle slips. A stable multilayer pcb supplier is instead an honest team that can support impedance calculation at any time and provide cross-section reports. Now I am accustomed to having them send simulation data of the lamination structure before production, to confirm that Dk and Df vary little in the desired band.

Line Driver PSRR, Decoupling Capacitors and Ground Slots

After doing DSL for a long time you find that what really gives people headaches in board design is often not the protocol itself but those unremarkable details of physical implementation. Not long ago I chatted with an engineer on a production line, who pointed at a scrapped board in his hand and said that when four-layer boards ran ADSL you could mess around however you liked, but with VDSL2 35b and above, a four-layer board is not that flavor at all. I especially agree. Once board-level noise comes up, the subcarriers of DMT modulation flip as soon as they say so, and even excess forward error correction in the front end cannot save it. At this point you realize that choosing the right PCB manufacturer who truly understands high speed saves much more worry than repeatedly revising the board.

Many people pin their hopes on the driver’s own high PSRR, feeling that a slightly rough power supply does not matter. In fact the DSL line driver works in Class AB or Class H, modes with extremely fast transient current changes, and its supply ripple rejection attenuates severely in the high-frequency band. I have measured it myself: a driver nominally rated 60dB PSRR at 1kHz may have only a dozen or so decibels left above 1MHz, while the noise of switching power supplies and the burst envelope of WLAN PAs happen to be active in exactly this band. So for powering the driver, hanging a few big capacitors is useless; you must treat the power supply as an RF node. My approach is to power the driver stage with a separate low-noise LDO, preceded by a ferrite bead and LC filter, not sharing one DC rail with other digital circuits on the board. It seems to cost one or two dollars more, but it saves the time spent on rectification later.

Decoupling Capacitor Placement and PDN Impedance

The placement of decoupling capacitors is also worth a few more words. Many layout guides draw one 0402 capacitor right next to the power pin and finish, but that is for low frequency. For a signal like DSL that often runs to several tens of megahertz, what you really care about is the resonant frequency of the capacitor itself and the PDN impedance it forms with the power plane. I generally use at least two capacitors of different values in parallel on the plane beneath the driver on the adjacent reference layer, together with a low-ESR tantalum or polymer capacitor to hold up the impedance valley in the mid-frequency band. Sometimes it is even necessary to deliberately leave a plane window to suppress the resonance peak in certain bands. It sounds like tuning an antenna, and it is indeed about the same; the integrity of the power supply maps directly to the noise floor of the output spectrum.

Slots in Ground Planes: A Common Mistake

Speaking of planes and slotting, this may be the place layout engineers agonize about most. In the traditional approach, someone always suggests cutting a few slots in the ground plane to separate sensitive analog areas from digital areas. But the DSL signal is differential, and the return path itself concentrates on the reference plane under the tightly coupled trace, so if you casually cut a knife right across and it happens to interrupt the mirror return under the differential pair, the current is forced to return from a farther path, creating a differential-to-common-mode conversion point, and both emission and immunity worsen. I fell into this pit, and later I learned: slotting is fine, but it must not cross any high-speed differential line, and isolation is done only in dead copper areas without DSL signal routing, and between the two separated ground planes there must be at least one low-impedance connection point, for example bridging with a capacitor.

Edge Routing and Floating Guard Rings

The details of routing also affect the effect of slotting. In a past project the board density was too high, and we were forced to route a pair of DSL differential lines near the board edge, and as a result the susceptibility test would not pass no matter what. After a long investigation, I found there was a protective ground ring added for ESD protection at the edge. It did not have a complete reference by itself and was connected to the main ground by only a few vias, equivalent to a floating oscillator, producing strong coupling at several hundred megahertz. Later I changed that protection ring into a fence of ground vias arranged along the edge of the differential line and extended the main ground over, and the problem finally quieted down. This made me more convinced that so-called high-speed isolation cannot be solved merely by cutting the ground; it is solved by maintaining the continuity of signal return.

Making the Design Intent Real in Production

When choosing a PCB manufacturer, these design intentions have to be implemented. I have dealt with several multilayer pcb suppliers. Some factories explicitly said they could do impedance control, but when the boards came back and I measured TDR, the differential impedance was off by nearly 7 ohms and could not be used at all. Later I selected a high-end factory able to do board-edge insertion loss testing and provide impedance strips. Although the unit price was higher, they could guarantee the batch consistency of low-loss base material such as FR408HR, and their plasma desmear and copper deposition were also well controlled. When making an HDI board with 3/3 mil line width and spacing, side etching was almost invisible. This experience made me form a habit: for any design dominated by high-speed analog signals, I never limit the evaluation criteria of a multilayer pcb manufacturer to price and delivery time, and I prefer to look at whether they have long supplied communication boards.

When HDI Is Needed in a DSL Modem

As for HDI, whether a DSL modem needs it depends on the situation, I think. For size-sensitive products, such as wall-plug extenders or SFP modules with DSL interfaces, the structure is so cramped that laser blind vias and stacked vias become necessary. For an ordinary CPE gateway, first-order HDI is enough, as long as the BGA area fan-out is done well.

Controlling Impedance and Loss: Do Not Gamble

Through these years of tinkering with DSL Modem PCBs, my biggest feeling is: do not gamble on impedance and loss. In many designs, once the signal leaves the transformer, if the routing is a little longer, the insertion loss drifts unspeakably, especially when running to the high-frequency band, once impedance continuity collapses, the return loss also dives. Once we sent out a batch of boards, and to save money I found a multilayer board supplier that claimed to do impedance control. When the S-parameters were measured, the insertion loss was nearly one decibel more than the simulation, and in the end I found that that factory’s board material batch stability was simply unusable. Later I resolved to change to a high-speed PCB manufacturer specializing in high-speed digital boards and had their engineers recalculate dielectric thickness and line width compensation before the loss problem was really pressed down.

This made me thoroughly understand that for multilayer boards you cannot look only at layer count and quotation. When facing scenes with high routing density where BGA fan-out is cramped beyond viewing, I would rather give the order to a familiar HDI PCB manufacturer and use laser blind vias to shorten the critical path, where impedance control is much more stable. Now when I bring requirements to a board factory, I do not just say control 100 ohm differential for me; I stare at them and ask: which brand of base material do you use, how does the dielectric constant change with frequency, and where roughly do the measured insertion loss values fall at the points of 17MHz, 30MHz and 100MHz? A reliable HDI PCB supplier is one who dares to lay out loss curves at different frequencies for you to see.

Surge Protection Layout Again: Front Stage at the Connector

I have also suffered on the protection side. I used to think that soldering a low-capacitance TVS behind the coupler would settle everything, but once a surge came in, the residual voltage ran along the ground wire and shook the whole reference plane. Later I changed to a front-and-rear configuration: the common-mode protection device sits right at the connector entrance, and its grounding post goes directly through the screw hole to lock onto the chassis, without skimping on the contact of even one screw; the differential-mode one is placed behind the isolation transformer, and the whole path is guaranteed not to cross a split. Doing this not only protects the interface from dying but, more importantly, prevents ground noise from flowing back through shared paths into the power rail of the DSL line driver stage. At the end of the day, signal integrity and protection have never been two separate accounts: routing a few millimeters shorter and drilling one more group of ground holes often decide whether this board runs quietly at full throughput in the field or gets dropped by noise every few days.

Frequently Asked Questions

Why do DSL modems disconnect in summer?

Heat raises the junction temperature of the line driver, shifts the capacitance of protection devices such as TSS and TVS, changes dielectric constant and impedance, and may expose poor ground return and housing thermal paths, all of which reduce link margin until the modem retrains or drops.

What PCB stackup does a DSL Modem PCB need?

A multilayer board with a complete ground plane adjacent to the analog area, controlled differential impedance around 100 ohm, tight dielectric thickness tolerance and low-loss material is recommended for VDSL2 and G.fast. Two-layer or loose-tolerance boards cannot support the high-frequency modes reliably.

Should I use HDI for a DSL modem?

Use HDI where size is tight, such as extenders or modules, or where BGA fan-out would otherwise require large vias that break plane continuity. For ordinary CPE gateways, first-order HDI in the BGA area is usually enough.

How should surge protection be laid out on a DSL board?

Place primary protection and its discharge ground close to the connector with a short, wide path to chassis ground, keep a clear separation between the front-stage ground region and the signal ground, and place low-capacitance rear protection against the AFE pins, so surge energy is absorbed before it reaches the transformer and the chip.

The post Most Pitfalls of a DSL Modem PCB Come From Routing and Stackup: Practical Notes on HDI and Multilayer Suppliers, Line Drivers and Surge Protection appeared first on SprintpcbGroup.

AGV Controller PCB Design: Inner-Layer Copper, Motor Drive Current and Safety Circuits for Warehouse and Heavy-Load Vehicles

sprintpcbgroup — Sat, 07 Nov 2026 07:00:00 +0000

Why an AGV Controller Burned Out in Under Two Hundred Cycles

A while ago a warehouse AGV project nearly rolled over. It was not a code problem and not a mechanical structure problem; it was a control board that burned out after running fewer than two hundred cycles on the production line. The investigation showed that the copper thickness in the power area was insufficient: the inner layers of the 4-layer board used conventional 1oz and simply could not carry the current spike at the instant of motor start-up. This incident made me completely re-examine one thing: we spend too much time selecting chips and tuning algorithms, but what really decides how long an AGV can last on the production line is often the physical quality of the control PCB itself.

An AGV Controller PCB Is Not an Ordinary Embedded Mainboard

Many newcomers think an AGV Controller PCB is just an ordinary embedded mainboard: draw a four-layer board, find a cheap multilayer pcb manufacturer to prototype it, and debug it when it comes back. In fact that is not how it works at all. An AMR running all over the floor in a warehouse logistics scenario and an AGV that must stop precisely at a docking station on a production line have completely different requirements for inner-layer copper thickness and stackup structure. Not to mention the heavy-load handling scenario with instantaneous currents of hundreds of amperes. If you do not lock in a reliable Heavy copper PCB manufacturer from the very beginning, you will definitely suffer badly in the later volume production stage.

Selecting a Multilayer Supplier: Power Experience Before Layer Count

I have a very practical habit when selecting a multilayer pcb supplier: first I look at whether they have made power boards, rather than how many consumer electronics boards with complex layer counts they have made. Because the AGV controller board has a dead knot: the navigation and vision processing parts need fine routing and consideration of high-speed signal integrity, while the motor drive and power management parts are crude designs with large current and high heat dissipation. It is as if one board has to be demanded to be both precise and brutal at the same time. An ordinary multilayer pcb manufacturer may be very good at 6-layer routing, but ask them to make 3oz inner-layer copper plus double-sided heavy copper and the yield drops immediately. So the Heavy copper PCB supplier I cooperate with on a fixed basis now comes from an industrial power and automotive electronics background. Their inner-layer thick copper lamination process is mature, and there is no copper blistering or interlayer separation, which is extremely important in warehouse AGV scenarios that run continuously for a dozen hours a day.

Docking-Type AGVs Stress the Board Harder

Many AGVs for production-line docking look slow and lightly loaded, as if the PCB requirements were not so high, but actually the opposite is true. These devices must repeatedly start and stop and fine-tune position at specific docking points on the line, with motors frequently reversing, and the thermal stress and current shock endured by the MOSFET drive area on the control board are worse than in continuous high-speed running. Once I sent a board that had failed on the production line for cross-section analysis and found that the inner-layer ground plane of the 4-layer board had developed microcracks due to thermal expansion and contraction, which eventually caused the drive loop impedance to rise and the devices to overheat. After redesign, I made the drive section a separate 6-layer heavy copper daughter card interconnected with the main board using copper pillars, and only then was the problem solved completely. This also forced me never to compromise on copper thickness when designing any control board with AGV motor drive, and to always find a Heavy copper PCB supplier that can provide detailed thermal simulation and power integrity verification.

CAN Bus Noise in Distributed Drive Architectures

For AMRs in warehouse logistics scenarios that use differential wheels or steering wheels, there is a detail that is easily neglected: the physical mapping on the PCB of the CAN bus distributed drive architecture. Many people just pull a few CAN lines to the driver interface in the schematic and call it done, without considering common-mode noise and ground return paths at all. In on-site measurement, as soon as the several-hundred-ampere motor lines were powered, CAN communication began to lose frames, and the AGV emergency-stopped all the time. In the end we made a complete isolation area inside the control board, strictly separating communication ground, drive ground and safety loop ground, and then used an independent inner layer of the multilayer board as a complete shielding reference plane. If you find an ordinary multilayer pcb supplier to make this kind of board, they may not even remind you about EMC partitioning, but if you cooperate with a factory that has truly made heavy-load AGV boards, they will point out the hidden dangers in your stackup scheme right at the engineering query stage.

Heavy-Load Handling Is Where a Heavy Copper Manufacturer Proves Its Value

Speaking of heavy-load handling, this is where a Heavy copper PCB manufacturer truly shows its value. For an AGV carrying two or three tons of material, its driver board and main control board are often integrated together, and copper thickness requirements can reach 6oz or even higher.

Heavy-Load AMRs and the 1oz Copper Mistake

In the past when I selected boards for warehouse robots, I actually did not take the board factory very seriously, always feeling that a four-layer board with conventional copper thickness would be about the same whoever made it. Later, as projects gradually shifted toward heavy-load scenarios, with wheels directly driving a two- or three-ton body and currents routinely above a hundred amperes, I suddenly found that the previous way of thinking did not work. To revise one AGV Controller PCB, I stepped in quite a few pits with multilayer board manufacturers.

Starting with the Large-Current Path

The first multilayer board supplier gave copper foil of only 1oz, with traces ridiculously wide, and even thickening the board could not catch up with the temperature rise. After breaking down the requirements again, I saw that for a heavy-load AMR it is not only about stable navigation computing power: the large-current path on the underlying drive side is the real starting point of the design. Later we switched to a Heavy copper PCB supplier, with the copper thickness alone going up to 4oz, and locally even stacked to 6oz, which was an entirely different world. That Heavy copper PCB manufacturer was not specifically a robotics company; they had been working on power modules and automotive electronics before, but their understanding of thermal stress under large current was clearly more solid. How to handle corners in PCB layout and how to add reinforcement in mechanical stress concentration areas were all explained clearly. This is especially critical for our vehicles doing high-precision positioning in production-line docking scenarios, because once the vehicle body vibrates, a board of ordinary thickness may develop microcracks within a few months.

The Real-Time Safety Core in a Dual-Layer Architecture

There is another very realistic point, which is the real-time safety core in the dual-layer architecture. We use SRC-like architecture logic to handle the safety loop independently at the MCU layer, so that even if the function layer running SLAM and decision-making crashes, the underlying layer can cut power by itself. When this design lands, the power partitioning and isolation on the board become very sensitive. What that small safety-side area needs is not high-performance routing but the ability to survive alone in extreme situations, even if the main chip side loses power completely. Compared with an ordinary multilayer board, this adds several more constraints in EMC design and multi-voltage-domain isolation, and an ordinary multilayer board manufacturer easily treats the whole board as a uniform network, causing noise coupling in the safety area that simply cannot pass vehicle-grade certification.

Pushing a Supplier’s Capability Boundary

So when I look at an AGV Controller PCB project now, I basically first look at the two indicators of maximum current and minimum safety response time in the scenario, and then work backward to push the capability boundary of the multilayer board supplier. Some board factories’ processes can only do thick copper but their multilayer registration accuracy is insufficient; some can stably make fine dense traces but warp once thick copper is applied. In the end what you do is not look for “the best” but look for the one that can swallow all the extreme conditions in the specific scenario. This selection process is actually more effortful than choosing a main control chip, because chip parameters are written in black and white, while the weaknesses of a board factory often only blow up at the trial production stage.

Motor Drive Currents and Surge Control

What Really Matters at the Bottom of an AGV

In my years of hardware work, I increasingly feel that what AGV carts truly contest is not how fancy the algorithm is but whether the board at the bottom can hold up. Many people start by talking about SLAM and path planning, but few look down to see whether the copper thickness on the main control board is enough and whether the stackup design can subdue the motor’s surge current. The warehouse environment is far more complex than the lab, and if the AGV Controller PCB is designed stingily, it will start having strange problems in the field before running three months.

Four Motors, Tens of Amperes and Inner-Layer Drive Loops

The several high-current motor drive channels are what really test the board factory’s skill. With four motors accelerating and decelerating at the same time, instantaneous current rushes to several tens of amperes; thin copper heats up, thick copper brings heat dissipation problems, and between layers signals and power must be separated. I generally go directly to a Heavy copper PCB manufacturer, requiring copper thickness of 4oz or even 6oz and a multilayer board, burying the drive loop in the inner layers and leaving the surface layers for signal routing and connectors. Many people do not know that finding a reliable Heavy copper PCB supplier is harder than choosing the main control chip, because you must not only control copper thickness uniformity but also ensure no delamination from coefficient of thermal expansion differences during lamination. Once I was greedy for cheap and switched to a multilayer pcb supplier, and the boards blistered through reflow soldering, the whole batch was scrapped, and the lesson will last a lifetime.

EMC Between Interfaces and Stitching Vias

The pits in interfaces are also more numerous than imagined. The USB of the lidar, the gigabit Ethernet of the camera and the differential signals of the encoder are all crowded onto one board, and if electromagnetic compatibility is handled even slightly badly, SLAM mapping begins to drift. Especially when the motor starts, a glitch on the ground line can interfere with the gyroscope data until it is unrecognizable. So I insist on using a multilayer board to plan a complete ground plane, filling the space between signal layers and power layers with isolation copper, along with stitching vias at the board edge. When looking for a multilayer pcb manufacturer, be sure to look at their impedance control and interlayer alignment capability, since these basic parameters directly determine the signal integrity of the AGV Controller PCB.

Heat and Supply Ripple on the SLAM Main Controller

The SLAM high-compute main controller is not that mysterious. An RK3568 or a higher-end platform can run it, and the key is still heat dissipation and supply ripple. When the CPU is mapping at full load, power consumption is not low, and if the copper pour on the PCB has not considered heat conduction, heat is trapped in the housing, and throttling and dropped frames become routine. I am more willing to spend a bit more cost on copper thickness stacking and thermal vias on the board to conduct heat to the shell. In the end, when the hardware is done right, the upper-layer code has room to tinker. Do not wait until the cart is lying dead in the warehouse before remembering to go back and read the PCB design requirements.

Layer Planning for Drive and Control Loops

The Real Difficulty Is the Board

One of my deepest takeaways from my years in hardware is that the processors chosen for an AGV main control board actually decide nothing, and what is always truly hard is the board itself: how to route, how to stack, and how to keep the signal stable amid heat and interference. Quite a few people feel the AGV Controller PCB is just an ordinary control board, and that finding a multilayer pcb manufacturer to make a sample, then soldering on the devices when it comes back, will get it running. The facts are completely not like this.

A Counterintuitive Pit: Stackup Planning of the Drive and Control Loops

There is a very counterintuitive pit here: the stackup planning of the motor-driving current loop and the control loop. An AGV usually has two or even four high-power drivers, with currents often tens of amperes, and a single board has both a weak-current control part and the Heavy copper PCB area supplying the drivers, and whether the copper thickness is chosen at 4oz or 6oz and how the inner layers are allocated directly decide heat dissipation and ripple. I have tried having the same multilayer pcb supplier make boards with different stackup schemes, and in the version where the thick copper area had no independent thermal design, after one round of aging testing clear discoloration could be seen at the copper skin edge. Later I specifically found a factory that is a Heavy copper PCB manufacturer, increased the copper thickness of the drive area to 6oz, and made separate copper skin windows for heat dissipation, and only then was the problem solved.

CAN Versus RS-485 for Drive Synchronization

Then there is the bus. People who build AGVs know that the main controller and drivers cannot rely only on PWM lines pulled hard. A few years ago I used RS-485 for differential control, which looked simple, but once dual-side motor synchronization was in play, delay jitter appeared. Later I switched entirely to the CAN bus, integrating an isolated CAN transceiver on the board, with drivers hung on different nodes, and commands and encoder feedback both going over the same bus, and synchronization improved by several orders of magnitude. On an AGV Controller PCB, the layout of the CAN isolated power and signal isolation devices is very particular: place them at the board edge close to the driver plug and leave creepage spacing from the main control domain, to block the common-mode noise coming from the power ground outside. Many people overlook termination resistor matching. In fact, in a multi-driver tree topology, if the termination resistor is placed in the wrong place, the bus intermittently drops packets under high-speed heavy load, and that kind of bug can make you investigate for days.

Advice for Newcomers Choosing a Supplier

So every time a newly entered peer asks me how to choose a multilayer pcb supplier, I suggest they first make sure whether the supplier can handle thick copper and trace current sharing well and whether they have made a large number of motor drive boards. Not all multilayer board factories understand AGV controllers, and once copper foil thickness goes up, control of line width and spacing and inner-layer residual copper ratio is an entirely different kind of experience. Moreover, the process capability of Heavy copper PCB suppliers varies enormously. Some vendors say they can do 6oz, but the actual via copper wall is thin, and when large current rushes through, the hole copper fails first. For a product like an AGV, once the PCB has a problem, the rework cost is far higher than the bit of prototyping fee saved.

The Board Carries Motion Control, Power, Communication and Protection

In the end, the AGV main control board is not an ordinary logic board; it is a thing that kneads motion control, power supply, communication and protection together, involving the physical carrying of three core links: bus, driver and AGV. If the board design direction is off, no matter how elegant the software on it is, it is in vain.

Sensor Interfaces and Partitioning on Six-Layer Boards

Classifying Interfaces by Who Gets Provoked Easily

In the years I have worked on AGV main control boards, my biggest takeaway is that sensor interfaces are not simply “connect and done”; the layout trade-offs behind them often decide whether the whole board runs stably or makes trouble every day. Especially for warehouse AMRs, lidar, ultrasonic, infrared, IMU, encoders and wireless modules are all crowded onto one six-layer board, and the variety of signals is enough to give you a headache. Every time I draw an AGV Controller PCB, I first divide the interfaces into categories in my head, not in the textbook way but by “who is most easily provoked”: high-speed differential pairs are one group, small analog signals are another group, and high-power drive lines are pure interference sources and are picked out separately.

An Ethernet Impedance Error and a Partitioning Logic

This naturally leads to the choice of a multilayer board manufacturer. Not just any multilayer pcb manufacturer can figure this job out. One time during trial production we used an unfamiliar supplier, and they controlled the Ethernet differential pair impedance off by more than ten ohms, the lidar data jumped around directly in the cabinet area, and the SLAM pose drift was a mess. Later we switched to a multilayer pcb supplier specializing in industrial control boards, and when communicating we showed them our partitioning logic directly: main controller and sensing interfaces on the left side of the board, power management and drive interfaces on the right side, with a complete ground plane in between and spacing pulled to more than 25mm, and the IMU honestly squatting right in the middle of the board, away from any DC-DC and motor pins.

Large-Current Drive Interfaces and Mixed Lamination Reliability

The large-current part needs even more care. The drive interface of an AGV is often more than ten amperes, and if copper is not thick enough, voltage drop and heating show their true form very quickly. The Heavy copper PCB supplier we specify now can make copper thickness up to 6oz or even 8oz, and traces do not need to be drawn as thick as an earthworm to withstand instantaneous current, which helps a lot in shrinking the board area and leaving more isolation space. When choosing a heavy copper pcb manufacturer, I particularly value their lamination process. If interlayer bonding is not good when thick and thin copper are laminated together, delamination and blistering appear after long use, and in an environment like a warehouse where temperature and humidity fluctuate greatly, that is a time bomb.

A Chain of Engineering Decisions

So you see, from a pile of interface signals to a board that finally runs stably, the difference in the middle is not some design trick but a chain of engineering decisions: which signals should cluster, which must be pulled far apart, how to thicken the copper skin, and what kind of factory to find to deliver your partitioning concept. These dirty and tiring jobs are the real foundation that keeps a main control board from dropping the chain.

Burned Copper and Steering-Wheel AGVs

A Fleet-Wide Communication Fault from Overheated Resin

I have been in this line for nearly eight years, and what I fear most is not algorithm deviation but an AGV suddenly lying dead in the middle of running, and finding in the end that the copper skin on the controller board had burned through. This happened last year: several forklift AGVs at the site reported communication faults at the same time. When I took apart the AGV Controller PCB back, the epoxy resin near the power area had discolored, clearly degraded by long-term overtemperature. That project used a conventional four-layer board made by an ordinary board factory, with copper thickness stated as 1oz, and in practice the local current-carrying capability simply could not withstand the instantaneous current during the emergency stop of dual steering wheels. Later everything was reworked, and we switched to a factory that specifically makes heavy copper PCB manufacturer boards, increased the copper thickness of the drive part to 6oz, and also gave the outline an aluminum-base reinforcement, and for more than a year until now there has been no repeat of such trouble.

Why Steering-Wheel AGVs Differ from Differential Drive

This is why I am especially careful in selecting multilayer PCB suppliers now. Many people think an AGV controller is an ordinary smart hardware product and any ordinary multilayer pcb manufacturer can make it, but it is not like that at all. In warehouse and production-line scenarios, AGVs do not only run differential; more commonly there is a steering-wheel mechanism, one steering motor and one traction motor, and when the two motors act together, the peak current output by the controller is quite large. Moreover, many AGVs are compact, so the PCB must have a multilayer stackup for signals, with CAN bus and encoder feedback all crowded together. If the multilayer pcb supplier has not made boards that truly drive industrial motors, they can easily mix power ground and signal ground into a mess, and in the end the AGV stops in front of the automatic charging station and cannot dock, and few people would suspect the problem of ground layer return.

Judging a Heavy Copper Supplier on the Production Line

My own experience is that when looking for a heavy copper pcb supplier, you do not necessarily need to look at how many qualifications are written on their website; it is best to go look at the production line directly. Half a year ago, when evaluating a new supplier, the salesperson told me they could make thick copper boards, but when I arrived on site and looked at the etching line, the nozzle pressure simply could not meet the deep-trench requirement, and the edges of the thick copper lines were as rough as if gnawed by a dog, so I directly told purchasing to pass. On the steering-wheel interface of an AGV controller, the inner layers of a four-layer board must carry large current while the outer layers have fine dense routing for encoders and magnetic navigation sensors, which demands extremely high line width compensation and lamination uniformity. If the heavy copper pcb manufacturer does not have sufficient lamination thickness control, uneven inner copper thickness makes impedance drift, which directly closes the eye diagram of differential pair signals, and then the AGV may stop 3 to 5 millimeters beside the target point during docking and not move at all, however much you try. This kind of problem is almost impossible to compensate by software, because the physical layer has already capped the signal quality.

The PCB Is Not an Ordinary Industrial Control Board

The market has too many people who treat an AGV PCB as an ordinary industrial control board, which is actually disrespect to the steering-wheel mechanism. The kinematic model of a steering wheel naturally requires the controller to process high-precision steering angle feedback and a high-response speed loop at the same time, the power stage on the board has to cope with frequent starts and stops and stall, and the ambient temperature often exceeds 40 degrees. A solid AGV Controller PCB must be taken seriously from the step of choosing the board factory. It is best to have the multilayer pcb manufacturer provide a cross-section analysis report, confirming that inner copper thickness and dielectric thickness are within tolerance bands, and power loop copper should be calculated according to actual instantaneous current values and temperature rise requirements, without trusting the recommended values in conventional design specifications. For my current projects, any involving heavy-load lifting and steering-wheel control, I choose PCBs from the few factories that can stably make thick copper above 6oz and with trace accuracy within plus or minus 0.025mm. It is more expensive, but the on-site troubleshooting time saved is enough for me to drink several more cups of coffee in the office.

A Dev-Board Factory Four-Layer Board and Intermittent CAN

In the past two years of making AGV controllers, what I have felt most deeply is not that programs and algorithms are difficult but how to get that board itself sorted out. Many newcomers start by focusing on what system the main control chip runs and how impressive path planning is, but what really chokes is actually the AGV Controller PCB. If this thing is not designed carefully, there will be pits everywhere afterward.

Take a steering-wheel AGV project in my hands. At first I wanted to save trouble and casually found a factory that made development boards to make a batch of four-layer boards. After the boards came back and were soldered, the CAN communication of the steering rotation channel kept disconnecting and the encoder feedback jumped numbers outrageously. After half a month of investigation, even changing to a better transceiver was tried, and finally when I checked ground plane noise with an oscilloscope, well, the power loop and signal loop were all tangled together. Only then did I go back and study the stackup process of multilayer boards, finding that it was not something an ordinary multilayer pcb manufacturer could solve. What I needed was a manufacturer that could separate the power layer and signal layer cleanly and also run heavy copper for the steering drive part. Ordinary thin copper skin carrying instantaneous current of ten-odd amperes produces such a large voltage drop that the motor has no strength, and positioning is erratic.

A Heavy Copper Manufacturer That Tunes Inner-Layer Copper to Load

Later I specifically found a factory that specialized in Heavy copper PCB, not the volume-oriented kind of supplier, but the kind of Heavy copper PCB manufacturer who can fine-tune the inner-layer copper thickness distribution according to your load. Their advice was very practical: partition the steering motor drive and logic control, do not route sensitive signals in the heavy copper area, match all encoder differential pairs strictly in length, and keep them away from inductors and large-current traces. After the change, it was incredibly stable, and even the clock waveform of the SSI absolute encoder was clear and bright, and all the previous glitches were gone. So I especially value this layer of selection now, and finding a reliable multilayer pcb supplier is much more effective than rewriting algorithms.

Interface Resource Planning Is a Kind of Black Art

Then there is the interface issue; interface resource planning on an AGV is a kind of black art. The board has only that much area, and you must fit the CAN interfaces of two steering drivers, one steering control channel, alternate interfaces for differential wheels, and all sorts of encoder connectors at the same time, and if you are not careful they collide. If an incremental encoder is open-collector output and the pull-up resistor is placed too close to the driver chip, the resistance value temperature-drifts as the chip heats up, and the counts lost make you doubt your life. My experience now is that I would rather reserve interfaces more aggressively early on and do the anti-interference routing thoroughly than patch with jumper wires later.

Choosing Board Factories That Understand Your Requirements

The board factory also has to understand your needs. Once I raised a mixed lamination requirement, and the salesperson was stunned and said they only make standard FR-4 with 18/18mil line spacing, and I replaced them on the spot. A factory that has really made AGV boards, when you say differential 100 ohm matching or heavy copper lamination, can answer right away, and may even remind you which high-voltage areas the solder mask opening should avoid. Once this communication cost is low, project advancement is faster by far more than a little. So do not squeeze the budget on PCB manufacturing, especially for a job like AGV where a major blunder may occur at any time; finding a factory that fits is worth more than anything.

Functional Safety and the PCB-Level Safety Loop

Safety Traces Too Thin for the Emergency Stop

I once drew a board for a team making warehouse robots. They used a differential wheel structure with a load capacity of over three tons, and the site often hit the safety bumper. The investigation showed the problem was not with the bumper itself: the safety loop traces on the controller board were too thin, and when a large current rushed through, ground bounce swallowed the emergency stop signal. In the later revision I directly found a heavy copper pcb manufacturer and made those few key traces on the safety chain 4oz copper thickness, and only then was the problem pressed down.

Separating Safety Signals from Power Loops in a Multilayer Board

Now, when I do an AGV Controller PCB, I habitually look at motor drive and logic control separately, but physically they are all squeezed on one multilayer board. The stackup of a multilayer board is not only for fitting in more traces; more often it is for separating safety-related signals and power loops by enough distance. For example, ISO 3691-4 requires the emergency braking function to reach at least PLd, Cat 3 architecture, which means dual channels plus diagnostics. If you do not do the isolation and cross-monitoring of the two channels solidly on the board, relying only on software redundancy is of no use. In a project I handled, I put the safety loop and the main control loop on the same 6-layer board, and when passing certification both EMC and safety diagnostic coverage passed in one go, relying on having found a multilayer pcb supplier willing to accompany us in stackup prototyping early on and tune the prepreg thickness and copper thickness distribution repeatedly.

Why Heavy Copper Is a Hard Requirement on AGVs

Thick heavy copper is a hard requirement on AGVs not only for the power and drive sections carrying current. For a heavy-load handling cart, the instantaneous current at start and stop can rush above 200A, ordinary 1oz copper foil simply cannot withstand it, and when the voltage drop is large, controller reset is common. The heavy copper pcb supplier I have long cooperated with showed me a temperature rise test of one of their 6oz boards: with 120A running continuously for two hours, there was almost no obvious hot spot on the thermal image. This kind of board is truly expensive, but if an AGV lies dead on the production line, the loss of one hour of line stoppage is dozens of times the board price difference, and customers will do the sum themselves.

Certification Is Not Won by a Safety Controller Alone

The industry now talks about ISO all the time, as if having certification settles everything. In fact the safety level is not finished by just one SRC-3000FS controller; the PCB-level hardware safety architecture decides whether you can obtain certification smoothly. I have seen too many solutions where the safety PLC hangs well in the schematic, but when drawing the PCB, the input filter capacitor of the emergency stop is placed beside the bumper socket, and once an external ESD hits, the safety input is pulled high directly, and then what PLd can be discussed? An experienced multilayer pcb manufacturer will, while doing DFM for you, incidentally point out these risk points, and that is what is truly valuable, not just comparing unit price and minimum line width and spacing.

An Independent Emergency Stop Chain and a Dedicated Safety MCU

For the control board of an AGV chassis, what gives me the most headache is not navigation algorithms but how to make the safety loop clean and decisive. Many solutions connect all signals, including the emergency stop button and the safety encoder, to the main control MCU in one go, which brings a hidden danger: once the Linux system freezes or a ROS node crashes, even the emergency stop fails with it. Later, on several revisions of the AGV Controller PCB, I deliberately made the emergency stop chain independent, using a dedicated safety MCU for logic processing, exchanging state signals with the main controller only through isolated IO and never through communication protocols.

Power Ripple, Contactor Noise and Thick Copper Consistency

This brings a direct requirement for the circuit board: safety-related traces and ordinary digital signals must be completely separated. If you are not careful in layout, the power ripple of the safety MCU can get into the emergency stop detection loop and trigger a false trip. Especially for some heavy-load AGVs, the motor peak current is outrageous, and the interference at the instant the emergency stop contactor cuts off can lift the ground plane by tens of millivolts of offset. I would never casually find an ordinary multilayer board factory and place an order for this kind of board. I once tried one or two multilayer pcb suppliers, and the finished copper thickness was uneven: a copper layer marked 3 oz measured half an ounce off, high-voltage trace spacing was not strictly enough, and during debugging the emergency stop loop was often reset for no reason. Later I switched to a Heavy copper PCB manufacturer specializing in heavy copper boards, and during communication the other party could clearly point out the compensation amounts for safety clearance and creepage distance, and could also make inner-layer copper thickness independently customized. After the boards came back, the stability of the whole safety loop improved by more than one grade.

Choosing a Safety MCU Without Over-Specifying

There are also considerations in choosing the safety MCU. The MCUs with dual-core lockstep on the market are indeed very suitable for functional safety scenarios, but I do not think you must blindly go for lockstep cores, otherwise the peripheral safety power supply and clock diagnostics alone can push up the BOM cost by a big chunk. On some single-steering-wheel AGVs with low speed, I prefer a Cortex-M4 with a hardware watchdog window as the safety core, pulling the dual emergency stop loop signals directly into the MCU GPIO, and then cascading them through hardware logic gates to the STO contactor coil. The MCU only does the decisions of speed monitoring and safety zone switching, and the execution layer still relies on safety relays and contactors, so even if the MCU hangs, the external hardware emergency stop button can still cut power directly. What really needs vigilance is those PCB designs of flexible solutions that entirely entrust safety outputs to main control software, which is not doing safety at all but drawing a pie.

Prototyping Pits: Prepreg Resin Content and Withstand Voltage

Let me say a few more words about pits in prototyping. An AGV Controller PCB often has to be made as a multilayer board carrying differential signals, with buried copper blocks in the interlayer for heat dissipation, and for this kind of board you must find a multilayer pcb manufacturer that truly understands the process. Once, in a hurry to get a prototype out, I found a nearby supplier to make a 6-layer 2oz copper board, and as a result the inner-layer isolation voltage broke down in the withstand voltage test. A cross-section showed that the prepreg resin content was insufficient. Later the stable supply channels were all Heavy copper pcb suppliers who proactively provided impedance test reports and withstand voltage test data. Although delivery is a few days slower, at least when a safety device sparks you know the problem is not on the printed board. These things cannot be learned from a datasheet and can only be exchanged for experience by blowing up boards again and again.

Power Rails, Voltage Swing and Thermal Structure

Why Battery Voltage Swing Makes Controllers Restart

In the years I have done AGV main control boards, I increasingly feel the design of the power section directly decides how long the whole system can last. We once had a batch of controllers that restarted repeatedly at the customer site. After checking all around, it was neither a program problem nor motor interference; it was that the power supply part of the board could not stand the voltage swing of the battery from full to depleted. Most AGVs use a 48V DC bus, which can go close to 60V while charging and drop to around 40V at the low-battery warning, while the processors, inertial navigation and lidar on the board need clean and stable low voltage. If this AGV Controller PCB merely draws a few DC-DC modules in the schematic and places them, without treating the power loop as one whole matter, trouble will keep finding you later.

Heavy Copper on the 48V Platform

For any new AGV drive control board now, I directly go to manufacturers that can do heavy copper PCB. Ordinary 1oz copper thickness can pass in systems below 24V, but on a 48V platform with sustained current often thirty or forty amperes, even if you pour large-area copper, the temperature rise is still hard to look at. A heavy copper board makes inner-layer copper thickness 3oz, 4oz or even higher, which not only spreads the heat of the current-carrying path more widely but also improves the reliability of pads and vias a lot, which is critical for carts that run on bumpy floors all year. If a multilayer pcb manufacturer only gives you boards by conventional process, without thickening for large-current inner layers, then after the board is used for a period, microcracks easily appear in the interlayer copper skin, and finally there are inexplicable short or open circuits. I have paid for this, so now when screening multilayer pcb suppliers, the first thing I ask is whether they have continuous shipping experience in heavy copper thickness processes; being able to prototype does not count as being able to make it.

Isolating the BMS Link and Drawing a Fault Boundary

Communication between the battery management system and the controller likewise cannot be left hanging. I have seen designs hang the BMS CAN interface directly near the DC-DC power loop of the main control board without even basic isolation, and as a result, when the charging pile was plugged in, a surge ran through the signal line into the microcontroller and broke even the sampling ports on the high-voltage side. Later we required that all signals docking with the battery pack be isolated, and that the isolated power supply go through a separate LDO path and not be taken directly from the 48V big loop. On one hand this prevents amplitude fluctuation from affecting communication, and on the other it forcibly draws a fault boundary at board level. AGVs work in factories, where batteries are plugged and unplugged and charge and discharge switch frequently, and if the hardware does not leave enough margin on this boundary, relying on software protocols alone for protection will certainly not hold.

Thermal Structure Matters More Than Module Efficiency

Many people think selecting a power module for a heavy-load AGV is just looking at the two indicators of wide input range and efficiency. I think the thermal structure design is more easily overlooked. The module’s nominal efficiency is based on its own heat sink or air-cooling conditions, but the AGV chassis is often sealed, with the internal air not flowing and all the heat trapped inside. This requires the PCB itself to take on a considerable part of the heat dissipation function. In multilayer stackup design, if the area directly under the power module can be made into dense copper-pillar vias conducting heat directly to the inner-layer copper, and then the thick copper inner layers provided by a heavy copper PCB manufacturer spread the heat across the whole board, the module’s case temperature can be pressed down. At high temperature, the module actively derates power, and even if its wide input range still has margin, the actual output has been castrated. So this part contests the board factory’s processing capability, not only component selection.

Honest Copper Thickness and Reliable Daily Operation

I have been in contact with several multilayer pcb suppliers, and some could only do conventional FR-4, hedging when asked about thick copper delivery time and yield. Later I found a factory that had long made industrial power boards and motor drive boards, which handled details such as opening of large-current via holes and thickened solder mask very maturely, and whose copper thickness on the board really was what it was marked, not like some factories that play word games with you using base copper thickness. Stable DC power is the foundation of all upper-layer functions of an AGV, and GPS, visual navigation and safety PLC all depend on this undisturbed underlying power architecture. To guarantee that a machine runs on the production line for sixteen hours without restarting, the places where hardware can put in effort are exactly in these copper skins and stackups that do not look so “intelligent.”

Contactor Drive Isolation and the Dirty Zone

Do Not Hang Contactor Drive on Logic Ground

After seven or eight years of AGV circuit design, there is one thing I am convinced of to this day: if you hang the contactor drive loop directly on logic ground, trouble is bound to come sooner or later. It is no mysticism; it is just a law of physics. A contactor coil is a big inductor, and the back-EMF at the instant of power-off can reach several hundred volts, and even if you use a freewheeling diode, arc noise still runs along the ground line into the main control. The most outrageous time I saw was an AGV spinning in place, and it took two days to find that the controller was being reset by interference from time to time, with the reset pin right beside a 3.3V trace on a multilayer PCB, and the ground bounce pulled the level crooked.

A Dirty Zone with a Single Ground Point

Later we unified our approach: all brutal circuits such as contactor coils and brake resistor switching are driven through optocoupler isolation, and the isolated power and ground are given a separate “dirty zone” on the multilayer PCB. The ground of the dirty zone is connected to the main ground at only one point, which I usually choose next to the battery negative terminal post. After doing this, a heavy-load AGV running with a 200 ampere charging current never had communication interrupted. I also never let CAN and 485 transceivers cross split zones, preferring to detour a bit more.

Why Inner-Layer Heavy Copper Must Be Real

When looking for a board factory, I have a very practical lesson: do not casually find any multilayer PCB manufacturer for prototyping. The main board of a heavy-load AGV controller is often six or eight layers, and large current in the inner layers must use thick copper, and some multilayer PCB suppliers quote “thick copper” that is actually plated onto the outer layers while the inner layers are still conventional thickness, and the current-carrying capacity simply cannot hold. A truly reliable heavy copper PCB supplier can give you 2oz or even 3oz copper foil right from the inner layers, and is also experienced in controlling fill and resin flow during lamination, without interlayer voids. The heavy copper PCB manufacturer we cooperate with on a fixed basis does cross-section analysis on every batch of boards, and only when they send me the micrographs do I feel at ease.

Terminal Selection and Teardrops

There is another thing many people do not take seriously: terminal selection on the AGV Controller PCB. If large-current through-hole pads are not given teardrops and reinforcing vias nearby, solder joints crack before long. I have stepped in this pit, and now I require all of them to be reinforced by the board factory according to the heavy copper PCB process, even if it costs a bit more. The underlying logic of the whole controller is actually just one sentence: isolate the crude power world and the fragile signal world cleanly, and be serious about every step from chip selection to PCB stackup.

Compound Robots: Four Motors, Lift and Rotation

Starting Current and Heavy Copper in a Mixed Lamination

People who have worked on compound robot projects should understand that the most troublesome thing about the controller is not computing power but how to get the physical layer running in a compact space. When I took this case, I stumbled at the very start on PCB selection. The AGV chassis plus the lifting and rotating mechanisms, the controller board must drive four motors at the same time: two for traveling, one for lifting and one for rotation, each channel drawing serious current, especially at the start of lifting, where the instantaneous current can rush to over three times steady state. With a PCB of ordinary copper thickness, the traces begin to heat in just a few days and the voltage drop is also obvious, and the motor output is soft and weak. Later I went directly to a manufacturer that could make Heavy copper PCB, pushed copper thickness to 4oz, and redrew trace widths, and only then was the heating on the large-current path pressed down. There was also a small pit in this, which is the effect of heavy copper boards on multilayer stackup. Many multilayer pcb manufacturers are unwilling to take small-batch mixed lamination of heavy copper and multilayer, because lamination alignment and thermal balance are hard to handle and the scrap rate is high. I had to ask one factory after another, and finally found a multilayer pcb supplier willing to cooperate. It was a four-layer board with the middle two power and ground layers at 2oz, and the top and bottom signal layers plus some power traces at 4oz. This cost slightly more, but board stability improved noticeably.

Modular Interfaces and CAN Termination

The interface side is even more trivial. If the limit switches, brake signals and encoder feedback of the lifting and rotating mechanisms were all stuffed onto the main control board with discrete connectors, the connectors alone could take up a small half of the board, and with too many wiring harnesses, interference and EMC become mines during assembly. My approach is to make the interface part of the AGV Controller PCB modular: the base board reserves two CAN channels and one 100M Ethernet, plus one set of high-speed USB, and the control of lifting and rotation is made into independent driver small boards hung on the bus through CAN. This way the main control board size can be kept compact, and the interfaces become cleaner, and later debugging and part replacement are also convenient. But note here that if the CAN bus termination resistor matching and the position of the common-mode choke are not handled well in the board layout, the bus occasionally drops frames when the motors start, and the lifting action may be delayed and stuttering, and on-site debugging nearly drove me crazy.

Hardware Synchronization Instead of Software Polling

Honestly, many people feel that the AGV Controller PCB is nothing more than a microcontroller plus a driver, and that copying a reference design is enough. But only when you do a compound robot do you know that the timing coordination between mechanisms relies on board-level hardware synchronization, not on software polling that can solve it perfectly. Later I connected the lift-in-position signal directly to the external trigger pin of the motion control chip, and the angle synchronization of the rotating platform also used hardware capture, so that the connection between the three actions of chassis traveling, lifting and rotation was pressed from tens of milliseconds of software fluctuation to the microsecond level, and the whole machine’s movement was noticeably crisper. This change looks inconspicuous, but I feel it was the step that raised the whole machine’s feel the most. If you want to save worry, when choosing a Heavy copper PCB supplier in the early stage you must ask clearly whether they can do impedance control and whether they can stack heavy copper with multilayer mixed lamination, clearing these physical limits first, so that later during system integration you lose less hair.

Half the Inner-Layer Copper and an On-Site Factory Visit

Not long ago I was making a controller board for a warehouse handling AGV and nearly got burned by a sample from a multilayer pcb supplier. The six-layer board they sent had inner-layer copper only half of the nominal, and they said “our default is like this.” You must know that the AGV Controller PCB has a motor drive section, and transient currents of tens of amperes are common; shrunken inner copper means concentrated heating, which at best derates power and at worst burns the board. For this I later made a special trip to the factory and watched them change the inner-layer copper to 3 ounces, and for the high-current loops I simply used heavy copper all the way: I found a heavy copper pcb manufacturer to make those several layers specially to 4 ounces, and only then did the inductive copper loss fall to an acceptable level. The scenario on this board was particularly complicated: a 24V supply came in and was split into several power rails, which supplied the lidar and also the isolated supply of the underlying motor encoder, and the core voltage of the processor had to run separately on its own DC-DC path. Two processors sat on the board: one ran ROS for global path planning, and the other, a low-cost MCU, handled only motor closed-loop and emergency stop signal acquisition, connected by isolated CAN, so even if the ROS side went down, the safety loop could still brake the vehicle. I have seen some teams insist on stuffing everything into one chip, and as a result heat dissipation could not be done, and in the sealed space of a forklift chassis the processor heated up within half an hour and throttled to 800MHz, the path stuttered, and that is what really delays things. Later that heavy copper pcb supplier chatted with me and said that a heavy copper board truly suited to motion control differs by one grade in the adhesion process between copper foil and base material, and the ability to control differential pair impedance is completely different. They showed several hybrid lamination structures that combined signal layers with multi-layer buried-copper power layers, which preserved high-speed line quality while withstanding the current shock of repeated motor starts and stops. Come to think of it, the thing most to be feared in choosing a multilayer pcb manufacturer is looking only at the online quotation without asking clearly whether they have made boards for industrial control scenarios. Because common-mode interference at an industrial site is simply not something a stackup of an ordinary consumer-grade PCB can withstand, and how to split digital ground and analog ground and where to open isolation slots all depend on the designer’s understanding of the scenario. For the version of the AGV Controller PCB I made, I pulled the spacing between strong and weak current to 25 millimeters, milled slots under the optocouplers used for bridging, and after the boards came back, when measured at the instant of motor start dragging two tons of cargo, the CAN waveform did not move at all. This feeling of being at ease cannot be bought with however many simulations are run.

Heavy-Load Differential Projects and Encoder Interfaces

Copper and Stackup Decide Drive Loop Heating and Signal Integrity

After these years of making mobile robot controllers, I increasingly feel that a reliable AGV Controller PCB is more headache-inducing than choosing which main control chip. It is not that finding any multilayer pcb manufacturer for a prototype and having it come back will do, especially when your chassis has to carry hundreds of kilograms of material boxes and also dock stably in narrow aisles, where the copper thickness and stackup on the board directly decide the heating of the drive loop and signal integrity. During one period I was responsible for a differential heavy-load project whose motor peak current rushed to over 60 amperes, and if the power traces on the board were not made by a professional Heavy copper PCB manufacturer, relying on conventional 1oz copper thickness widened by brute force, the temperature rise simply could not be pressed down, and the EMI test afterward could not pass either.

Encoder Isolation Links and Multi-Axis Robots

Many people put the emphasis entirely on encoder selection and decoding. This is of course critical: the SSI or BiSS interface of an absolute encoder is connected directly to the motion coprocessor, and a real-time deficit of 2 microseconds in closed-loop control affects navigation accuracy of plus or minus 5mm. But at the hardware level, if the isolation link of the encoder signal introduces noise through interlayer coupling in the PCB, even the best control algorithm will be repeatedly hit by frame-drop interrupts. My habit now is that when I meet a multi-axis compound robot scenario, for example with a six-axis arm on top and lifting rotation below, I honestly find a multilayer pcb supplier that can support multilayer boards and understands impedance matching, strictly partition analog ground, digital ground and power ground, and never save cost on this.

Embedded Copper Blocks Instead of Fans

Another easily overlooked thing is the application scenario of heavy copper. It is not only large current that needs heavy copper: in some occasions, motors start and stop frequently, and the inrush of contacts repeatedly produces thermal stress on the board, and the fatigue life of ordinary copper foil simply cannot withstand it. I tried several Heavy copper pcb suppliers. Some merely thickened the copper skin without helping you optimize the heat conduction path, and as a result the heat piled up in the inner layers, and even the copper skin beside the M4 screw hole warped. A truly experienced manufacturer will suggest burying copper blocks directly in the multilayer board under the drive bridge, or adopting a metal-core hybrid structure, which for a compact sealed chassis is far more reliable than forced fan cooling.

Understanding the Scenario Before Taking the Order

In the end, the system control of an AGV finally lands on the contest between a few layers of copper and FR-4 on the board. If the multilayer pcb manufacturer you choose cannot understand your scenario, does not know that a safety PLd level requires isolation creepage distance, and does not know how large the electrical clearance between the encoder control loop and the power loop should be, then the pits you step in later are far more expensive than the bit of prototyping fee you spent. Now I would rather spend half of my R&D time going back and forth with the factory to tune the stackup and compare ripple noise tests than wait until the vehicle runs onto the production line and randomly reports emergency stops because of ground bounce interference.

The post AGV Controller PCB Design: Inner-Layer Copper, Motor Drive Current and Safety Circuits for Warehouse and Heavy-Load Vehicles appeared first on SprintpcbGroup.

Network Attached Storage PCB: How Signal Integrity, Hot-Swap Power, Thermal Paths and Fab Choice Decide NAS Reliability

sprintpcbgroup — Fri, 06 Nov 2026 07:00:00 +0000

A NAS is only as dependable as the board that connects its drives, and many dropped disks, CRC errors and silent data problems trace back to PCB details rather than software or drives. This article collects first-hand lessons on a network attached storage PCB: impedance control for SATA and PCIe Gen3 to Gen5 differential pairs, return paths on SFF-8639 backplanes, power integrity and pre-charge for hot-swap drives, heat paths around the drive bays, and the realistic value of ECC memory. It also explains how to judge a multilayer, HDI or high speed PCB manufacturer by measured impedance reports, stackup advice and mass production history instead of price or layer count. A short FAQ follows at the end.

Why the Board Under the Drives Matters More Than the Spec Sheet

I have been fiddling with the NAS at home for some time. At first, like most people, I thought it was just a matter of assembling a low-power machine and stuffing in a few disks. Then one day, on a whim, I decided to design a custom backplane myself, and only then did I find that the real threshold of storage is not in the system or the hardware configuration but entirely on that unremarkable Network Attached Storage PCB. Unless you have personally drawn a version or two of a board that combines power for several hard drives, data links and status LED signals, it is hard to understand why it can make the quotes from a multilayer pcb manufacturer differ by several times.

At first I took the easy route and found a multilayer board factory with cheap prototyping. When the board came back and was soldered and powered up, the NVMe drive occasionally dropped, and the SATA link reported CRC errors. It took three or four days to figure out that it was not the drive; the several high-speed differential traces on the board had no impedance control at all, and the return loss was outrageous. After that I began to screen high speed PCB manufacturers seriously, and found that many suppliers claim verbally to be able to make 6 or 8 layer boards, but once you ask them to clearly state the dielectric constant tolerance in the stackup and provide measured differential impedance reports, they start to be vague. Such multilayer pcb suppliers are often good only at low-frequency digital boards or power boards, and when they meet a board like a NAS that mixes high-speed serial signals and large-current power, they quickly show their limits.

Later, because space was too tight, I wanted to try HDI routing and put blind and buried vias on the board. Result: another pit. There were quite a few hdi pcb suppliers who came knocking, but those who could truly guarantee the reliability of microvia fill while keeping laser drilling precision stably within ±50μm were not many. One sample sent over, when cross-sectioned, showed uneven hole-wall copper thickness and micro-cracks on the pad. Used in a storage device that is powered up for a long time, such a thing would have problems in less than three months. So now I am more cautious in choosing an hdi pcb manufacturer than in choosing hard drives, and without personally visiting the production line and seeing their test reports for high-speed board material, I absolutely do not place an order.

What torments a NAS most is actually not performance benchmarks but 24/7 data integrity. Unreasonable PCB stackup design, insufficient copper temperature tolerance, casual glass cloth selection: these trivial details slowly ferment after half a year of continuous running. I personally saw a seemingly intact NAS PCB whose impedance drifted by nearly 8% in a damp heat environment, directly causing the eye diagram of the Gen4 link to close completely. From that moment I understood that for storage boards, the process consistency and material control of the multilayer board manufacturer matter far more than price. Now when someone asks how to choose a PCB supplier for a DIY NAS, I tell them directly: do not spend your time comparing prices; first see whether the other side has long experience making server carrier boards or storage backplanes, and whether they can produce multi-batch impedance consistency data under the same stackup. There is not much mysticism in this trade; it is purely reliability piled up from engineering details, and finding a reliable multilayer pcb manufacturer and high speed PCB manufacturer saves far more than the little money you save thinking about which NAS system to use.

How a Cheap Mainboard Taught Me Where the Real Hardware Is

A few years ago, when I built my first NAS for my own use, all my attention was on the CPU and memory. I casually picked a cheap ITX board for the mainboard, thinking that running file sharing could not involve much load. What I got in return was a very real lesson: frequent disk drops, an array rebuilt again and again, and I-O errors in the log too many to flip through. Later I took the whole machine apart to find the cause, and found the problem was not in the hard disks and not in the system but that the SATA traces on that four-layer board were inherently unable to withstand long periods of full-speed read and write. From then on my view of a NAS changed: the core hardware is not the processor but the PCB that connects everything together.

Many people do not realize that when we assemble a proper NAS, what truly decides the ceiling is the quality of a Network Attached Storage PCB. Whether you mount four disks or eight, insert U.2 or SATA SSDs, those are only surface configurations. If the stackup design and routing margin of the board itself are not enough, no number of drive bays helps. Especially now that PCIe lanes keep increasing, when Gen4 or even Gen5 signals run, the requirements on board dielectric loss, impedance continuity and via stubs are completely different from those of an ordinary low-speed board. Expecting a cheap four-layer through-hole board to run a 16GT/s data stream steadily is not much different from asking a grocery-shopping car to run on a racetrack.

There is a place that is especially easy to step into a pit: the backplane. In the whole system, the backplane is the layer that does its work most quietly physically, with the health status, link training and power management of all the hard drives pressed on it. The SFF-8639 connector looks simple, but making the same slot compatible with SATA, SAS and NVMe means the back-side traces must serve low-speed differential and high-speed differential at the same time, and the 36dB loss budget of Gen5 is consumed extremely fast between backplane traces and connectors. A few third-party backplanes I have used, some of which, to save cost, used ordinary FR4 without any signal conditioning, would drop speed once after two weeks when fully populated with NVMe drives. This is not the hard drive’s fault; the board itself eats the signal.

So in these two years, when recommending that friends build a NAS, I directly ask them to first find a reliable multilayer pcb manufacturer or an hdi pcb manufacturer with high-speed board experience. It is not necessary to pursue any-layer interconnect, but at least the backplane should have eight or more layers, key high-speed traces should be strictly impedance controlled at 100Ω differential, and those with the capability can have the board factory do back drilling or blind and buried vias to minimize the effect of residual via stubs. Some multilayer board suppliers will brag to you about how many layers and how fine a line width and spacing they can do, but high-speed signal design is not about who has more layers; it is about who can truly land the signal integrity constraints from the drawing onto the board. A truly experienced hdi pcb supplier will give you concrete rectification suggestions on your stackup, trace spacing and reference plane splitting, instead of throwing out a confirmation of “can do” and finishing.

One more point worth mentioning is the thermal design on the PCB. We always stare at the CPU cooler and forget that hard disks are the most heat-averse parts in a NAS. The datasheets from WD and Seagate both say the recommended operating temperature should not exceed 50°C, but the airflow design of many finished machines does not take the drive area seriously at all, with fan policy tied only to CPU temperature, and the drives, stacked in a corner, quietly run above 55°C. Later I modified a four-bay chassis for a friend, cutting a row of air guide slots under the backplane to force airflow across the control chip and motor on the back of the drives, and the effect was immediate. The root of this actually lies in the board layout stage: if the PCB stacks capacitors and inductors behind the drive slots and blocks the air path, then no amount of fiddling with thermal pads later can save it.

After slowly finding the pattern, I look at only two things when choosing a board for an assembly: first, the board factory must be a high-speed digital board manufacturer with batch delivery experience, since not every multilayer pcb supplier can handle a PCIe 4.0 or higher backplane; second, the heat dissipation path of the whole system should be determined when the board is drawn, and the thermocouple-measured temperature of the hard disks is the data I trust. A NAS is not like a gaming console: running 24/7, 8760 hours a year, every bit of silent error may slowly pickle the whole storage pool. Compared with updating firmware or changing a cable, choosing a carefully designed NAS PCB at the very start is the highest value-for-money investment.

Atom C3000 Lessons: SATA and PCIe on the Same Board

I have tinkered with several NAS mainboards that I drew myself, and slowly found that what really chokes is often not chip selection or software tuning but that most unremarkable Network Attached Storage PCB itself. When many people talk about the Atom C3000 platform, they love to stare at how its several PCIe flex lanes are split or whether the number of SATA ports is enough, but few are willing to look down and see whether the board factory’s work can run Gen3 speed stably.

The year before last I built an 8-bay machine on the C3000. At first, to save effort, I directly found a multilayer pcb manufacturer that often made ordinary four-layer boards, thinking that SATA is just routing a differential pair, how hard could it be. When the sample board came back and the eye diagram was measured, a few lanes had bit error rates drifting to an unbearable level under full read-write load. After a long search I found that the board factory had handled the reference ground of the SATA and the neighboring group of PCIe far too casually, and the return path took a big detour. Later I honestly switched to an hdi pcb manufacturer capable of high-density interconnect, who directly used laser-drilled microvias to press the signal layers tighter, and the boards that came back, even stuffed in a stuffy chassis running resilio sync for several days in a row, never made another error.

This incident taught me a lesson: SATA III looks slow, but a 6Gbps signal is actually quite sensitive to dielectric loss and impedance continuity, let alone when the board also runs PCIe Gen3 and a 10-gigabit network port SerDes. Many multilayer pcb suppliers apply the same 100Ω template to all differential pairs and do not distinguish whether you run 6G or 8GT/s, and still use ordinary FR-4, with outrageous loss. A reliable high speed PCB manufacturer will first take your stackup requirements and run an SI simulation, tell you which layer is safer for SATA and which for PCIe, and may even suggest pressing the key traces onto a low-loss base material. When choosing an hdi pcb supplier, also do not look only at how many orders of any-layer they can do; ask whether they have made high-speed backplanes or at least can produce a TDR test report. Otherwise, with a chip like the C3000, originally equipped with AT hardware offload, in the end the bottleneck is stuck on those few centimeters of copper foil you drew yourself, and that would be a truly wronged feeling.

PCIe-to-SATA Bridges, Differential Impedance and the Cost of a Casual Board Factory

I have fiddled with NAS for several years, from the earliest finished chassis with a low-power mainboard to later drawing my own boards for custom storage schemes, and the pits I have stepped in were basically all concentrated on the PCB. When many people discuss NAS they always circle around software, such as TrueNAS, Unraid and file systems, but in fact it is the unremarkable circuit boards at the hardware bottom layer that decide the bottom line of whether your data is safe.

Once I built an eight-bay home NAS. The CPU had only two native SATA ports, which certainly would not be enough for eight hard drives, so I added an external PCIe-to-SATA bridge chip. My thinking at the time was simple: just route the lines according to the reference design. When the board came back and was tested, two SATA channels dropped disks frequently in 6Gbps mode and were barely usable when stepped down to 3Gbps. At first I suspected the driver or hard drive compatibility, and only after a week of struggle did I find the problem lay in the impedance of the PCB itself. The board maker I had found was not really a professional high speed PCB manufacturer, and their experience handling differential pairs was clearly insufficient: the impedance discontinuity caused by vias when a trace changed layers was never compensated, and the 100Ω differential impedance that SATA needs drifted badly. I unplugged the drive and measured the eye diagram, and the open “eye” was nearly closed. Later I made up my mind to switch to a multilayer pcb manufacturer that really had high-speed board experience and re-prototype. They adjusted the stackup structure very reasonably, with the signal layers right next to the reference ground, precisely controlled the line width and spacing of differential pairs, and then used back drilling to remove the useless via stubs, and the interference dropped by an order of magnitude. With the same schematic, different PCB factories produce two entirely different things.

This incident made me fully understand that when you play with drawing your own Network Attached Storage PCB, you cannot look only at whether the routing connects; you must first use the right manufacturer. Now, whenever a design involves PCIe Gen3 or multiple SATA 6Gbps channels, I absolutely do not casually look for a cheap multilayer pcb supplier. It is not that their attitude is bad; it is that this kind of high-speed board needs the deviations of the manufacturing process to be considered from the design end, and a reliable hdi pcb manufacturer will verify it for you with simulation and will tell you how many mils of layer shift will cause how much impedance change. For example, when you route PCIe differential lines on a 6-layer board, the core thickness and prepreg parameters they tune for you are in a completely different dimension from those of an ordinary double-sided board. The hdi pcb supplier I have worked with steadily for several years has quite stable laser drilling precision and copper fill when making HDI boards, so the return paths of high-speed signals are much easier to control, and the bit error rate during SATA transfers can be pressed extremely low.

There is another easily overlooked point: the heat dissipation path of the board. The PCIe-to-SATA chip on a NAS mainboard does not consume much power, but if the soldering void rate of the ePAD is on the high side, the heat cannot get out, temperature slowly builds up, and the jitter of the SATA link worsens. After being burned, I always ask the PCB side to issue an X-ray report to confirm that the pad void rate is below 25%. These details are not finished by simply making the board; you must find a factory that understands high-speed design and HDI process. So when friends around me say they want to build a NAS PCB themselves, I always add one more sentence: put software and CPU aside for a moment, find a high speed PCB manufacturer that has done signal integrity verification, and your data will be truly safe. Otherwise those boards that seem to run will, over time, produce mysterious data checksum errors that can drive you crazy.

SAS Expanders, Passive Backplanes and Connector Reality

Some time ago, when building a high-density storage unit, I was nearly dragged down by connector selection. The problem was in the backplane. I originally wanted to reuse an active backplane with an expander chip, but during testing I found that after a SAS 12G signal passed two sections of connectors and a long trace, the eye diagram simply could not open, and the link budget was extremely tight. Later I simply made the backplane passive, moving all signal processing onto the main controller board and using only SFF-8643 cables to connect directly to the drive slots. The most direct consequence was that the complexity of the main controller board jumped up two levels, with the stackup going from 10 layers to 14 layers at once, and it had to use high-speed board material. During that period I asked a number of multilayer pcb suppliers for stackup schemes, and finally managed to control the impedance at a manufacturer that had made high-speed backplanes many times.

Many people think the storage topology is just a matter of whether to place a chip on the backplane, but it is not that simple. Once you integrate a SAS expander or PCIe switch onto the mainboard, the whole Network Attached Storage PCB turns from an ordinary multilayer board into a high-speed digital board requiring strict control of differential pairs, length matching and crosstalk. The signal leaves the high-speed connector on the mainboard, passes through a long cable to the drive slot, and insertion loss and return loss must be calculated all the way. At that point you are not picking an ordinary multilayer pcb manufacturer; you are choosing an hdi pcb manufacturer that can do high-density interconnect and has real debugging experience with SAS links. The pit I stepped in was that some factories can make you a 10-layer through-hole board, but their yield starts to drop as soon as laser blind vias and any-layer interconnect come in, and in the end I had to switch back to a conservative scheme and swap the connector between the main controller board and the backplane for a lower-loss model to barely stabilize 12Gb/s.

Actually, putting active devices on the mainboard has another benefit, which is centralized thermal management. Previously I used a backplane with chips, and the system fan had to specially take care of that small area at the back, so a slightly disordered airflow made the SAS controller report speed reduction, whereas the several expander chips on the mainboard side, because their heat was spread out, had stable lifetimes. Later, in the new NAS PCB scheme, I put all four SAS expanders on the mainboard near the fan wall, and the backplane became a purely passive adapter board, keeping only high-temperature connectors. In production it is no different from a conventional double-sided multilayer board, so processing difficulty dropped immediately. This approach is very demanding on high-speed signal routing: several groups of x4 SAS channels run in parallel on the mainboard, and the crosstalk among them must be isolated by interlayer shielding and ground vias. I was glad that the hdi pcb supplier I found at that time could make the stackup this thin while also guaranteeing batch consistency of the dielectric constant.

The choice of connectors was also found by trial. In the early days, to save money, I used a compatible Mini SAS HD port, and in repeated insertion tests pin retraction appeared, which turned out to be insufficient spring contact force. Later the whole slot was changed uniformly to a model with a metal shell and latch, and although the cost per unit was higher, the whole machine never had another link drop on the vibration table. When I design a storage backplane now, even if it is only a passive adapter board, I add a separate poured copper area beside the connector and reinforce it. These details cannot be written into a spec sheet, but their influence on actual reliability is more real than how many Gb/s it runs.

Looking back now, many decisions in storage architecture are not isolated: the division of signal and status management among the backplane, connector and mainboard determines how much engineering cost you pay at which link. For people who build NAS, rather than chasing the theoretical bandwidth of the link, it is better to first find out what grade of process the hdi pcb manufacturer and multilayer pcb supplier in your hands can deliver stably, because the design that finally lands is bound to be limited by the quality of the boards you get and the consistency of the connectors. Experimental data does not lie; run the eye diagram once and everything is clear.

SFF-8639 Return Paths: Why the Pin Map Is Not the Hard Part

When I used to make a custom storage unit for a small company, I drew a six-bay NAS backplane, and at that time it was purely being driven like a duck to the roost. At first I thought that if I placed an SFF-8639 connector on the board and pulled out the SAS and PCIe differential lines according to the pin definition, it would be done. The first version came back, and the NVMe drive occasionally dropped when inserted, while the SAS drive was rock steady. After three days of investigation, I found it was not a trace length problem and not impedance control; it was that I had cut the ground return path into a mess myself.

Many people feel that the double-sided pin design of SFF-8639 is natural, with one side for SAS and one for NVMe, isolated physically. But on a Network Attached Storage PCB, what is really deadly is not how the pins themselves are arranged but how you handle the reference layer of those few outer high-speed differential pairs after they come out. To save layers, I had merged the reference grounds of SAS and PCIe on that board, and as a result the PCIe traces crossed a split, the return path took a big detour, and the eye diagram closed directly at the receiver. The multilayer pcb manufacturer I found then was an ordinary multilayer board factory, and their technical support told me a four-layer board was completely enough. They did not know that what I was running on the board was PCIe 4.0 signaling, with the edge rate right there, and it was not a layer-count problem at all but a stackup and return-path design problem.

Later I switched to a factory that specializes in high-speed boards, and as soon as they got my Gerber they asked: do you plan for the backplane of this NAS PCB to be compatible with both SAS and NVMe drives? If so, the reference planes of the two protocol domains must be physically separated, even if it costs one more layer. I half believed and half doubted, and revised one version, hollowing out the bottom ground under those few SAS differential pairs and the PCIe link, each returning with its own reference to the corresponding controller chip, and it immediately became stable. Only then did I understand that how the pins map one to one inside the connector is not critical; what is critical is that in that short fan-out region after leaving the connector, the signal return path must not be pulled along by someone else’s ground plane.

If I were to choose a partner again now, I would have several hard indicators for an HDI PCB supplier: whether they can keep the PCIe 5.0 loss within budget on a six-layer or eight-layer board, while also reserving redundant routing for four SAS lanes. I have met some hdi pcb manufacturers who, on hearing that you want to mix 12G SAS and PCIe 5.0 on the backplane, directly suggested adding layers to ten, doubling the cost. But later I found a multilayer pcb supplier very familiar with high-density interconnect who was willing to spend time with me adjusting line width and spacing and using low-loss board material, pressing the backplane insertion loss within -12dB, and it passed the compatibility test without adding layers. For small-batch NAS products, the cooperativeness of this kind of high speed PCB manufacturer matters far more than simply touting layer count and technical specifications.

In the end, SFF-8639 is only a mechanical-level convention; it tells you how pin 1 to pin 4 correspond but does not tell you how the actual signal is seen at the other end of the board. My habit is that before drawing the backplane I first run through the complete signal path in my head, starting from the PCIe contacts of the hard drive, through the connector pads, to the PCB trace, and then to the BGA fan-out under the switch chip. If the reference layer breaks anywhere, whether you are an HDI board or an ordinary multilayer board, the result is a lost eye diagram. So when I talk with colleagues now about NAS backplanes, I basically do not mention pins and ask only one thing: have you drawn the complete return path?

Why a PCIe 5.0 Backplane Lives or Dies on the 36dB Loss Budget

Not long ago I tinkered together an all-flash NAS for myself, planning to run all eight bays over NVMe, and I took a big fall on the backplane. Many people feel that a backplane is just a matter of soldering on the connectors and pulling out a few differential pairs, and that any multilayer pcb supplier can prototype it. I used to think so too, until the first board that came back ran PCIe 3.0 fine but dropped drives frequently as soon as it reached 4.0, with link training failing in all sorts of ways. Only then did I realize this thing cannot be covered by an ordinary process at all.

After a long detour I understood that the key was stuck on that 36dB. On paper the PCIe 5.0 loss budget looks 8dB bigger than the 28dB of 4.0, which seems nice, but the Nyquist frequency doubles to 16GHz, and the dielectric loss of ordinary FR4 in this band climbs outrageously; a slightly longer trace or a couple more vias and the insertion loss of the whole link shoots upward. Those board makers who claim they can run high speed simply pick some material and build the board, and the actual loss measured at 16GHz turns out to be nothing like what they said. One earlier prototyped backplane had a differential impedance that was off by only two or three ohms, yet the eye diagram closed directly. It was only rescued, barely, after I added a retimer, which doubled the cost, and the backplane had to be revised again as well.

Later I learned my lesson and went straight to a dedicated high speed PCB manufacturer, and specifically one with real vector network analyzer measurement capability in-house. After talking with four or five companies, I settled on one that did full impedance simulation and a custom stackup, going over everything from the type of glass fiber weave to the roughness of the copper foil, and only then was the loss pressed within budget. Along the way I also tried an hdi pcb manufacturer, because their process can make vias extremely short with well-controlled parasitics, so the residual signal on the backplane dropped a lot, and in a crowded drive-bay layout like this it was actually cleaner than routing on a traditional multilayer board alone.

Another easily overlooked item is the connector. Whether it is SFF-8639 or SlimSAS, the place where it mates with the backplane is often treated as a lumped load, but at Gen4 and Gen5 even slight reflections here pile up into a considerable return loss. On one backplane version I narrowed the shape of the connector pads slightly, and the return loss came down by several dB, something nobody at an ordinary multilayer board factory will remind you of. A particularly meticulous hdi pcb supplier did differential-mode impedance compensation at the connector location for me, and in the end the channel consistency measured much better.

So later, when someone asks me who to go to for a NAS backplane, I say: do not rush to compare unit prices, and do not assume that the production capability of a multilayer pcb manufacturer can necessarily cover signals of more than ten GHz. A board that powers up is one thing; a board whose bandwidth does not collapse under full-load random reads and writes is another. Those cheap prototype shops do not even guarantee batch stability of the dielectric constant, and they struggle even to make a four-layer board, so holding the channel-to-channel deviation within 3dB at 16GHz basically depends on luck. In the Network Attached Storage PCB system the backplane is the blood vessel; if the board factory is not up to standard, the whole NAS ends up a lump of scrap iron.

Hot-Swap Pre-Charge: The Resistor and MOSFET That Fought Back

The deepest pit I stepped in while fiddling with my own NAS backplane was trusting the reference design too much. I figured a pre-charge circuit is just a resistor plus a MOSFET, so I would put them down according to the datasheet and be done. When the first batch of boards came back and I hot-swapped a 16TB enterprise drive, the drive repeatedly spun up and spun down, and the 3.3V rail on the backplane shook like an electrocardiogram. At the time I thought the drive’s power requirement was simply too high. Only later, when I captured the current waveform at the instant of pre-charge with an oscilloscope, did I find that the power loop routing area was too large, and the parasitic inductance amplified the voltage spike enough to make the supply monitor chip trigger protection again and again. The multilayer PCB manufacturer only cared about connectivity when making the board and paid no attention to the real physical path of the current, and that is how I was taught a lesson.

Hot-swap on a backplane goes far beyond the on/off logic at the circuit level; its hard layout constraints are even more tormenting than signal integrity. The position of the pre-charge resistor is not wherever there is free space; it must form an extremely short loop connection at the root of the gold fingers, otherwise at the instant the inductance stores energy an unexpected resonance will appear, enough to repeatedly stress the MOSFET outside its safe operating area. When I re-prototyped, I simply pressed the spacing between the power plane and the ground plane down to 3mil and had a certain hdi pcb supplier laminate it with a high-speed board process, stitching the power plane with denser vias, which forced the parasitics of the power loop down by an order of magnitude. After that change hot-swap never gave trouble again; no matter how many drives were plugged and unplugged at the same time, the backplane stayed rock steady.

Many people choosing a NAS PCB manufacturer look only at routing density and differential impedance, which is of course not wrong, but I feel power integrity is the real litmus test in hot-swap scenarios. Especially for a hybrid board like the Network Attached Storage PCB, which runs high-speed signals and also supplies tens of watts of power, it is extremely important to find a proper high speed PCB manufacturer who is willing to discuss stackup simulation details with you. Quite a few multilayer pcb suppliers only ask you how many layers and what line width and spacing, and the moment you ask them to do a PDN simulation they start making excuses. With a backplane made by this kind of factory, when you connect an oscilloscope and look at the transient response during pre-charge, the waveform jumps up and down, so how could the drives not develop strange problems.

If I start a new NAS backplane now, the first thing I do is not draw the schematic but confirm the dielectric thickness and copper foil roughness again and again with the hdi pcb manufacturer, because the loss tangent not only affects the PCIe eye diagram but also quietly influences the damping characteristics of the pre-charge loop. To me every shape of copper on the backplane is a deliberately designed electromagnetic structure, not simple connectivity. Many DIY players may feel such minute matters are overcorrecting, but when your NAS is filled with eight drives running data 24 hours a day and does not drop a drive for months, you will understand that all those nights spent fighting parasitics in the layout were worth it.

Four-Layer Defaults and the SATA Eye That Nearly Closed

In my years of playing with NAS, many people around me start by talking about hard drives and operating systems, as if the backplane inside the chassis and the routing on the mainboard were not worth mentioning at all. But if you have taken apart a few machines, drawn boards yourself or hunted for a contract factory, you know the water here is much deeper than choosing a hard drive. Take the Network Attached Storage PCB I struggled with the longest: at first, to save trouble, I found an ordinary multilayer board factory to prototype it, and I directly reused the default parameters they gave for the four-layer stackup, and in operation it dropped drives from time to time. Later, probing with an oscilloscope, I found that the impedance control of the SATA differential pairs was a mess and the eye diagram was nearly closed. This made me fully understand that making a NAS PCB is not a matter of throwing it at any multilayer pcb manufacturer and being done; for the high-speed part, especially now that six or seven bays are common and the backplane traces are squeezed together, you must honestly find a high speed PCB manufacturer that really understands high-speed design, because the stackup structure and dielectric thickness they provide are the ones that can be used. Later I switched to a supplier that specializes in hdi pcb, who used laser blind vias to compress the signal layers more tightly, and the crosstalk between bays immediately dropped by several dB. Since then I have developed a habit: when choosing a multilayer board supplier, I do not look at their ordinary business volume; I ask directly whether they have a production line that has long made hdi pcb as an hdi pcb supplier, and whether they can provide back drilling, which is crucial for reducing via stubs.

ECC Memory: The Bit Flip That Changed My Mind

Speaking of this, I think memory follows the same logic, except it torments people in a more hidden way. Many people feel ECC memory is for servers and that on a home NAS it is purely a waste of electricity. I used to think so too, until one time my ZFS pool suddenly reported several checksum errors; the SMART data of all the drives was healthy, and changing cables and changing the backplane did nothing. Finally I ran memtest for a whole night and actually caught an occasional bit flip. That bit, of all places, happened to land on a dirty data page of the in-memory file system and was read out before it could be written back to the drive. For a home environment, a bit silently flipping in memory may be just a strange stripe on a photo or an archive that cannot be extracted, but on a NAS running 7×24, such soft errors accumulating could well scramble the metadata structures. This is exactly the value of ECC: it does not make your system faster, it catches the bit flipped by cosmic rays or power supply ripple before the data is actually corrupted. Afterwards I replaced all the non-ECC memory on that machine with ECC-capable modules. That does not mean I can rest easy from now on, but at least I can be sure that the next time an anomaly appears, the direction of my troubleshooting will not be ghost stories inside the black box of memory.

ECC Is Icing, Not Foundation: Where Board Quality and Power Really Count

In my years of playing with NAS, I increasingly feel the community has raised ECC memory too high, as if data would evaporate at any time without ECC. In fact, among the machines I have built myself, from the J3455 to the custom NAS PCBs I later drew myself, what really gave me headaches was never memory bit flips but the signal quality and power stability of the board itself. A multilayer PCB has to carry several groups of high-speed differential lines, with SATA, PCIe and USB 3.2 all squeezed together, and if you have it prototyped by a manufacturer that has only done simple double-sided boards, crosstalk and impedance control are a mess, and even with two mechanical drives it drops a disk from time to time. At that point, talking about ECC really makes no sense.

After all, ECC is icing on the cake, not the foundation. I have seen too many people force ECC onto an old server platform, with standby power of fifty or sixty watts, higher than my whole machine running with four drives. They ignore a fact: for a home or small studio NAS, drive spin-down and low system load take up the vast majority of the time, so the platform’s idle power is the key that decides the electricity bill. Even if you pick an SoC that is labeled as supporting ECC, if the traces on the multilayer board are slightly longer, the few extra check bits may fail first at high temperature, and ECC instead becomes a destabilizing factor. My own solution is very simple: directly find a reliable high-speed PCB manufacturer, have them step in from the stackup design stage, use a 6-layer board to make all the memory signals length-matched with reference ground isolation, and non-ECC memory still runs memtest for hundreds of hours with zero errors.

Speaking of PCB material selection, and I will say it from the heart, an ordinary multilayer board supplier and a manufacturer that can do HDI are two completely different worlds. Although most NAS PCBs still do not need microvias, if you want the board to be compact, with four routing layers able to hold dual M.2 and a 10GbE port, and with line width and spacing pressed to 3mil/3mil, you must find an HDI PCB supplier who understands impedance simulation. My most recent batch of boards was sent to a factory in South China that specializes in high-speed multilayer boards, and they suggested I change the topology from the SoC to the memory from T-type to fly-by, which saved a signal layer, and the bit error rate was actually lower. The cost saved from this one layer of board thickness had long since covered the small extra expense of a controller chip that uses higher-grade memory.

The power consumption issue also has to be split apart. The extra current that the ECC logic itself consumes is actually negligible; the big part is always a wrong platform choice and an unreasonable power network design. On a PCB, if the inductors of the buck circuit are placed too densely and the copper is not wide enough, the heat they bring is an order of magnitude more than the dynamic power of that little ECC controller. I once drew a NAS baseboard for an RK3588, and the first version used only four layers, with the DC-DC area stacked over the SATA signals, and as a result the 12V line loss was so large that the voltage drop when the mechanical drive motors started was severe and the system rebooted directly. After changing to a six-layer board with a complete ground layer laid down, still without ECC, the temperature of the whole board dropped by seven or eight degrees and the idle power consumption actually dropped by 3 watts. What does 3 watts mean for a 7×24 device? A year of electricity is enough to buy an 8GB DDR5 stick.

So when I talk with friends about how to choose for a self-built NAS now, I always put the PCB design specification and the partner factory’s capability first and only then talk about ECC. If the multilayer PCB manufacturer you find cannot even provide an S-parameter report for high-speed signals, then the ECC traces on the board most likely do not meet the specification either, and enabling the function is just deceiving yourself. In fact, an experienced HDI PCB manufacturer can fully help you optimize those few ECC traces properly, and can even use buried vias between layers to reduce return loss, but the premise is that you must be clear about what you want. I would rather spend the budget on a high-speed PCB manufacturer that can provide complete simulation and testing than pay double the platform license fee for the label of “supports ECC”. In the end, once the board runs stably and power consumption calibration is done, the reliability of the whole NAS is far more solid than relying only on memory error correction.

Drive Cage Thermals: The Heat Is in the Disks, Not the CPU

When I first started tinkering with my own four-bay machine, all my attention went to the CPU. I always felt I had to pick a processor with a good performance-per-watt figure, with a stepping that wasted less silicon, so that standby power could be held down. After running for a week, one drive suddenly gave a bad-sector alert, and when I opened the monitoring, the CPU was only a bit over 40 degrees while the hard drives had all collectively climbed above 56. The fan policy followed only the mainboard CPU temperature, and the drive cage was stuffy as if there were no convection at all. At that moment I realized that the center of thermal design for a NAS is not the CPU at all but that row of continuously spinning hard drives.

The heat tolerance of hard drives is far less robust than many people think. In the mechanical structure the head flies only a few nanometers above the platter, and when high temperature comes the head almost skims the surface and the deformation of the platter itself is amplified. Many individual users assembling machines habitually apply desktop thinking and feel that putting a tower cooler on the CPU settles everything, forgetting that the several drives packed tightly together are the area that really needs active airflow to carry the heat away. The mainboard of some finished NAS products reserves an independent temperature-sensing chain for the disk backplane: instead of simply relying on the drives’ own S.M.A.R.T. reports and waiting for the BMC to poll, it attaches thermistors directly to the backplane, and the fan control algorithm then weights the CPU temperature and the backplane temperature together. Put plainly, the cooling curve has to change from single-point protection of the CPU into a moat that also takes care of the continuous load of the disk array zone.

This forces you to examine that unremarkable NAS PCB. In many small chassis the bays are packed tightly and the backplane connects to the mainboard, or it is simply an all-in-one board, with signal traces and power all piled onto it. Hard drive vibration and electromagnetic interference are actually quite sensitive to the routing layers, especially for SATA links or higher-speed SAS and U.2 channels; if the board layers are designed even slightly roughly, the bit error rate climbs at high temperature. Later I went around a circle of multilayer board suppliers and found that some manufacturers focused on high-speed PCB handle impedance control and the inter-board heat dissipation structure more carefully, with higher copper thickness and inner-layer current-carrying capacity, and noticeably better thermal uniformity of the whole board. HDI PCB manufacturers with high-density interconnect are also increasingly common in storage devices, because to squeeze multiple bays and a compact system circuit together, a conventional through-hole board simply cannot fit the routing.

There is a very interesting phenomenon: even if you use a very expensive high-speed PCB, if the loop filtering and power partitioning are not done well, then when the disk array does a heavy verification or high-throughput reads and writes, the ripple on the board will still smash the eye diagram directly. I stepped in this pit myself: a board that looked richly built, after scrubbing continuously for five or six hours, began to report link recovery warnings. In the end the cause was that the DC-DC loop supplying 5V and 3.3V to the drives had degraded load regulation at high temperature, and the upstream had no targeted compensation. So it is not enough to find just any multilayer board supplier or HDI PCB supplier; you have to talk with engineers who understand storage scenarios and see whether their board design has really been considered for the long-duration, high-load working mode of a drive cage.

Looking back now, when choosing a board or designing the circuit for a NAS, never look only at the parameters of the CPU cores. The hard drives are the components that need to be served first with cold air. A good NAS PCB is not finished when the traces are connected; it must take in the thermal path, power integrity and signal quality dimensions all at once. It is like building a house: if the foundation is not solid, however beautiful the decoration, it is all for nothing.

Mainboard Quality, Drive Power Rails and Idle Consumption

I have been tinkering with NAS for four or five years myself and have changed quite a few boards, and most of the pits I stepped in had to do with the mainboard itself, not the CPU or the hard drives. Many people start by staring at power consumption and bay count, not knowing that a badly laid out NAS PCB can ruin the whole machine experience. I have a board whose memory traces had no impedance matching; with two high-frequency sticks inserted it was never stable, and on a board made by another high speed PCB manufacturer there was no problem at all. Ordinary people cannot fix this kind of problem; the root lies in not taking high-speed signals seriously at design time and casually finding a cheap multilayer board factory to prototype.

Choosing a board is really choosing the vision of the multilayer pcb manufacturer. The signal types inside a NAS are mixed, with SATA, USB, PCIe and network all squeezed onto one board, and if the stackup is not planned, once crosstalk arises, the occasional dropped disk or speeds that cannot be reached make you think the hard drive is broken. The multilayer pcb supplier I later settled on for regular work will at least sandwich a complete ground plane between the key signal layers and apply length-matching constraints to the differential pairs, and only then is it stable. If you are still using those entry-level boards that do not even spring for six layers, replace them early: the power consumption looks low, but in reality signal integrity problems cause constant retransmission and the processor wastes a lot of electricity for nothing.

Hard drive power should not be skimped on either. On some boards the SATA power is taken directly from the mainboard, with 12V and 5V ripple outrageously large, and at the instant the motor starts the voltage can be pulled below 4.6V, which cuts the drive’s life directly. Before assembling a machine now I first measure the power loop, and I pay particular attention to what parts the DC-DC section on the board uses. If you do not want to fiddle, simply find a factory that can do HDI PCB and order a backplane, separating power and signal, so that routing and via density are both looked after, which is especially necessary for many bays. I have dealt with several hdi pcb suppliers, and basically the ones that can do laser blind vias can solve the width of the heavy-current copper, so no matter how high the drive start-up current is, it will not burn the board.

On power consumption, I do not blindly chase numbers that look good. A well-designed Network Attached Storage PCB will itself help you save power in every link, for example by not needing to hang a pile of compensation components at chip selection, and by keeping power conversion efficiency above 90% instead of getting a standby figure of a dozen watts by lowering the CPU frequency. On a four-bay mainboard I use myself, loaded with enterprise drives, as long as the PMIC on the board does branch control, idle drives really can be shut off completely, and the whole machine under light load can be pressed below 18 watts. That is more direct than swapping in any gold-rated power supply. In the end, do not treat the NAS board merely as a carrier that can hold parts; it is itself the key that decides power consumption, stability and lifespan. Spending a bit more to find the right multilayer pcb supplier is far more worthwhile than fumbling around with cooling and power later.

Home NAS Reality Check: Not Every Board Needs the High-Speed Playbook

After tinkering with a private NAS for these years, my deepest takeaway is: not every NAS PCB has to follow the high-speed signal integrity playbook. Many players on the market start by staring at the board material’s loss factor and insist on contacting a high speed PCB manufacturer or an hdi pcb manufacturer, as if not stuffing in a few HDI blind and buried vias would be unfair to those hard drives. But in a home environment, with four to eight bays, the links run basically SATA 6G or even gigabit networking, and the backplane trace length is only ten-odd centimeters, so at this frequency an ordinary FR-4 multilayer board can fully cope. Back then, to save trouble, I had a batch of NAS backplanes made at a low-price-oriented multilayer pcb supplier, and as a result, when the mounting screw torque was slightly too high, the board warped directly by 0.3 millimeters, and after power-up it occasionally dropped drives; it took ages to find that the connector coplanarity was the problem. This kind of mechanical-level reliability is far more deadly than fluctuations in the dielectric constant of the board material.

Later I specially found a multilayer pcb manufacturer that came from making server backplanes. Process-wise there was no HDI microvia business, just an honest 8-layer board, with copper thickness increased to 2 ounces and lamination uniformity tightly controlled. After that batch of boards was installed in chassis, the drive backplane connectors never had a poor contact even after over a hundred rounds of brutal plugging and unplugging. This made me completely change my thinking: for a NAS PCB, what determines lifespan is often unremarkable details such as batch consistency of the gold plating and the wall adhesion of the screw-hole annular rings. As for scenarios that truly need PCIe 4.0 x4 to connect an NVMe array, only those are worth considering the laser drilling scheme and high-speed materials of an hdi pcb supplier. For an ordinary home NAS, spending the money on a reliable multilayer pcb supplier is the best investment, and there is no need to hang on to the high-end high-speed PCB manufacturers for the sake of a string of theoretical eye diagram metrics.

As for the design focus of a Network Attached Storage PCB, I have always felt it should fall on two things: thermal stress relief under years of standing still, and fault tolerance under repeated plugging and unplugging. I know a team that makes enterprise NAS, and before finalizing, their backplane was specially run through an 85℃ high-temperature aging test, not to verify the eye height of the signal but to see the delamination risk of the laminate. In that environment, if the multilayer pcb manufacturer saved dwell time in the lamination step, open and short circuits will blow up three months later like a time bomb. So now when I recommend a NAS backplane scheme to others, I first ask how the frame structure is fixed, then talk about the board material category, and only last take a glance at differential pair length matching. The order must not be reversed.

Never Skimp on the PCB: NVMe Pools, Drive Temperature and Server-Grade ECC Platforms

The biggest lesson from building my own NAS over these years is: never save budget on the PCB. Many people feel this is just a circuit board, that finding a cheap multilayer board supplier to prototype it at random is enough, and that as long as the board runs, all is well. But after going through two inexplicable cases of hard drives going offline and array errors, I realized that PCIe signal integrity in a NAS is far harsher than on an ordinary desktop.

During that time I kept tinkering with a baseboard I drew myself, planning to integrate four hard drives and a compute board, and to save space I used very compact routing. The first one I found was a multilayer board manufacturer with an extremely low quote, and the prototype did light up when it came back, but once the load rose in summer, the NVMe pool would drop speed intermittently, and PCIe correctable errors would pop up in the log from time to time. After I had it redone by an HDI PCB manufacturer that came from high-speed signal work, with all differential pairs under strict impedance control and back-drilled stubs made as short as possible, the problem disappeared directly. This made me understand that a NAS PCB is simply not something that is qualified just because it lights up, especially when you start using PCIe Gen4 or even Gen5 to expand the backplane, where the dielectric loss and glass fiber effect of the board material will close your eye diagram tight at high frequency. The conventional FR4 stackup that ordinary suppliers use eats most of your budget in insertion loss at 16GHz, and with a slightly longer trace the SerDes is constantly compensating, so drive throughput becomes very unstable.

Another point is that many people think NAS thermal management can be settled by the CPU fan curve. In fact hard drives are much more sensitive to temperature. On the board I redid, I ran an independent thermocouple trace to each drive bay, feeding directly into the EC, and the fan policy was fully tied to the drive body temperature, so however hot the CPU got the fan would not spin up like mad. When I later had this board made by the HDI supplier I directly used 2-step stacked vias, which freed up space to squeeze in several more PWM and SMBus isolation channels, and noise and cooling both settled down at once. Put plainly, for a NAS that is online 7×24, the layer planning of the PCB and the capability of the outsourced factory almost decide the reliability foundation of the whole system, from the length-matched serpentine routing of differential pairs to the target impedance of the power plane PDN; every detail directly affects whether your data quietly flips a few bits in the middle of the night.

Let me also talk about ECC. My own NAS was locked in from the beginning to an ECC platform, because I have seen more than once a ZFS scrub on non-ECC memory repair mysterious checksum errors, with no way to trace whether it was the drive or a memory bit flip. At the board level you must settle the RDIMM VTT power and address line topology in advance, and some high-speed PCB manufacturers are very familiar with the layout guidelines of this kind of storage-dedicated platform and can save you a lot of debugging time; if you take the trouble-saving route of a low-price supplier, they simply do not engage with you on these details. So now when I recommend a DIY NAS to others, I always stress getting the choice of PCB board factory right first, going directly to those high-layer-count manufacturers that have really built high-speed backplanes and server-grade mainboards; even if it costs a bit more, it beats spending all day staring at SMART errors with your heart in your mouth.

Backplane Layer Count, ECC as a Physical Guardrail and Decoupled Cage Cooling

When I first assembled my first NAS, what I valued most was board layer count and routing, which many storage players tend to overlook as they stare only at the CPU and network ports. In fact, how good the foundation is of a NAS mainboard, or rather a NAS PCB designed specifically for storage, directly decides whether your data stays stable after you fill it with drives. I stepped in pits later and learned that an ordinary desktop mainboard and a qualified Network Attached Storage PCB are two completely different things, and the backplane alone can produce countless problems.

Machines that plug SATA cables directly into the mainboard are still fine, but once a drive backplane is used, the requirements on the PCB rise steeply. Some cheap backplanes on the market are just a single-layer or double-layer FR4 board with not even impedance control, with a miserable signal eye diagram, yet they still claim to support mixed SAS and NVMe hot-plugging. I have seen a backplane that claimed U.2 compatibility with a row of 12V power lines running in parallel right next to the high-speed differential pairs, with spacing of less than 15mil, and after plugging in a PCIe Gen4 enterprise drive the Correctable Error entries flooded the system log. This is something forced out by a multilayer pcb supplier who does not understand high-speed design, which is gambling with your data on luck. When I choose a backplane afterwards, I would rather pay a bit more to find a high speed PCB manufacturer that specializes in making boards for storage, who at least have measured experience in controlling differential pair length matching and controlling the glass fiber effect.

Speaking of multilayer boards, many people do not understand why NAS boards need to use 8 or even 10 layers. Fanning out the CPU and memory is one reason, but the more practical need is power integrity. For a backplane supporting 12 bays, if only one layer of power copper is provided, the voltage drop of 5V and 12V under heavy current is frightening, and the voltage sag at the instant the drives start can directly cause drives to drop. I once took apart a backplane that had worked for two years and found the copper beside the tantalum capacitor on the 5V loop had already changed color; this was because the multilayer pcb manufacturer back then did not calculate the current-carrying capacity properly, and running 7×24 over a long cycle it was bound to fail sooner or later. Later I changed to a backplane that uses heavy copper inner layers and separately split an auxiliary power plane, and never had a drop at startup again.

As for ECC, my personal attitude has gone through a reversal. In earlier years I felt that for storing photos and movies at home, why would you need ECC, it just costs extra memory money. Until one time a ZFS scrub reported a pile of permanent checksum errors, and that machine had had a non-ECC stick replaced half a year earlier; that very stick had a bit that would randomly flip at a particular address, silently polluting the metadata, and in the end all three snapshots were corrupted. My view now is very clear: as long as you run any file system with checksums on a NAS, ECC is not optional; it is a physical guardrail. Those who say it is not important should go search the web for actual recovery cases of silent data corruption and take a look. When choosing a platform, you must confirm whether the SoC itself truly supports ECC; some boards, even with ECC memory inserted, only work in non-ECC mode, and you must check this directly in the datasheet and not believe the manufacturer’s marketing copy. The worry about power consumption is entirely unnecessary: the few extra watts of ECC are not worth mentioning next to the motor power of mechanical drives and the fans.

As for cooling control, my approach now is to manage temperature with the drive cage completely decoupled from the mainboard. In many ready-made NAS chassis the fan speed is tied only to the CPU temperature, which is absurd. When the CPU is lightly loaded but the drives are rebuilding the array at full speed, the drive temperature can soar above 55 degrees while the fan speed does not budge. My scheme is to attach an NTC probe separately at the backplane air intake and run a small board that controls the fan behind the cage directly by PWM, with no logical dependency on the mainboard whatsoever.

Frequently Asked Questions

Why do NAS drives drop out even when the drives and the software look healthy?

In my experience the cause is very often on the board. Poorly controlled SATA or PCIe differential impedance, a broken return path at the connector, ripple on the 5V and 3.3V drive rails during hot-swap or spin-up, and heat trapped around the drive bays can each produce dropped disks, CRC errors and link recovery warnings while SMART still reports the drives as healthy. Measure the eye diagram, the pre-charge current waveform and the drive temperature before replacing drives.

Does every NAS board need a high speed or HDI PCB manufacturer?

No. For a home NAS with four to eight bays, SATA 6G links and backplane traces of ten-odd centimeters, a well-made FR-4 multilayer board is usually enough, and mechanical reliability such as warpage, connector coplanarity, gold plating consistency and lamination quality matters more. A high speed PCB manufacturer or an HDI PCB supplier becomes necessary when the backplane carries PCIe 4.0 or 5.0 NVMe, mixes SAS and NVMe, or packs many bays into a tight layout, because dielectric loss at 16GHz, via stubs and return paths then decide whether the link trains.

What should I ask a PCB manufacturer before ordering a NAS backplane?

Ask whether they provide impedance simulation and a custom stackup, measured S-parameter or vector network analyzer reports, back drilling to remove via stubs, and PDN simulation for hot-swap power. Ask about batch stability of the dielectric constant, copper thickness and lamination quality for inner layers that carry 5V and 12V current, and X-ray results for thermal pad voids. A manufacturer that can only quote layer count and line width is not enough for a high-speed backplane.

Is ECC memory necessary on a NAS, and does the PCB matter for it?

If you run a checksumming file system such as ZFS, I treat ECC as a physical guardrail rather than an option, because a silent bit flip can corrupt metadata before it is written back to the drives. Check the SoC datasheet to confirm that ECC really works on the platform, since some boards accept ECC modules but run in non-ECC mode. ECC does not replace a good board: memory topology, length matching, reference ground isolation and stable power rails are still decided by the PCB, and the extra power ECC draws is small compared with a poor power network design.

The post Network Attached Storage PCB: How Signal Integrity, Hot-Swap Power, Thermal Paths and Fab Choice Decide NAS Reliability appeared first on SprintpcbGroup.

Wireless EV Charging PCB: How Heavy Copper, Parasitic Inductance, Gate Loops, and Foreign Object Detection Shape Board Design from 85kHz to 6.78MHz

sprintpcbgroup — Thu, 05 Nov 2026 07:01:00 +0000

For wireless EV charging, the PCB is not a passive wiring board; it is the crossing point of high-frequency power, small signals, magnetic fields and heat. Moving from 85kHz to 6.78MHz makes skin and proximity effects, parasitic inductance in the commutation and gate loops, dielectric loss, copper thickness tolerance and etch compensation decide whether a prototype runs or burns. This article collects hands-on lessons on a Wireless EV Charging PCB: why heavy copper is only the first step, how GaN gate loops and hidden diode footprints are handled, how foreign object detection (FOD) is corrupted by switching harmonics, and how thermal shock, Z-axis expansion and lamination quality cause delamination. It also explains how to question a heavy copper PCB supplier or a multilayer PCB manufacturer, using cross-section reports, measured resonance data and mass-production experience instead of quoted copper weight or price. A short FAQ follows at the end.

Why the MHz Band Is Harder on the Board Than It Looks

In the past two years I have handled several wireless EV charging projects, from the earliest 85kHz all the way to the later ones where someone insisted on 6.78MHz, and the pits I stepped in could fill a laboratory. I find that the industry is now a bit over-enthusiastic about the MHz band, especially some start-up teams, who talk about magnetic resonance and spatial freedom at every turn, as if all problems would be solved once the frequency goes up. Anyone who has really built a Wireless EV Charging PCB knows in their heart that on the 6.78MHz road the board-level cost of implementation is much larger than imagined.

Take finding suppliers as an example. In the 85kHz days, the market was full of factories that could make a power board well, and finding a reliable Heavy copper PCB supplier was not hard: four-layer and six-layer boards with thick copper, carrying a main loop of around one hundred amps, had stable yields. But once you switch to the MHz band, at the same power level the copper thickness cannot come down, while skin effect and proximity effect squeeze the current into the few microns at the surface of the copper foil, and the heat generation climbs in somersaults. If you then contact those conventional multilayer pcb manufacturers, they hesitate even when sending the quotation once they hear you want to run tens of amperes at 6.78MHz while controlling interlayer capacitance and magnetic coupling. Later our team stubbornly tried four different Heavy copper PCB manufacturers: either the dielectric loss tangent was too large, or the lamination structure could not take thermal shock, and the boards that came back from prototyping blew the GaN gates in less than half an hour after being installed.

Many people think that going to MHz lets the magnetic coupler be made as thin as a cookie and makes vehicle packaging much easier, but the difficulty of physically reconstructing the board itself eats up that little space advantage directly. I remember one time vividly: we tried to integrate the resonant inductor and the coil into a “magnetically integrated module”. At 85kHz a standard LCC topology was used, and in the PCB layout it was enough to reduce the commutation loop area, with the parasitic parameters basically still within the prediction of the lumped model. But at 6.78MHz, even the mutual inductance between two parallel copper traces can distort the gate drive waveform beyond recognition, not to mention the leakage inductance of the planar transformer that you thought had no effect. A young engineer spent a long time measuring with a network analyzer and found that the frequency characteristics of the whole power loop deviated completely from the simulation result. The problem lay in the stackup data provided by the multilayer pcb supplier, which had omitted the actual dielectric constant of one prepreg layer. At low frequency this would at most cost a point or two of efficiency, but at MHz it directly triggered chained oscillations, and the drive power consumption was far higher than the theoretical value.

The gate drive account cannot be left out either. Both are wide-bandgap devices, but SiC and GaN perform very differently in this band. We measured a 1200V SiC module whose gate drive circuit alone, at full power at 6.78MHz, consumed auxiliary power almost equal to the power adapter of a small laptop, and that small driver area glowed in the thermal imager in a way that made you uneasy. After switching to GaN the auxiliary power came down, but the gate’s tolerance of layout stray inductance came down too, forcing you to squeeze the drive loop to the limit, and the board material and process all had to be re-evaluated, so in the end you are back in the endless loop of negotiating with a Heavy copper PCB supplier: you want thick copper, ultra-low dielectric loss and mixed multilayer lamination all together, and the factories that can do all of these can be counted on two hands.

In my view, pushing in-vehicle wireless charging fully to MHz at this stage looks more like a collective failure of physical assumptions. Nobody dares to trust results computed by looking up tables in the lumped parameter system, and EMI cannot be fudged by adding a common-mode choke and a few Y capacitors as in the low-frequency band. Every trace on the board turns into an antenna, coil near-field leakage and PCB edge radiation crosstalk mix together, and when you do a pre-compliance test, the spectral components on the analyzer are dense as a fence. So when I talk about this topic with people now, I always first urge them to think clearly: unless you must pursue extremely small volume or dynamic alignment capability, there is nothing shameful in honestly staying at 85kHz. If you really go to MHz, then treat the PCB itself as a microwave device, and from the very beginning drag the most knowledgeable Heavy copper PCB manufacturer and multilayer pcb supplier into the project together, throwing out and redoing everything from material selection to stackup design, and only then do you have the capital to continue. Otherwise, however beautifully the drawing is made, what comes back from the fab is a pile of burnt-black copper foil.

Parasitic Inductance, Heavy Copper Etching and Multilayer Loops

When I worked on wireless charging projects for electric vehicles in the past, I did not at first feel that a PCB could cause any trouble. Just carry the current and connect a few components; find a multilayer board supplier, stack it up with the conventional scheme, and as long as the inductance value is calculated right, it is done. As a result, when the board came back and was powered up, pushing the frequency to 6.78 MHz gave me a lesson directly: the driver area was so hot I could not hold my hand on it, and efficiency dropped outrageously. Only then did I realize that in the MHz band the PCB is no longer what it used to be.

On review, the pits were basically all stepped in the same place: parasitic inductance. Think about it: the commutation loop is only a small ring of trace, and a few extra nanohenries can drag the loss up a great deal through the spike at the switching instant, and all that heat is trapped in the small area of the gate drive and cannot get out. At that time I tried three or four heavy copper pcb suppliers in a row. Some claimed they could make thick copper, but above 10oz they could not control etching precision, the line width consistency was unwatchable, and at high frequency the skin effect directly made the equivalent impedance drift. Only when I met a heavy copper pcb manufacturer that specialized in heavy copper, who knew how to balance high copper thickness against interlayer bonding strength, did thermal management finally have a foundation.

There are also many twists and turns in multilayer boards. It is not the case that any multilayer pcb manufacturer can stack a few layers and have it work; how to arrange the traces and how to interleave the layers so that part of the magnetic flux cancels out is a very deep subject. I tried an ordinary four-layer board and the radiation exceeded the limit directly; later I changed to six layers and flattened the current loop, and only then barely passed. Many people tend to think too narrowly about inductance, staring only at the few turns of the coil, but at MHz the inductance of vias, pins and even the return path of copper pour all jump out to make trouble. The magnitudes are not large but their sum is enough to give you a hard time. So later I directly required the multilayer pcb supplier to run simulation that models proximity effect and skin effect together, instead of looking only at the nominal inductance value and calling it done.

Put this way, moving a wireless charging PCB toward MHz is not a simple matter of swapping the components for high-frequency ones. From copper thickness to stackup, from trace inductance to the underlying magnetic materials, the whole physical model has to be rebuilt in your head. The old “good enough” way of thinking does not even need to be tried in this band.

The 11kW Prototype That Blistered: Copper Thickness as a Cooling System

In these years of wireless charging work, I have increasingly felt that a good PCB is not finished when the circuit is drawn. Many people put all their energy into the topology and magnetic coupling design and just find any multilayer board factory to prototype the board, feeling that as long as it connects, it is fine. As a result, once power is applied, especially when running near megahertz, the whole board turns into a scalding heat source and efficiency drops in a mess. I did the same at the start, and only after paying for it did I understand that a Wireless EV Charging PCB is a completely different thing from an ordinary digital board: with high current density and high frequency, many physical laws will hard-pull you back to reality.

At that time we were building an 11kW demonstration prototype, and the transmitter side used a standard four-layer board with 1oz copper. Everything was normal during the earlier debugging, but when the frequency was raised from 85kHz to several MHz for testing, the power consumption of the auxiliary supply section suddenly ran out of control, and the copper ring around the drive resistor blistered directly. When taken apart, it was not a device selection problem; the PCB itself could not stand the heat accumulation on the current path. Put plainly, the copper was not thick enough. The copper foil of an ordinary board offers too much resistance to pulsed current of tens of amps, all the power loss turns into heat, and it piles up in an area of a few tens of square millimeters and cannot spread. The temperature did not rise slowly; within a few seconds it could blister your hand.

After that I started looking everywhere for a Heavy copper PCB manufacturer, and the process was not easy either. There are many factories that can do thick copper, but very few that can do thick copper while also controlling line width accuracy and withstand voltage. What we needed was not simply to thicken the copper but to let thick copper carry the large-current bus in the multilayer structure while other layers still carried small signals and logic control, which needs a multilayer pcb manufacturer with a very mature lamination process; otherwise the layers delaminate easily, or uneven copper thickness causes local impedance jumps. The biggest advantage of a thick copper board is not only lowering DC resistance; its ability to absorb and carry transient heat is far stronger than thin copper, equivalent to installing an invisible heat-dissipation skeleton in the PCB. The power consumption is still there, but hot spots do not accumulate easily, and only then did thermal design at high frequency have room to operate.

Once the frequency reaches the MHz level, multilayer board design is no longer only mechanical stacking. You find that the integrity of the ground plane and the length of the return path directly decide whether the board works stably. In one of our designs, just changing the position of a via reduced the loop inductance a lot, and the ringing on the switching waveform clearly converged. Only then did I realize that parameters such as the interlayer dielectric thickness and glass cloth type provided by the multilayer pcb supplier are not decorations; they genuinely affect impedance continuity and signal return. Many times a sharp rise in power consumption is not because device losses increased but because parasitic parameters on the PCB rang away the energy during switching, and all that loss is concentrated on the board, so the power consumption you measure is naturally outrageous.

I think the supplier selection step cannot be overstated. A reliable heavy copper pcb supplier dares to tell you what uniformity they can achieve in copper thickness, instead of thumping their chest and saying the maximum can reach 20oz when the actual deviation is huge, with thin corners and a thick middle. Likewise, if a multilayer pcb supplier proactively discusses with you the influence of solder mask ink on high-frequency dielectric loss rather than just hurrying you to place an order, that factory is basically dependable. Sometimes a price difference of twenty or thirty percent hides a huge gulf in yield and long-term reliability. The one we later cooperated with regularly has quite strict acceptance standards for the plating capability of thick copper boards, and every batch of boards comes with a copper thickness distribution map under thermal imaging; a supplier like this saves you a lot of worry.

Looking back now, the essence of a Wireless EV Charging PCB is a physical channel for energy: it must not only connect the devices but also bear and channel large current and high-frequency fields. If heavy copper and multilayer design are done well, power consumption and heating are much lower and a large cooling cost can be saved. Many people ask me why I do not use a more complicated cooling scheme, and I say do not rush to add fans and water cooling; make the PCB copper a bit thicker and the stackup a bit more reasonable, and many thermal problems disappear on their own. The board itself is the first cooling system, and if you ignore this, no matter how much you spend on peripheral cooling, you cannot repay the debt owed to basic physics.

GaN Gate Loops, Hidden Schottky Footprints and Supplier Requirements

Some time ago I helped a customer evaluate a receiver-side board for in-vehicle wireless charging, with a power of about 11kW and a frequency set at 6.78MHz. The customer asked right away whether an ordinary multilayer board could be used directly, and as soon as I heard it I knew this was not so simple. Under high-frequency heavy current, if the copper is not thick enough the temperature rise of the traces cannot be held down, especially near the gate loops of the GaN devices, where if the current path is a bit longer the parasitic inductance rises and ringing makes the switching waveform shake until it cannot be looked at. I had been burned before: I had stubbornly used a multilayer board with standard 2oz copper to run a MHz-class half bridge, and the gate voltage showed a very obvious dip on the rising edge; later investigation showed that the drive loop and the device input capacitance were resonating near 9MHz and just happened to be excited.

Later I learned my lesson: for any design of this type, I first talk with the heavy copper pcb supplier and ask clearly how thick they can really pile the copper, and it must be a factory that can still guarantee line width accuracy after plating. Some factories claim to make thick copper, but actually only the outer layers are laminated with thick copper foil while the via barrel copper is still a thin layer, and once heavy current hits, the via becomes the weak point. I habitually look for a heavy copper pcb manufacturer who can make copper thickness of 4oz or more and whose via wall copper thickness can keep up with the outer foil thickness, so that the thermal capacity and impedance of the whole current path stay consistent.

Another experience also left a deep impression. That board used GaN, and I deliberately shortened the gate drive loop to within 5 millimeters, but at the beginning of layout, because it was a multilayer structure, the gate signal changed layers back and forth, two extra vias were added, and the simulated inductance of the whole loop came to nearly 3nH. Do not look down on this small number: at the switching edges of 6.78MHz the gate spike could rush above 6V, and a normal 5V driver chip simply could not stand it. In the end, with no way out, I put the whole gate drive loop on the top layer, hollowed out the inner layer to use as a shielded reference, placed the driver chip tight against the GaN gate pin, and then enlarged the local copper area for natural-convection cooling, and only then was the junction temperature brought to an acceptable range. This also made me realize that many multilayer pcb manufacturers do not understand the special needs of high-frequency power; they just do conventional impedance control, and you have to watch the layout and stackup strategy yourself.

Speaking of cooling, many people now blindly believe in the “low auxiliary power” of GaN, feeling that if it generates little heat you need not consider it further. In fact, once the reverse conduction loss at MHz is added to the extra switching loss caused by high-frequency oscillation, the total heat is not small. I do not much believe in the claims that “no freewheeling diode is needed”; every time I leave a place for a Schottky diode on the PCB, and even if I do not solder it during early debugging, I must build the pad and the thermal vias in. This “hidden space” is something many people designing a wireless charging transmitter for the first time overlook, and if you add it after finding that the waveform is wrong, the board revision cycle is lengthened for nothing.

So my habit is that when looking for a multilayer pcb supplier to make this kind of board, I state requirements directly: the interlayer dielectric thickness must support a low-inductance loop, certain layers must be designated thick copper, and I confirm that their solder mask opening and copper spreading capability can meet the cooling need of my local hot zone. Business has to be this direct; do not expect that throwing over one Gerber file will run smoothly by itself, especially for something like the Wireless EV Charging PCB where power, high frequency and heat are interwoven. You have to find out the supplier’s process boundaries in advance, otherwise when the board comes back and is powered up you will be dumbfounded.

Why Simulated Inductance Reductions Rarely Survive the Fab

I have been in contact with the electric vehicle wireless charging business for almost three years, and my biggest feeling is: never be fooled by the pretty percentages in papers. If you read a review, it will tell you that an interleaved layout can make loop inductance drop by 63% in an instant, and the calculated result makes it seem as if changing to a different stackup structure would solve the deadly switching noise at MHz. But when you actually draw the board and look for a manufacturer, you find that not one of the earlier pits has gone away.

Inductance, on the drawing, is just a simulation result, but once it reaches the factory, if the heavy copper PCB supplier slightly changes the etch factor of the copper foil, or the dielectric thickness drifts a little during multilayer lamination, the magnetic field cancellation effect you spent two weeks optimizing is directly discounted. What is more annoying is that many multilayer PCB manufacturers are simply unwilling to cooperate on tightly coupled layer pairs with extremely thin dielectric, because they do the yield arithmetic. You have to state your requirements clearly, and even go to the production line yourself to watch the first article, before you can get a board whose impedance and coupling are barely acceptable.

One more point left a particularly deep impression: stress amplification from resonant voltage. Many people in the industry, whenever MHz wireless charging is mentioned, stare only at switching loss and EMI, but what really breaks people during system debugging is the voltage spikes on the resonant tank. If the Q of the resonant network you design is high, under light load or slight detuning the voltage across the capacitor can directly threaten the device voltage rating. At this point it is not the case that finding a multilayer pcb supplier to make the board in thick copper settles everything. Thick copper only solves the over-current temperature rise problem; if the layout of the inductor traces makes the loop area large, the voltage ringing brought by the peak of the resonant current makes the whole board radiate outward like a small radio station.

On the Wireless EV Charging PCB, I now tend to see stackup design as an art of compromise rather than the optimal solution in a paper. The so-called enhanced current distribution and magnetic field cancellation are in essence trading space for performance. If you want to free up adjacent layers on the board for power commutation and gate return, it means some signals have to take the long way around or add vias to change layers. And the parasitic inductance of a via in the MHz band is not a negligible small quantity at all; sometimes, in pursuit of a tightly coupled commutation loop, the two or three extra vias punched in pull the effort right back to the origin.

So I gradually lost the patience to argue with people about which layout can reduce inductance by what percentage. What I value more is whether the multilayer board manufacturer can stably control the dielectric thickness for me, whether they can combine thick copper with fine lines, and whether the waveform I measure myself on the bench after the board comes back is clean. Because in the end, the ideal conditions in the simulation fall onto real materials, and the only thing that can verify them is the board you have just taken out of the reflow oven.

Resonant Current, Etch Compensation and What a Heavy Copper Maker Should Understand

Over the past few years I have been following all kinds of wireless charging projects, from the initial few watts of phone charging to the tens or even hundreds of kilowatts of electric vehicle wireless charging now, and my biggest feeling is that the PCB has been seriously underestimated. Everyone talks about topology, coil design and magnetic coupling, but few treat the PCB itself as a key component. I think this is a huge blind spot, especially when you push the frequency to the MHz level.

In earlier years we used industrial-grade multilayer boards for a 6.78MHz system and went to a conventional multilayer pcb manufacturer, whose stackup recommendation followed the signal integrity way of thinking, with copper thickness basically at 1oz or 2oz at most, because they assumed high-frequency traces only need to consider skin depth. But what runs on a Wireless EV Charging PCB is high-power resonant current, whose peak is several times the average; skin effect is only one of the troubles, and the bigger problem is that the copper cross-section is simply not enough. Later we were forced to find a Heavy copper PCB supplier ourselves, trying from 4oz to 6oz, and in some extreme projects even evaluating 10oz copper thickness, and only then found that the fundamental problem was not insufficient copper thickness but that, at such non-standard copper thicknesses, the etching accuracy and layer-to-layer alignment of the multilayer pcb supplier fell off a cliff. Once the line width loses control, the frequency of the resonance point drifts, and at MHz a deviation of a few hundred kHz makes efficiency drop to an unwatchable level.

There is an experience here that is rarely mentioned: traces cannot be sized by current-carrying capacity alone. At MHz the trace itself is part of the resonant network, and its parasitic inductance directly participates in the resonance; tiny deviations in copper thickness and line width change this inductance value. We paid for this: we found a factory that claimed to make heavy copper PCB, and when the prototypes came back and were measured, the resonant frequency differed from the simulation by nearly one megahertz. After two months of investigation we finally found that when doing trace compensation they had applied a large etching factor based on DC experience, so the actual line width was narrower and the inductance went up. Nobody cared about such deviation in the kHz era, but at MHz it is fatal.

So when I look at PCB suppliers now, I no longer care how many oz of copper they claim they can make; I care more about their understanding of the process for high-frequency heavy current. A qualified Heavy copper PCB manufacturer should be able to discuss with you the simulation of current density distribution at different frequencies, instead of tossing you a chart from IPC-2152. If the internal copper thickness of a multilayer board at MHz is allocated unreasonably, the inner-layer copper foil becomes a hot spot because of poor heat dissipation, and that position happens to be the trace node from the resonant capacitor to the coil, where a high temperature collapses the life. These are things you only understand after really blowing up a machine, and textbooks do not write them.

Vehicle Heat, Lamination Drift and Process-First Coil Design

I have been involved in quite a few in-vehicle wireless charging projects, and after going round and round I found something rather helpless: everyone spends too much energy on magnetic circuit simulation and coil optimization, but what actually drags things down is often that unremarkable PCB. However exquisitely the coil is designed, once it is installed in the vehicle and runs, heat and vibration can throw off every precise calculation. Especially as output power goes up, you find that the copper thickness of an ordinary PCB simply cannot bear the temperature rise brought by large current, not to mention that after long-term thermal expansion and contraction, tiny delamination between the copper foil and the base material quietly changes the electrical parameters of the coil. This is why I am now more willing to deal with suppliers who have done heavy copper PCB: not because they can print thick copper, but because they understand how the change in thermal path brought by thick copper reacts back on the stability of the coupling coefficient.

Many people treat misalignment purely as a problem of mechanical installation tolerance, and I think that is a bit narrow. Seen from the PCB manufacturing step, misalignment actually starts when the copper-clad laminate is cut. If expansion and contraction control in the multilayer lamination process is not stable, the relative position of each layer of coil shifts, and this kind of microscopic misalignment simply cannot be made up by external mechanical adjustment. You get a multilayer PCB whose coil pattern looks right, but when it works in a high-frequency magnetic field, the capacitive coupling between inner layers produces extra loops from that tiny misalignment. I have seen a team spend three months tuning the resonant network, only to find in the end that the supplier’s layer-to-layer registration accuracy was insufficient, making the mutual inductance of the two coils differ from the design value by nearly seven percent. For a manufacturer that has never run vehicle-grade heavy copper PCB in volume, this is almost an unsolvable pit.

From another angle, coupling is not only the magnetic coupling between coil and coil; electric field coupling inside the PCB is equally thorny. The side of a thick copper trace has no extra solder mask protection, and when current density concentrates at the edge of the line, the electric field stress between adjacent conductors makes the parasitic capacitance deviate from the theoretical estimate. The more you try to raise the coil’s quality factor by shrinking the line spacing, the more obvious this parasitic effect becomes. I have seen the impedance control reports given by several multilayer PCB manufacturers, and the deviation between measured and simulated values at 6oz copper thickness is far larger than on a conventional 1oz board; in other words, which heavy copper PCB supplier you choose determines, to some degree, whether your coupling design is running simulation or just talking on paper.

So when I take part in such projects now, my thinking is reversed: first find a multilayer pcb supplier that can run a stable thick copper process, get their actual machining tolerances, dielectric thickness and copper foil profile data, and only then do the coil simulation and matching. The shape and number of turns of the coil are actually not the first things to be fixed, because process capability decides the ceiling of the design. Some manufacturers can make 10oz copper thickness and guarantee line width accuracy within ±20%, and this capability directly widens your current-carrying boundary, so that you can even consider greater misalignment tolerance and cover worse parking scenarios. Otherwise, however perfectly you align in the lab, once the position moves the impedance of the whole system drifts, and that is not an inherent problem of the coupling coil; it is that the foundation was not laid firmly.

Mixed Heavy Copper Lamination, Ground Planes at MHz and Near-Field Surprises

Not long ago a customer building a wireless charging post talked with me about a ground-side transmitter PCB that behaved fairly stably at the prototype stage, but problems appeared in clusters once it reached small-batch production. The most troublesome were heating and inexplicable signal drift, especially when the load changed and the resonance point of the whole system wandered. They sent me the schematic and layout to look at, and at one glance I noticed that the board used conventional 1 ounce copper and was made by an ordinary multilayer board manufacturer specializing in quick-turn prototyping.

This choice, in a scenario transmitting hundreds of watts, was simply digging a pit for themselves. The transmitter coil of wireless charging works at tens to hundreds of kHz, but once GaN or SiC switches are involved, the harmonic content of the current waveform can easily climb to the MHz level. Many people think this is just a simple low-frequency high-current circuit that can be handled by a standard PCB process, but in fact the skin effect and proximity effect brought by high-frequency harmonics make the effective conducting area of the copper foil shrink sharply and the loss multiply. At these frequencies the temperature rise curve of conventional-thickness traces is not linear at all, and local hot spots may be twenty or thirty degrees higher than the simulation result.

I advised them to switch at once to a supplier dedicated to heavy copper PCB, not the kind that can only make single or double-sided thick copper but one that can handle mixed lamination of multilayer thick copper. Because the power loop of wireless charging needs 4 ounce or even 6 ounce copper on inner layers to carry large current, while the signal control part needs fine dense traces, and this mixed structure demands outrageously high lamination process and layer alignment. When an ordinary multilayer board manufacturer meets such a requirement, it either refuses the order outright or barely makes it with terrible inner-layer copper uniformity, leading to loss of impedance control. Later they found a supplier with real mass-production experience in heavy copper multilayer boards, and after re-prototyping, the heating problem dropped by two orders of magnitude.

Thickening the copper on the PCB is only the first step. Another easily overlooked point is that the ground plane strategy of the whole board is a completely different game in the MHz band. In the kHz band we are used to relying on large-area ground pour to handle EMC, hoping a low-impedance return will swallow all the noise. But at tens of MHz, that ground pour itself starts to become a resonant cavity. The most typical case I have seen was the auxiliary power module on the receiver side, whose switching frequency happened to fall on a certain MHz point and, through the distributed capacitance between the internal ground layer of the PCB and the power coil, formed a perfect common-mode radiation path, directly interfering with the CAN bus next to it until it reported errors constantly.

The way to solve this is no longer something simply thickening the copper can do. In the layout stage you must deliberately dismantle the current loops and split the ground layer under the high-speed switching nodes, forcing the noise current to follow the path we designed instead of running everywhere. This tests the process capability of the multilayer board manufacturer very much, for example the flatness of copper filling in buried and blind vias and the tolerance control of dielectric thickness. Once I took an eight-layer design to suppliers, with third-order blind and buried vias inside and also an attempt at locally mixed FR4 and high-frequency board material; almost all prototype shops shook their heads and said the yield was too low. But a multilayer board manufacturer with real R&D capability will sit down with you to discuss stackup optimization instead of throwing out a flat “cannot do”. They will tell you between which two layers to add a high-resin-content prepreg, which can significantly improve the contradiction between high-voltage isolation and distributed capacitance.

On inductors, I think there is a very interesting trend now. To relieve the PCB, many people want to move the resonant inductor off the board and integrate it directly into the magnetic coupler, for example in topologies such as LCC-S. The direction itself is fine, but the consequence is that the PCB design pressure at the front end is not reduced; it is converted into a more complicated electromagnetic compatibility challenge. Because once most of the physical volume of the inductor leaves the PCB, the switches, rectifier bridge and sampling circuits remaining on the board are exposed to a stronger near-field magnetic field. Especially in detuned conditions, the magnetic field around the coil redistributes and may couple a mV-level offset into the ADC sampling pin, which varies with the load phase and cannot be filtered out. I scanned with a near-field probe on the test bench: for the same board, between the coil aligned and deliberately offset by 5 centimeters, the MHz envelope measured under the crystal differed by as much as 15dB. A conventional conducted emission test cannot measure this at all, because the excitation conditions are so special.

So the real chokepoint in wireless charging PCB design, especially automotive grade, is not how ingenious the schematic is but whether you can find the supplier who is willing to grind through the process details with you.

Inductors on the Board or Inside the Coupler

Now many people talking about wireless charging stuff the inductor into the magnetic coupler at every turn, as if moving work off the PCB were technical progress. I think this approach is a bit like abandoning the root for the branch. What is really hard about a Wireless EV Charging PCB is not that there is no place to put those few inductors, but whether the copper foil can hold up and the heat can be dissipated while tens of amperes run continuously. Instead of racking your brains winding wire on a magnetic core and worrying that air gaps and assembly tolerance affect the coupling coefficient, it is better to honestly find a reliable Heavy copper PCB supplier. With a thick copper board, ceramic base or directly 4oz or 6oz copper thickness, the inductor can stay honestly on the board, with traces and coils as wide as small streams, and impedance and temperature rise controlled tightly at the drawing stage.

Coupling is even more subtle. If you move the inductor into the coupler, the PCB looks smaller on the surface, but in essence you just transfer an electrical problem into a magnetic circuit problem, and the uncertainty of the magnetic circuit is much larger than that of the circuit. When temperature changes, the ferrite characteristics drift, the resonance point runs along with them, and a seemingly clever dual-switch topology, if not matched with compensation capacitors, loses soft switching in a moment.

A Heavy copper PCB manufacturer I have worked with made 8-layer boards for an in-vehicle wireless charging module, building the resonant inductor and the primary coil directly in thick copper on the board and then, through the multilayer structure, working out the electromagnetic field paths clearly, and the consistency of the coupling coefficient was much more stable than the external magnetic core scheme.

There is a point here that many people overlook: today’s multilayer pcb manufacturers have long been able to bury a magnetic substrate in an inner layer, or to build the inductor in using a planar transformer approach. What you need is not to push the problem out of the PCB but to make the PCB itself part of the solution. The advantage of a multilayer board is that signal layers, power layers and shielding layers can be arranged alternately, and the cross-coupling between inductors can be almost eliminated through the ground plane, with no need to go around in circles building reverse compensation coils.

Find the right multilayer pcb supplier and let them step in from the stackup design stage, pressing current paths, inductance values and even part of the magnetic circuit into the board, and what you end up with is a board with its own coupling capability, a clear heat dissipation path and fully electrical interfaces, instead of an “electrical plus magnetic” monster that needs repeated core adjustment and centering.

In the end, where the inductor goes is not a dogma. If you throw the inductor into the coupler, you appear to save PCB area but in fact push the most fragile link in the system onto mechanical structure and material properties. Managing the inductor on the PCB takes a bit more area, but what you get is fully controllable electrical parameters and high mass-production consistency, which for a car that really runs on the road matters far more than a board a few square centimeters smaller in the lab.

Foreign Object Detection Thermal Shock, Delamination and Board Splitting

In my years of working on wireless charging PCBs, my biggest feeling is that many people in the industry put their eyes entirely on topology and switching frequency, but hardly discuss whether the board itself can withstand the heat. On the power loop of a Wireless EV Charging PCB, the current peak is routinely over a hundred amps, and especially when foreign object detection suddenly triggers a power reduction, the transient temperature rise on the copper foil can burn ordinary FR-4 to a different color. Once, in a system-level temperature rise test, a metal bottle cap was mistakenly placed at the edge of the coil, and the system pulled the transmitter current from 120A to almost zero within 300 milliseconds and then re-swept the frequency to try to recover. In the course of this sharp current collapse and recovery, the temperature of a 2 ounce inner copper sheet directly under the coil shot from 85°C to over 210°C, the board surface bulged instantly, and with the glass transition temperature of FR-4 only around 130°C, this kind of thermal shock simply cannot be withstood.

If you go to some multilayer pcb suppliers, they only produce a scheme based on conventional thermal simulation and do not understand at all how the local hot spots brought by magnetic field coupling should be absorbed in the stackup. For example, the magnetic field under the coil induces eddy currents on the inner-layer reference ground copper, with eddy current density highest in the projection area directly facing the coil conductors and decaying quickly at the edge, so the heating there is completely non-uniform.

Unfortunately many board factories’ simulations do not even set the anisotropic thermal conductivity of copper foil correctly, let alone couple in the hysteresis heating of the magnetic core and shielding layers.

A truly useful heavy copper pcb supplier will open their mouth and ask whether the inner copper directly under the coil should reach 4 ounces or more, instead of pushing a pile of expensive high-frequency board materials and calling it a day. They understand that cooling relies on heat capacity and thermal diffusion, and the thermal time constant of 4oz copper is more than twice that of 2oz copper, enough to press down the transient peak temperature.

But I also tried several heavy copper pcb manufacturers, some of whom clearly could not control resin fill when laminating thick copper boards, so the multilayer boards delaminated directly after a power burn-in and the resistance drifted to an unwatchable level. Once a supplier used a prepreg of high resin flow to laminate two layers of 6 ounce copper; at first glance the laminated thickness looked uniform, but after 48 hours of high-temperature burn-in, because the roughening treatment of the copper surface was insufficient, the bonding force between resin and copper foil decayed, and the Z-axis thermal expansion directly pulled off the roots of the inner-layer connection holes. The on-resistance drifted from 0.2 milliohm to over 10 milliohm, the efficiency of the whole power loop dropped a notch, and it took two whole weeks to pin down that it was PCB delamination.

Later I simply split the power loop and the signal loop onto two boards: the signal board went to a multilayer pcb manufacturer specializing in high-density interconnect, using laser blind vias and any-layer interconnect to keep digital and small analog signal routing clean; the power board went separately to a thick copper factory, with all inner layers required to be 4 to 6 ounces and with black oxide treatment of the copper surface mandatory to enhance bonding force, and the two boards were connected with high-current pin headers or soldered copper bars, which saved a lot of worry at once.

As for foreign object detection, do not listen to device makers boasting about how sensitive their detection chips are; on a real vehicle the parasitic coupling of the PCB can push the detection threshold into a state of repeated false alarms. I measured on a four-layer board prototype that the dv/dt of the power copper was injected directly into the input of the detection loop through interboard coupling capacitance, forming a common-mode pulse of nearly 20 millivolts amplitude and 50 nanoseconds width, while the wake-up threshold of the detection chip was only a nominal 15 millivolts, so with the vehicle parked at the charging position doing nothing, the system frequently triggered the foreign object detection state and the whole charging session was interrupted.

Only after being burned did I learn my lesson: I hard-reserved a 3 millimeter isolation slot between the digital circuit and the power copper, together with several deeply buried shielding copper layers, and only then got the interference under control. Specifically, along the edge of the isolation slot I punched a via fence going down through to the inner-layer shielding copper, and the shielding copper layer used a crossed mesh with 0.1 millimeter line width, which both kept the impedance of the eddy current discharge path below 2 milliohms and did not affect magnetic field penetration. The PCB here is not simply a current-carrying tool; it is a skeleton integrating the power path and the thermal path, and if the skeleton is not built right, the best control algorithm afterward is just watering flowers in the mud.

Later I summarized a quick method for judging a board factory’s level: ask them directly whether they can provide cross-section reports on resin-fill uniformity for three or four layers of inner copper of 4 ounces or more, and whether they have measured how the dielectric loss factor between copper foils changes at different frequencies. Those who can answer are basically worth deeper cooperation, while those who only stress the dielectric constant of the board material have most likely never crawled around in the pit of wireless charging boards.

FOD Coil Accuracy and Phase Detection on Heavy Copper Boards

On foreign object detection for wireless charging, many system people tend to think the work on the board is simple. You assume that placing the coil on the transmitter and connecting an analog front end to sample the amplitude is the end of it, but when you really do the work you know where the pits are. Two years ago I cooperated with a Heavy copper PCB manufacturer on a 7kW wireless charging transmitter board, in which the coil section directly used thick copper, a four-layer board with 4 ounce copper on both the top and bottom layers, used for the transmitter coil and the detection coil. Only then did I realize that the winding of the coil cannot be designed independently; it must be considered together with the PCB routing, otherwise the sensitivity is miserable.

At that time, because the power loop current was large, I found a Heavy copper PCB supplier in Shenzhen that specialized in thick copper boards and promised copper thickness uniformity within ±10%. When the first batch of samples came back, we found that the induced voltage of the detection coils at different positions differed enormously, with the maximum and minimum differing by nearly 9V. Investigation later showed that it was not an asymmetric winding of the coil but insufficient etching accuracy of the thick copper board, which caused deviations in coil line width and an uneven inductance distribution. Only then did I understand that the so-called improvement of magnetic field uniformity cannot be achieved by optimizing the coil winding alone; you have to pin it down from the manufacturing accuracy of the multilayer board. Later I changed to a multilayer pcb manufacturer who could do 6-layer mixed thick copper lamination, putting the detection coils on inner layers, using 0.5 ounce precision traces for the phase detection circuit, and still using thick copper on the outer layers to carry the large current, and the problem finally improved. This experience made me completely stop believing the claim that pure simulation optimization can solve FOD; there are too many variables on the hardware implementation side.

Phase detection is much harder to serve than amplitude detection. Because you need to sample the amplitude and phase of multiple coils at the same time, the phase consistency among the analog front-end channels must be controlled within 1 degree, otherwise the joint decision is useless. We used an eight-channel synchronous sampling ADC, but the trace delays were different and phase matching was not done well in the PCB layout, so a 2 to 3 degree phase change caused by coil position offset was misjudged as a metal foreign object. Many teams who look only at algorithms would never think of this kind of problem; they assume the decision algorithm is poor, but actually the multilayer pcb supplier had a deviation in microstrip impedance control during processing, which gave a poor phase specification between channels. Later we were forced to calculate the trace lengths ourselves and hard-match the path length from the coil to the ADC front end, pressing the error to within 0.2mm, and only then did it stabilize.

Another easily overlooked item is power stage noise coupling. On a high-power Wireless EV Charging PCB, the interference radiated by the GaN switches when they work goes straight into the detection coil loop, especially the narrow MHz pulses, whose spectrum overlaps exactly with the FOD detection band. Shielding cans and isolated grounds are only remedies; the fundamental method is still physical isolation. We pulled the detection coil and the receiving coil’s power loop apart on the board, with the ground layer of the multilayer board between them for isolation, and added a tunable band-pass filter at the front of the detection circuit, whose center frequency the MCU adjusts dynamically according to the operating frequency. This filter could not use an integrated part and was built from discrete devices, and it is very sensitive to PCB layout.

Overall, in wireless charging foreign object detection, choosing the right Heavy copper PCB supplier and multilayer pcb supplier is only the first step. What really tests your skill is how you communicate process tolerances with the board factory and how you make physical compromises among the coil, phase, detection circuit and power stage later on. Without having stepped into the pits, it is hard to appreciate these details.

6.78MHz Harmonics Versus the FOD Chain

For the last two years I have been tinkering with an MHz-class in-vehicle wireless charging system and have stepped in many more pits than in the earlier tens-of-kHz projects. At first I thought it was nothing more than raising the frequency, making the coil smaller and improving efficiency, but PCB-level problems almost overturned the whole scheme. The biggest conflict was in foreign object detection (FOD). The threshold measured nicely in the lab with a signal source, but once the whole machine ran at full power, FOD reported randomly: either it did not recognize a coin, or it tripped protection even with no load. At first I suspected that the software filtering was not enough, but a spectrum analyzer later showed that the harmonics generated by the power loop switching at 6.78MHz went straight into the FOD detection chain, with high-order components even soaring to tens of MHz, completely overlapping with the resonance point of the FOD coil. Only then did I realize that PCB layout at high frequency is not at all a matter of “connecting the wires”.

We were forced to re-prototype the entire Wireless EV Charging PCB and went around a circle of Heavy copper PCB suppliers, because the current density of the transmitter coil and the power loop was so large that ordinary copper thickness could not take the temperature rise. But a Heavy copper PCB manufacturer can handle current carrying and has difficulty guaranteeing impedance control on the signal layers, especially with the FOD small-signal traces right next to the large-current commutation loop, where inductive and capacitive crosstalk are a complete mess. Later we worked deeply with a multilayer pcb manufacturer who could do both high-copper-thickness multilayer and high-density interconnect, trying to build the FOD coil into an inner layer, isolated with an independent reference ground layer and shielding copper, with power ground and signal ground joined at a single point. That multilayer pcb supplier repeatedly adjusted several versions of the stackup structure, pinned the consistency of dielectric thickness and pressed with several different base materials before bringing the common-mode noise coupling we required down to an acceptable level. Because even a difference of a few microns in dielectric thickness makes the resonant frequency of the FOD coil drift, and the instability of Dk and Df, once amplified at MHz, directly raises the noise floor of the detection chain so that the threshold cannot be set stably.

Many people think a high-frequency PCB means choosing the most expensive high-frequency board material and adding a few layers of shielding cans, but at 6.78MHz and high power, structural parasitic coupling is the gate of hell. If the FOD coil routing sits slightly close to the drain copper pour of a switch, or the return path of the gate drive loop is not enclosed tightly, high-order harmonics pour into the receive chain, and however you tune the software decision threshold it is all in vain. For our last version of the design, even the screw hole positions and the grounding path of the shielding can were ground over with the board factory for half a month, otherwise a slight change in inductance would waste the entire FOD calibration. Running a few hundred watts now, FOD can basically be held stable, but every time I think of those months I still feel that planning the PCB of a high-frequency high-power wireless charger according to distributed electromagnetic behavior from the start saves much more of your life than going to suppliers for remedies after problems appear.

Supplier Lottery, Resonant Capacitor Placement and EMC

When I used to tune wireless charging boards in the lab, what annoyed me most was not that efficiency would not go up but that every time I changed suppliers, the boards that came back were like opening blind boxes. Once, in a hurry to prototype a 6.78MHz transmit-receive board, I took the easy way and found a manufacturer that had only made ordinary multilayer boards, and as the copper thickness went up, the interlayer bonding force collapsed, and after high-low temperature cycling the board blistered directly. Only after that did I understand that for something like a Wireless EV Charging PCB, finding the right Heavy copper PCB manufacturer is much more important than getting the topology to run.

Many people think the benefit of MHz-band wireless charging is that the coil can be made small and magnetic integration is easy, which is true, but there is a hidden pit: high-frequency current crowds even harder in the copper foil. In EV scenarios the resonant current is routinely over a hundred amperes, and even if you use Litz wire for the coil, the current-carrying density in the resonant capacitor loop and the rectifier bridge section on the PCB is several times that of the kHz band. Ordinary 2oz copper thickness simply cannot take long-term heat accumulation, and local hot spots can bake FR-4 above its Tg, after which the dielectric constant drifts and the resonance point flies away. This is also why I increasingly rely on Heavy copper PCB suppliers: not because of process mysticism but because the cost of trial and error is too high. Once we did a 12-layer board at a multilayer pcb supplier, requiring 6oz inner copper and, after the outer layers were finished with immersion gold, copper pillar reinforcement, and their engineer said directly it could not be done, because their lamination line simply could not control the shrinkage of such a thick core. Later we changed to an old-brand multilayer pcb manufacturer that specialized in power and automotive boards, who gave three lamination cushioning schemes alone and passed the prototype on the first try. Anyone who has done it knows that feeling of relief.

Another underestimated point is the influence of interlayer layout of the PCB on the magnetic leakage inductance. Under MHz conditions, the trace inductance between the coil and the resonant capacitor has reached a level that cannot be ignored, and if you add two more vias, the soft-switching range of the LLC is offset. On an EV wireless charging module I once tried moving the resonant capacitor from the top layer directly to an inner layer referenced in the same layer as the coil, and the leakage inductance dropped by about 18%, with switching loss visibly coming down. But this kind of change places very high demands on the layer shift control of the multilayer pcb manufacturer, because if the alignment of the coil pattern and the inner-layer capacitor pads is slightly off, the equivalent series resistance jumps up. I have heard that some peers use any-layer interconnect and stepped copper thickness on 6.78MHz EV prototypes, embedding the resonant cavity completely in the board and even eliminating the external capacitors, shrinking the volume by multiples. However, for that kind of board, the multilayer board suppliers in China who can stably mass-produce it can be counted on two hands.

Mentioning EMC opens another blood-and-tears library. Near-field coupling in the MHz band is slippery: you think adding a shielding frame settles everything, but in fact common-mode noise runs along the DC bus to the vehicle ground, conducts to the antenna end and stirs the magnetic field uniformity into a mess. I once ran into a case where FOD could not pass the automotive consistency test at all, and after a month of investigation we found that the ringing of a piece of copper on the PCB coupled with the parasitic mode of the coil near 6.78MHz. We changed to another multilayer pcb supplier to redo the stackup, split the power ground and used multi-point grounding, and only then pressed it down. Teams that have verified EMC only in the kHz band and then dare to boast that MHz is no problem are not something I much believe.

In the end, the role of the PCB in EV wireless charging is no longer that of a simple “circuit board”. It is the crossing point of high-frequency power, signals and magnetic circuits. Many times, however beautifully the scheme is fixed, one unreliable board can delay the whole project by half a year. When I choose suppliers now, I do not look at the quotation and lead time first; I first ask one sentence: have you shipped MHz-class Heavy copper PCB in volume, and for in-vehicle wireless charging? Those who answer vaguely, even if they call themselves the leader of multilayer boards, I dare not jump in with. In this trade, being bold is not what gets things done; it is the intuition accumulated from every failed prototype.

Gate-Loop Tolerance, Distributed Coupling and Partner Selection

People who build electric vehicle wireless charging sooner or later hit a wall: it is not that the principle is so hard, but that you have to fight to the end with the PCB in the MHz band. Many people think that finding a multilayer board manufacturer to pile on a few layers of copper is enough, but once it comes to the gate drive loop, they are exposed immediately. A deviation of a few nanohenries is enough to make the switching waveform shake like an electrocardiogram, and the heat directly eats up the efficiency; this is no joke.

During the time I was looking for a Heavy copper PCB supplier, my biggest feeling was that factories that can do thick copper are a dime a dozen, but those who understand high-frequency coupling and thermal margin balance are few. What you want is not just to raise the copper thickness but to keep the power layer and the gate return layer close enough, with the dielectric thin enough that the magnetic field is tightly coupled, while still surviving thermal cycling. In this kind of multilayer board design, where to place the vias and which two layers are laminated tight against each other matter much more than merely choosing a heavy copper PCB manufacturer. Some factories stress how strong their current-carrying capacity is, but once put on a vector network analyzer, the parasitic parameters are a mess, and the gate voltage overshoot cannot be reined in at all.

Why is coupling at MHz so deadly? Because at low frequency you can still play with components as lumped parameters, but at several megahertz the whole board becomes a distributed system, and the little energy coupling between traces converts directly into EMI. The metal enclosure, the coil bracket and even the heat sink all take part in this field, and resonance points that do not exist in the kHz band all appear at MHz. At this point you rely on the traditional EMC way of testing, measuring conducted and radiated emissions, but without a near-field scan you are just guessing, yet many multilayer pcb suppliers in China do not even raise this with you; what they deliver is electrical connectivity, not field distribution.

The inductance of the gate loop is something with zero tolerance. A dual-switch topology looks like it can save gate drive channels, but once single-device stress goes up, the worst-case junction temperature calculation cannot allow the slightest wishful thinking. I once got careless in a thermal calculation, and at high temperature the gate drive waveform collapsed at the edge, the half bridge shot through, and the board burned in a miserable state. Afterward I would rather make the topology “dumber” and physically press the gate loop to the shortest, while forcing myself to choose multilayer pcb manufacturers who truly understand high-frequency electromagnetic and thermal coupling, who are willing to spend on thin prepreg in the stackup coupling and optimize the via anti-pads. These details are worth more than the numbers on the quotation.

A last word on the detection side. For foreign object detection, if the analog front end is not physically isolated from the high-power traces, MHz switching harmonics pour straight into the detection signal, and whatever FOD algorithm you use is in vain. You must separate the signal return paths completely and, at the PCB level, make their grounds free of common-mode interference. I also gained this understanding only after burning a pile of boards. So when I pick a wireless EV charging PCB partner now, I first look at their high-frequency cases and then at their production line capability; if they have not done actual near-field coupling measurements around 6.78MHz, I basically do not talk to them.

Frequently Asked Questions

Why does a wireless EV charging PCB get so much harder when the frequency goes from 85kHz to 6.78MHz?

At MHz, skin effect and proximity effect push current into the outer few microns of the copper, so heat rises even though the nominal copper thickness is unchanged. Traces, vias and copper pours stop behaving as lumped elements, and a few nanohenries of loop inductance can distort gate waveforms, cause ringing and shift the resonance point. Dielectric constant and loss tangent, stackup data and etch compensation also matter much more than at kHz frequencies.

Is thick copper enough for a wireless EV charging power board?

No, it is only the first step. In my projects, thicker copper solved over-current temperature rise but created new risks: poor etch accuracy and line width drift that move the resonant frequency, resin fill and Z-axis expansion problems that cause delamination, and weak via barrels. Ask a heavy copper PCB supplier for cross-sections showing copper uniformity and resin fill, via wall copper thickness, bonding strength after burn-in and measured results at the actual operating frequency.

How does PCB layout affect foreign object detection (FOD)?

FOD coils and their analog front end are very sensitive to switching harmonics, dv/dt coupling through board capacitance, and tolerance in coil line width and dielectric thickness. In my experience, physical isolation of detection circuits from power copper, separate reference and shielding layers, single-point ground joins, matched trace lengths between coil and ADC, and tight layer registration all improved stability more than software threshold tuning.

What should I ask a multilayer or heavy copper PCB manufacturer before ordering a wireless charging board?

Ask whether they have shipped MHz-class heavy copper boards in volume for in-vehicle wireless charging, whether they can mix thick copper inner layers with fine-line signal layers, and how they control lamination, layer shift, dielectric thickness and etch compensation. Ask for cross-section reports, thermal imaging or copper thickness distribution data, and for any measured results of dielectric loss across frequency. A supplier who can only quote layer count, maximum copper ounces and dielectric constant is not enough for this application.

The post Wireless EV Charging PCB: How Heavy Copper, Parasitic Inductance, Gate Loops, and Foreign Object Detection Shape Board Design from 85kHz to 6.78MHz appeared first on SprintpcbGroup.

Programmable Automation Controller PCB: How Copper Weight, Lamination Quality, Bus Topology, and Thermal Design Decide PAC Backplane Reliability

sprintpcbgroup — Thu, 05 Nov 2026 07:00:00 +0000

In a multi-slot PAC, the backplane is the silent spine of the control station: it distributes tens of amps of 5V and 3.3V current, carries the module bus, and has to survive hot-swap surges, vibration and years of thermal cycling. Many unexplained drop-outs, module resets and communication errors trace back to the Programmable Automation Controller PCB rather than to the controller or the program. This article collects field lessons on inner-layer copper weight, base copper versus plated copper, layer registration, Z-axis expansion in thick copper laminates, slot power planning, bus topology (parallel, serial, Ethernet-style), thermal paths and hot-swap soft start. It also explains how to question a heavy copper PCB supplier or a multilayer PCB manufacturer, using cross-section reports and proven industrial backplane history instead of price or layer count. A short FAQ follows at the end.

The Silent Backplane That Took a Whole PAC Station Offline

In the early years, when I was doing equipment control cabinet integration, I stepped into a pit that still hurts to remember. We were retrofitting a production line. The PAC controller itself was a big-name brand with a very stable reputation in the market, and the logic program had been verified over and over, so everyone felt nothing could go wrong. After the equipment ran, stations dropped every few days, and expansion I/O modules would go offline for a few seconds and then recover by themselves. We checked the communication cables, the termination resistors and the grounding, and struggled for two whole days. In the end we took the whole backplane off and sent it to the board factory for cross-section analysis, and what a find: the inner-layer copper thickness was uneven, in places it did not even meet the most basic current-carrying capacity, and a whole ground-return copper trace had micro-cracks from lamination. This backplane PCB was the most silent bone in the whole system. It makes no sound, but once it has a problem, no amount of advanced controller modules plugged into it will help.

People in industrial control who talk about PAC habitually put their energy on visible parameters such as CPU clock speed, scan cycle and redundancy switchover speed. But after dealing with a dozen or so PAC systems of different brands over these years, I increasingly feel that what really decides whether a PAC can run steadily for ten years is that backplane whose shape has almost no presence. Especially when you face a rack-type PAC with multi-slot expansion, the physical bus topology on the backplane, the copper thickness of the power traces and the lamination reliability of the multilayer board cannot be remedied by software at all. You might say that the base board of a programmable automation controller is just a few rows of connectors and a few layers of copper. In fact it is the spine of the whole control station: modules plug into it, data runs through it, and tens of amps of 5V and 3.3V current are also distributed through it. Once the spine develops bone spurs or turns brittle, a clever head is useless.

Base Copper, Layer Registration and Backplane Heat

So later I set myself a particularly stubborn rule: when I get the drawings of any PAC backplane, I do not talk about slot count or which high-speed bus it supports; I first look at copper thickness and layer count. Whether the 3 oz a manufacturer claims is truly 3 oz base copper or a plated-up “equivalent 3 oz” must be asked clearly of the heavy copper PCB supplier. In a large automation line, one backplane may carry a dozen analog modules and motion control modules at once, and the transient current on the 5V bus is far fiercer than the nominal rating; if the copper is 20 microns thinner, local heating is completely different. Not to mention some heavy-duty applications where the backplane directly distributes 24V field power: if such traces use ordinary 1 oz copper, after two or three years the copper foil fatigues in thermal cycling, line resistance drifts up, module supply voltage drifts down, and the fault symptoms become all sorts of strange things that the electrician on site simply cannot locate.

When choosing a multilayer backplane there is another easily neglected blind spot: layer-to-layer registration error. I have personally seen a fourteen-layer backplane in which one inner layer shifted by less than 3 mils during lamination, so the impedance of a whole differential bus drifted over its entire length, and the bit error rate exploded during high-speed serial communication. Only after cross-sectioning did we find that the multilayer pcb manufacturer, to meet the delivery date, had simplified the positioning step for inner-layer punching. A board like this passes the continuity test one hundred percent and looks perfectly intact, but the moment a signal of several hundred megahertz runs on it, it starts to haunt you. The PCI Express or high-speed serial backplane bus running on a PAC backplane is actually very demanding on the eye diagram, and if the board factory has not made special compensation in the lamination process for thick copper and multilayer structures, measuring static continuity of the finished board in the lab is meaningless.

There is also the matter of thermal structure. Most hardware engineers focus on the gaps between modules and the fan air duct design, but rarely think that the copper foil distribution of the backplane itself is a passive heat-spreading plane. On a high-current heavy copper PCB backplane, making the power layer and ground layer thick enough and wide enough is not only to lower resistance but also to cool the modules plugged into the backplane from below. We once compared two PAC backplanes with the same slot count specification: one used an ordinary multilayer board process, and the other used a heavy-copper scheme made by a real heavy copper PCB manufacturer. After two hours at full load, looking at the temperature distribution near the slots with a thermal camera, the hottest point of the latter was almost nine degrees lower. What nine degrees means on a semiconductor life curve I do not need to explain.

So whenever a new control cabinet scheme needs my final decision, I always pull the backplane’s multilayer pcb supplier out of the BOM and audit it separately. What I look at is not whether they can make it but whether they have long experience supplying industrial control base boards. Thick copper drilling, desmear, hole-wall copper thickness and flatness after resin plugging are not at all the same difficulty level as for an ordinary multilayer board.

Thermal Cycling, Lamination Strength and Ethernet-Style Expansion

After so many years in industrial control integration, I increasingly feel that the most underestimated item in a PAC system is that backplane. Most people stare at the CPU scan cycle and the module channel count, but however lively the bus runs, in the end everything has to pass through the traces on the backplane PCB. Some abnormal crashes and inexplicable module dropouts on site, traced to the end, are often not protocol problems but tiny deformations of the multilayer board under temperature and vibration that shifted the impedance of the differential pairs.

This is especially obvious in heavy-duty settings. For example, in a steel mill or underground belt control, the inrush current at the moment of module hot-swap can make the copper foil temperature rise sharply. Without a thick copper process, an ordinary board below 3 oz cannot survive many insertion cycles before the edges of the copper bubble and delaminate. I paid for this early on: to save trouble I picked a multilayer pcb manufacturer with a low quote, but they had never handled industrial-grade thick copper lamination, and copper cracks inside the vias disturbed the current distribution of the entire backplane until the whole rack collapsed. Since then I have remembered firmly that I must find a heavy copper pcb supplier with real experience; being able to print a board is not the same as being able to do industrial control. A qualified heavy copper pcb manufacturer will consider the match between copper thickness and the prepreg between layers, and can even calculate the coefficient of thermal expansion for you, which is what guarantees the current-carrying capacity and mechanical life of the power bus on the backplane.

The mechanical stress of a modular structure then tests the interlayer bonding strength of the multilayer board. Many PAC PCBs are 8 or even 12 layers, running high-speed differential buses and tens of amps of power at the same time, with the ground layers also having to share the cooling duty. Without a solid multilayer pcb supplier, if the lamination deviation is just slightly larger, then once the few mounting screws are tightened the board develops micro-strain, and after some time certain through-hole pads fracture. I have seen a backplane where the oblique force of module insertion alone tore a crack several millimeters long in a split ground layer inside, leaving the reference plane of a local bus floating. It sounds mysterious, but on a site with strong vibration, this kind of physical failure is harder to troubleshoot than electrical problems.

More and more controllers now turn the backplane bus into an Ethernet-style expansion architecture, in which the physical distance between modules can stretch to hundreds of meters. This seems to ease the burden on the backplane, but it actually demands more precision in board-level power management and clock distribution. Because each expansion backplane is in essence a local Programmable Automation Controller PCB subsystem, it still needs a stable power network and a low-jitter reference clock. In such cases, the stackup planning of the multilayer board and the attenuation that copper foil roughness imposes on high-frequency signals become a hidden threshold. If it goes wrong, check-retransmissions on the bus become frequent and the processing cycle is falsely inflated; it looks as if it is working, but efficiency has actually dropped a great deal.

So when I organize project specifications now, I never treat the backplane as just a “passive connector”; it is itself the skeleton of the system. When choosing a supplier, I always go over copper thickness, layer count, dielectric thickness tolerance and wear resistance of the surface finish face to face with their engineers. Besides, the upgrade cycle of a modular PAC system is often longer than that of the whole machine, and a backplane with enough design margin can span two or three generations of I/O module replacement, saving far more retrofit money than the small purchase price difference saved by cutting cost at the time. Anyone who has run a few sites can work out this account.

Bus Architecture, Slot Zoning and Module Compatibility

Working on PAC controller hardware, I have never considered the backplane a simple adapter board; it basically decides the expansion capability of the whole system and how reliable its signals are. Years ago I took apart a GE RX3i rack, whose PCI backplane and the traditional 90-30 serial backplane share one connector, which physically forces you, when making a Programmable Automation Controller PCB, to plan the pin partitions very clearly, otherwise plugging in a wrong-generation module either burns the interface or drags the bus levels down. Later, when I drew a PAC PCB of my own, I went directly to a supplier that could make heavy copper PCB for the backplane section, because the supply traces have to pass more than ten amps of continuous current, and with or without hot-swap, the current-carrying capacity of the copper foil must be considered. A conventional 1 oz board visibly changes color and shows rising voltage drop after a few live insertions. In cooperation, a reliable heavy copper PCB manufacturer will not just pile on copper thickness; it will also remind you of the effect of thick copper on interlayer bonding strength and of the release of thermal stress along the large-current path.

On the bus side, the ideas of different manufacturers actually conflict a lot. An architecture like the Aotuo NA400, which uses a serial backplane, simplifies the number of slot signal lines but places higher demands on protocol processing on the module side, while the Delta AH500 insists on the high-speed throughput of a parallel bus, and behind it the multilayer pcb designer has to fight over length matching and impedance. At this point choosing the right multilayer pcb supplier is critical: some manufacturers control four and six layers very well, but beyond ten layers the layer shift and dielectric thickness tolerance start to drift, and the eye diagram of high-speed differential pairs cannot be looked at. One of our batches of boards ignored the lamination parameters of the multilayer pcb manufacturer at the initial prototyping stage, and as a result the jitter of the PCI clock exceeded the limit and the whole batch was scrapped. Only after paying this tuition did I understand that a multilayer board is not a mere processing step; it is itself part of the physical characteristics of the backplane.

Module compatibility is also an increasingly headache-inducing matter. Modular I/O like the APAX-5000 must not only be powered when plugged into the backplane but also deal with devices of different logic levels and different response speeds. This means your backplane cannot pull just one set of buses; it must be segmented and isolated by functional area, and some low-speed modules simply go through independent serial channels so as not to slow down the arbitration efficiency of the whole PCI bus. Reflected in the PAC PCB, this means making one board into a collection of several “virtual backplanes”, with power, ground, high-speed bus and low-speed scan lines all split apart; otherwise, once there are many modules, interference and reset problems can debug you into doubting life. In the end, whether the backplane form is PCI, serial or mixed bus, what finally holds the bottom line is the board material process and the care in layout, which is why I later put the evaluation of the heavy copper PCB supplier and the multilayer pcb manufacturer at the same priority as chip selection.

Remote I/O, Heavy Copper Rework and Serial Backplane Extension

After running a few factory projects I finally understood that the hardest part of a PAC system is not the control algorithm but that seemingly dull backplane. Every time I/O has to be thrown out dozens or even hundreds of meters, the backplane immediately changes from a simple circuit board into the ganglion of the whole production line, and once the signal stumbles here, the most powerful controller behind it is useless. I took part in a renovation at a plastic woven bag factory where the original control cabinet space was entirely occupied by new equipment, and the digital modules had to be moved to an auxiliary cabinet at the other end of the workshop, at a distance of nearly eighty meters. During that period I spent almost all my time choosing the backplane scheme and communicating with PCB manufacturers.

At the first prototype I took the easy way and handed the Programmable Automation Controller PCB design straight to a quick-turn factory I often worked with, thinking that drawing a four-layer board following the reference circuit would do. When the board came back and was powered up, the power layer on the backplane simply could not take it. I had forgotten that the dynamic current these modules need was much larger than before, and the copper was made only 1 oz thick, so in less than two weeks faint burn marks appeared on the copper. Later I urgently switched to a factory that had really done work as a heavy copper PCB manufacturer, redid the inner layers with 2 oz thick copper, and also laid generous heat-spreading copper areas near the bus interfaces of the backplane. The engineer told me that a PAC backplane often has to carry module power, and thin copper looks like saving cost but is actually planting a mine for later maintenance. Since then I have formed the habit that before ordering such a PAC PCB, I must first confirm whether the manufacturer has mass-production experience with thick copper backplanes, not just whether they can make multilayer boards.

The multilayer board itself is also a watershed. Whether to use blind vias or back drilling directly affects the signal reflections after expansion. Once, we used an Ethernet backplane for remote expansion, with the racks connected by standard Ethernet cables, and each backplane had to carry its own Ethernet PHY and pulse transformer, so the board instantly went from a simple four layers to six or even eight layers. At that point, choosing a multilayer pcb manufacturer cannot be only about price; you must look at their ability to control the impedance of high-speed differential pairs. The time I paid for it, a certain multilayer pcb supplier gave a stackup that looked standard, but the dielectric constant deviation was actually large, and in the long-distance expansion test the CRC errors rose until my scalp tingled. In the end it was by changing the supplier and readjusting line width and spacing that the bit errors were pressed down, but the delivery was delayed by a full five weeks.

On expansion, my inclination now is: if internal high-speed serial bus cascading can be used, use Ethernet as little as possible for long distances. It is not that Ethernet is bad; it is indeed flexible, and you can throw in a switch to make a branch, but determinism really is a black hole. Even with the 1588 clock synchronization patch, once the frequency converters and motors on the factory floor start, microsecond-level scheduling still gets disturbed. Later I changed the scheme to a directly connected backplane expansion cable, running a synchronous serial protocol over a manufacturer-customized shielded differential cable, so that the signal from the main backplane to the expansion backplane travels almost as if on the same board, and determinism and jitter both look much better. Of course, this requires the PCB design of the backplane to reserve enough high-speed interface resources, and requires your heavy copper PCB supplier to do precision multilayer work at the same time. Such manufacturers are hard to find, but if you can find one, the electrical skeleton of the whole system becomes much more solid.

In the end, the backplane is not just a collection of passive connectors; its supply thickness, stackup strategy and the topology of the expansion interface directly define how far a PAC system can go and how much load it can carry. Many times the extra money we spend on the controller would be better spent first on doing the homework on the backplane scheme and the PCB partner. The sentence I have written most often in my own project notes is: design the backplane as an active device, and do not treat it merely as a routing board.

Power Integrity and Bus Topology on a Large PAC Backplane

In my years of PAC system integration, what gave me the most headaches was not choosing the processor but the few layers of copper on the backplane. I have met several inexplicable module-reset faults where changing the power module did nothing, and in the end a thermal camera showed that the 5V plane on the backplane was abnormally hot locally, and the voltage at several slots near the power module had dropped below 4.6V. This made me fully understand that when designing a Programmable Automation Controller PCB, power integrity and signal integrity are simply two sides of the same coin.

Many people think a backplane is just a big ribbon cable with a few slots and that having an ordinary multilayer PCB manufacturer make a board is the end of it. But the real power flow of a PAC is far more complicated than imagined. In a medium-sized system, forty or fifty amperes on the 5V rail is routine, and a 1 oz copper plane simply cannot carry that; the IR drop will undervolt the modules at the far end from the power supply, causing lost pulses in a mild case and a whole station dropping offline in a serious one. So when I later asked a factory to make backplanes, I had only one hard requirement: the copper thickness of the power layer must reach 2 oz or more, and a vented pattern must be used to avoid blistering during reflow. Ordinary multilayer board suppliers argue about cost with you and say 2 oz copper is hard to route with fine lines, but a truly professional heavy copper PCB manufacturer knows how to balance line width and current capacity, and will even design a bus-bar copper area at the power entry to lower the contact resistance.

The placement of the power module also deserves care. Do not always plug the power supply into the far-left slot for convenience; I make a habit of putting the power module as close to the physical center of the backplane as possible, which minimizes the maximum voltage drop. Sometimes, when there is really no other way, I add a solid copper bus bar layer inside the backplane and use the thick copper layer to connect the power pins of all slots directly. This is not the conventional routing approach and needs the process of a heavy copper PCB supplier to pull off; with a copper layer of 6 oz or more carrying the large current, the temperature rise can be held to single digits, far more dependable than adding a fan.

The choice of bus architecture is also strongly tied to the backplane. In the past I used a PCI backplane on a local rack, and the few nanoseconds of delay were great, but once the cable stretched to two or three meters the signal was finished. Later I tried serializing the bus and running LVDS over differential pairs on the backplane, which latched stably over a dozen meters without repeaters, at the cost of higher transceiver power and a more complicated backplane design. For still longer distances, an industrial Ethernet backplane is of course possible, but the jitter and protocol stack latency make real-time tasks torturous to write. I have learned my lesson: for any PAC PCB project, I specify the upper limit of the expansion distance at the start. If it is a closed-loop application within ten meters, I run a custom serial bus directly on the backplane, with the synchronization clock on a separate differential pair, so that the whole system is like a virtual local backplane, which guarantees determinism and also saves the trouble of switches.

Back to the link between power and bus. A high-current backplane will certainly produce magnetic fields, and if the power plane and the high-speed signal layers are stacked improperly, ground bounce and crosstalk can make you debug until you doubt life. I usually bury the most critical signal lines between two ground layers and keep the power layer apart in a separate lamination structure, which requires the multilayer pcb manufacturer to have full electromagnetic simulation traceability, not merely to stack layers for you. Some multilayer pcb suppliers, to meet the delivery date, blindly apply a generic stackup, and the PAC PCB they make shows a bus bit error rate soaring under an 80A load; in the end it still comes down to board material and copper foil uniformity.

The backplane is not a supporting role; it is the skeleton of the whole PAC system. Signals run on it and current flows through it, and the collapse of any one detail turns the most expensive CPU module into a decoration. So I would rather spend more effort in the prototype stage on several heavy copper PCB suppliers, running the thermal simulation and voltage-drop simulation thoroughly, than take the blame after the site goes down.

CPU Slots, Thick Copper and Choosing Board Makers Early

What I feel most in the industrial controller trade is that the backplane looks like a “slot board” but is full of compromises inside. Most of the pits I stepped in during one project were not chip or code problems but a board that had not been properly looked after.

Many people think a PAC backplane is done once the slots are connected and the power is wired, but when you really draw a Programmable Automation Controller PCB, you find things are far from that simple. CPU modules and I/O modules are crowded on the same backplane, and once power consumption is large, ordinary FR4 and 1 oz copper foil simply cannot take it. I once had a prototype board where, after the CPU ran at full load for less than half an hour, the thermal image near the CPU slot on the backplane was red all over, and the area was hot to the touch. The copper there was too thin, and when current density concentrated, voltage drop and temperature rise both arrived. So now, whenever I design a PAC PCB, I do not dare to economize at all on the copper thickness at the power junction and under the CPU module. At the very least I find a reliable heavy copper PCB manufacturer to make it, and some local areas even need 4 oz or thicker; otherwise heat dissipation and current-carrying capacity are hard defects.

Speaking of finding a board factory, that is another matter. With thick copper and multilayer together, many factories simply shake their heads, afraid of lamination blistering and inner-layer shorts. For one 12-layer backplane, I ran through four or five multilayer pcb suppliers, and most of them shrank back on hearing that thick copper and standard copper had to be mixed in lamination. In the end it was a friend’s introduction to a heavy copper PCB supplier who specialized in heavy copper and had made similar industrial control base boards that dared to take it. That time I learned one thing: do not look for the factory after the drawing is finished. Before drawing, you must communicate clearly with the multilayer pcb manufacturer about their process capability, such as how fine the thick copper line width and spacing can go and how thick the prepreg between layers must be, because these directly decide your stackup and routing, otherwise changing the design later can make you lose a layer of skin.

The physical size of the CPU module is also an easily overlooked variable. Never plan by assuming a single slot. One CPU module I have used had a heat sink tall enough to take up the space of the adjacent slot, and after it was inserted the neighboring module simply could not be installed. At that point you can only reserve an empty slot on the backplane or offset the CPU module structurally, but doing so disturbs the signal flow and power distribution of the whole backplane. Many people are used to filling all the slots first and dealing with it later, but for high-power modules you have to spread them out deliberately, staggering the CPU, communication modules and heat-heavy analog modules, so that the air duct in the chassis can truly sweep across every one, instead of only extracting air at the two ends while the middle cooks in one pot.

Another point is the backplane power strategy. I do not like pinning all hopes on one power feed. Even if the modules themselves have redundant designs, your backplane must also separate the power cleanly at the physical layer. The PCB traces near the current junctions must be wide enough, and the thick copper area must extend to the roots of the slot pins, not rely on a short length of thin trace to carry the load. Diagnostic circuits are secondary; the key is that under fault conditions the freewheeling path must not drag down the good feed as well. I saw a failure case where a paralleled MOSFET on the backplane broke down and went short, which directly pulled another healthy power module into protection, and the whole station was paralyzed. A hard flaw like this cannot be solved by changing the power module; it is a hidden danger in board-level design.

In the end, a qualified PAC backplane looks on the surface like a place to insert a few modules, but underneath it is a tangle of power supply, cooling, manufacturing process and physical layout all stirred together. You have to think from the viewpoint of the layout engineer, the purchaser and the field maintenance person at the same time, otherwise it is easy to make something that runs well in the lab and drops its tools on site.

Auto-Addressing, Expansion Racks and Power Zoning

When I used to do PAC system integration, what gave me the most headaches was not choosing which CPU but the backplane layout fighting with the expansion modules. At that time I always felt that as long as the signal types were right, the slots could be filled any way, but when power was applied on site, the modules near the power supply were hot enough to fry an egg, and the modules at the far end dropped offline now and then because of insufficient power. Later I talked with a supplier that makes heavy copper PCB, and one sentence from them woke me up: if you do not use thick copper on the high-current traces of the backplane, you are making the modules work on an empty stomach. From then on I paid extreme attention to the copper thickness and stackup design of the backplane. Especially for something like a Programmable Automation Controller PCB, which has to carry both high-speed differential signals and large currents, finding the right heavy copper PCB manufacturer matters far more than skimping on that bit of cost.

You see many modular PACs now, where once the local backplane is full you hang an expansion backplane next to it, and a single pull is dozens of meters. There is a detail easily overlooked here: address assignment. In the early years we used DIP switches to set addresses by hand, and as a result an electrician on site, with one shaky hand, set two modules the same, and a whole section was paralyzed; troubleshooting took three hours. Only later did I understand that the real pain point is that the backplane must have an “auto-discovery” layer of logic. It is not at the software level: the PCB hardware itself must embed identification pins for every slot, for example using the resistance of a resistor to ground to tell the CPU “I am slot number so-and-so”. This approach demands much more of the multilayer pcb manufacturer, because the identification lines have to run mixed in with high-current lines without crosstalk, which cannot be managed without solid multilayer stackup experience.

Once the expanded backplane has a dozen more modules hung on it, power supply becomes another hidden killer. I saw a project where an expansion backplane had nearly twenty analog input modules connected, and the accuracy drifted like a feather in the wind. Investigation showed that the 5V supply at the end of the backplane had dropped to 4.3V. After I changed to a reliable multilayer pcb supplier who recalculated the bus copper width and via current carrying, the problem was solved. In this kind of heavy-load scenario, the backplane must have its slots divided by power consumption level, with high-power modules placed in front and low-power sensors behind, so that the thermal budget has some margin. Many people treat the backplane only as a physical bracket and never realize that it is itself the combination of power bus, communication bus and heat-dissipation skeleton. The idea of thermal zoning does not require separate heat sinks; it means separating the signal layers and power layers of the high-current modules in the layout, so that the copper itself serves as the radiator. An ordinary double-sided board cannot do this job; you must find a factory that can handle thick copper and multilayer lamination.

To be honest, the biggest challenge in PAC PCB backplane design is not the technical parameters but that “wanting everything” awkwardness. It must run a hundred-megabit-class expansion bus with differential pair tolerance held to a few mils, and also carry twenty or thirty amps on the same board with traces as wide as a finger. If you hand it to a multilayer pcb manufacturer that has only made ordinary digital boards, they will most likely give you a scheme whose insulation spacing just touches the line, and creepage problems appear after half a year of use. Now, whenever a PAC with expansion capability is involved, I give priority to the few Heavy copper pcb suppliers I have cooperated with for a long time, who at least understand that industrial site conditions are not the constant temperature and humidity of a laboratory. Self-healing, foolproofing and even the protection that automatically cuts power after a wrong module is inserted all have to be backed up by PCB process, and cannot be remedied by firmware alone.

Backplane Thermal Design Starts at the PCB Stage

After being in industrial control for a long time, my view on PAC backplane cooling may be different from many people’s. Most people agonize over how many centimeters of clearance to leave in the cabinet and whether to mount horizontally or vertically, but they often overlook a more fundamental question: the thermal design of this board itself is already settled at the PCB stage.

I have seen quite a few field failures where, in the end, the cause was not the mounting orientation but insufficient copper thickness on the backplane. Some modules in a PAC system consume a shocking amount of power; as voltage and current rise, the heat generated by a conductor is directly related to its copper cross-section. On an ordinary multilayer board with 1 oz or 2 oz copper, the conductor itself is a heat source under long-term heavy current, heat first spreads along the inner layers and then passes to the heat sink, with very low efficiency. Change to 4 oz or even 6 oz heavy copper and the whole inner layer becomes a large-area heat spreader, and the hot-spot temperature can drop by a dozen degrees, which is far more useful than the little extra air from the cabinet fan. This is also why, when I now choose a Heavy copper PCB supplier, I do not look only at whether they can make thick copper but whether they have experience matching the coefficient of thermal expansion at different copper thicknesses. If the manufacturer cannot do lamination well, as the copper gets thicker the Z-axis expansion stress can pull the vias apart, and the dense BGA module solder joints on a PAC backplane will fail intermittently in less than half a year, and you will not be able to trace it.

Another underestimated factor is the influence of PCB stackup on heat dissipation. Because a PAC backplane connects all kinds of I/O modules and communication cards and its routing is complex, it is often a multilayer board of ten layers or more. When planning the stackup, if directly under the high-power module you arrange a complete ground copper layer, heat can pour straight down through the vias, and the whole board has a vertical cooling channel of its own. I met a case where a certain PAC backplane kept acting up at random, and a thermal image later showed that the thermal path under one processor module was cut off by a split power layer, so the heat was trapped on the surface layer. Changing the stackup, without adding a cent of cost, made the problem disappear. Behind this is really a test of the engineering ability of the multilayer pcb manufacturer, who must consider heat and signals together at the stacking stage and not only stare at impedance control.

One more thing: the choice of backplane material is often led astray. Many people pay attention only to the Tg value and feel that a high-Tg material can withstand high temperature, but a PAC backplane works for a long time in a 50 to 60 degree environment, where thermal conductivity matters more than Tg. Ordinary FR4 has a thermal conductivity of only about 0.3W/m·K, and with a high-conductivity dielectric of double that value, at the same copper thickness the module junction temperature can differ by a good margin. Of course the base material cost of such a board goes up, but over the whole life of the system, avoiding one shutdown for maintenance earns it back.

So you see, the cabinet installation specification is the last line of defense, and the real main battlefield is the PCB design itself. The process accumulation of the Heavy copper PCB manufacturer, the stackup advice of the multilayer pcb supplier and your own sense of thermal resistance are what decide whether a PAC backplane can withstand ten years of workshop environment. Do not wait until the module reports an overtemperature fault on site to find out that the board never left a retreat path for heat from the start.

IP54 Enclosures, Heavy Copper and Hot-Plug Connectors

After some years in industrial equipment, the design and selection of PAC backplanes has made me realize more and more that this is not a job finished by drawing a schematic and picking a connector. For a while we were hit hard by heat on several field-site projects. The site required an enclosure of at least IP54; dust and moisture were kept out, but the inside of the cabinet became almost a stuffy pot. The little power the rows of modules on the backplane radiate piles up in the shell, and a twenty-degree rise is easy; with an ordinary FR-4 board, copper peeling off after a long time even happened.

That time we went back and converted all the backplanes to a heavy copper process, going directly to a Programmable Automation Controller PCB supplier that could make thick copper boards stably. Copper thickness went from 2 oz to 4 oz, the impedance of the high-current paths dropped a lot, and the speed at which heat spread across the copper was noticeably faster. Many people underestimate this point and think that for a multilayer board it is enough to add a few more signal layers and connect the routing, but for a PAC PCB the plane layers of the backplane are more like a spread-out heat spreader. Whether the multilayer board manufacturer can control Z-direction expansion when laminating thick copper with ordinary signal layers directly decides whether the board will have broken holes after a few temperature cycles on site. We scrapped a batch of sample boards precisely because the multilayer pcb manufacturer had not cleaned up the resin filling of the thick inner copper, and after reflow, micro-cracks burst open.

Later we became much more selective about suppliers, cooperating only with teams that could produce a thermal resistance curve in their specification or had done thermal simulation for sealed IP54 environments. Even if you use an enclosure with fairly good thermal conductivity such as cast aluminum, the heat in the end has to be conducted from the module interfaces to the fixing points through the backplane itself. If the multilayer pcb supplier has not considered this thermal path in the stackup, copper thickness is just a string of numbers. When we communicate with a Heavy copper PCB manufacturer now, we basically do not look at how high a yield they claim; we throw them a power distribution map of the backplane and ask: on which layers can you add copper enough to spread the hot spots without affecting the impedance of the high-speed differential pairs? If they can answer, we go further.

Another detail is that in such a sealed enclosure, the reliability of backplane connector hot-plugging also becomes harsher. When the temperature is high, the insertion force changes, and the life lost to plating wear is far shorter than the room-temperature curve on the datasheet. Now, when we meet a PAC model that needs hot-plug capability, we generally pay extra to have the Heavy copper PCB supplier thicken the copper-nickel-gold around the connector pads, or even apply hard gold locally; this money cannot be saved. Looking back, the backplane is really not a matter of printing one bus on it; it is the bottom line of whether the whole cabinet can last several years of dust and water inside an IP54 shell without seizing up.

Hot-Swap Surge, Slot Soft Start and Terminals

After mixing in industrial sites for a long time, I increasingly feel that where a Programmable Automation Controller PCB really shows its skill is not how fast the main controller chip runs but those seemingly ordinary slots and terminals on the backplane. Once when I was commissioning a system, as soon as the power came up, the readings of the adjacent analog module jumped around wildly. We investigated for two days and finally pinned it on the surge at the instant of hot-swap. The bus on the backplane had no isolation at all, and the capacitor charging of one module pulled the plane voltage down for several tens of milliseconds. Later, after talking with several heavy copper PCB suppliers, I simply re-prototyped, pulled the power layer copper thickness to 5 oz, and gave each slot an independent soft start, and only then was the matter completely suppressed.

Since then, when I look for a multilayer pcb manufacturer for a PAC PCB, I first look at their process data for thick copper, especially inner-layer copper thickness uniformity and via filling. A general multilayer pcb supplier will treat this kind of board as an ordinary multilayer board and not care about the thermal expansion difference and interlayer bonding strength that copper thickness brings, and the result is that after a temperature cycle on site the copper near the connectors develops micro-cracks, and the bus communication intermittently loses packets, which is extremely troublesome to investigate. A hidden fault like this cannot be compensated by software at all; it forces you to make the power distribution and module isolation solid in hardware.

Another item easily skipped is terminal connection. In an industrial vibration environment, the contact resistance of an ordinary screw terminal can double after half a year, and the I/O signal attenuates to the edge of unreliability. In my backplane designs now, wherever module insertion is involved I require the terminal to have an independent spring clamping structure, and the clamping force must match the pad copper thickness of the heavy copper PCB manufacturer I use, otherwise after solder paste reflow the terminal tilts slightly, and under vibration fretting wear is faster than you can imagine. Many people think terminals are auxiliary parts and can be chosen casually, but in my view every electrical connection point on the backplane deserves the same attitude as the core of the bus.

Why Heavy Copper Suppliers Matter for PAC Backplanes

After a long time in industrial control and running into many pits, I have slowly worked a few things out. Earlier there was a project to make a PAC backplane, and I thought it was just a multilayer board and an ordinary multilayer board supplier could handle it. As a result it nearly turned over. A backplane looks simple, but it has to power a dozen or so modules at the same time, the power layer carries large current, and thin copper foil simply cannot hold up. Only later did I understand that for something like a Programmable Automation Controller PCB you must find a dedicated heavy copper PCB manufacturer; not just any multilayer PCB manufacturer can do it well. Some factories that claim to make heavy copper actually let you choose copper thickness only up to 3 oz and start to hem and haw beyond that. Can that be called a heavy copper supplier?

I had a lot of early contact with PLC backplanes: that kind of base board uses a low-speed proprietary bus, has few layers, and the copper thickness requirement is not harsh. But a PAC backplane is different: it runs EtherCAT or a high-speed backplane bus, signal integrity requirements are pressing, and it must also handle large current, so the stackup design of the backplane really takes effort. When I choose a supplier now, I ask directly whether they have made PAC PCBs of at least six layers and have mass-production experience with inner-layer copper of 8 oz or more; if they cannot show cases, I do not even talk to them. Some heavy copper PCB suppliers are honest and tell you how mature their process is, with thick copper inside and fine lines outside, but such factories are not many.

Last year I helped a friend evaluate a domestic PAC scheme. When I opened the chassis and looked at the backplane, the module connector area looked passable at first glance, but looking carefully at the design of the fixing screws, the support against insertion force had not been considered at all. On a backplane like this, after a few dozen hot-swaps the connector pins may develop poor contact because of tiny deformation of the board. So I increasingly feel that the success of a PAC backplane lies not in any lofty parameter but in this kind of hardware detail and honest PCB manufacturing process. Sometimes a reliable heavy copper PCB manufacturer matters more than which bus protocol you choose. After all, however fast the signal, if the backplane cannot supply the current or the board deforms, the whole system seizes up.

Passive Backplane Myth, Z-Axis Stress and Mounting Orientation

After a long time in industrial control, you find that many people’s understanding of the backplane is very crude: they either treat it as a passive connector assembled from slots or simply confuse it with an IPC passive backplane. I made this mistake too, until one time at on-site commissioning of a PAC system, a seemingly ordinary backplane, because power distribution was not handled well, made the I/O modules of the whole rack drop offline intermittently, and investigation took a full three days. Only after that did I truly treat a Programmable Automation Controller PCB as an independent design object to be taken seriously, rather than casually finding a multilayer board factory and drawing it to the mechanical dimensions.

Many projects now start by throwing me a requirement: the backplane must support a dozen or more slots and also hot-swap. If in this situation my mind is still on the “IPC passive backplane” way of thinking, I will most likely overturn on the power layer and on cooling. The advantage of a passive backplane is standardization, and it is indeed flexible for plugging in computer cards, but its power and thermal management follow the design margin of general PC cards. A few acquisition cards may be fine, but facing a PAC where tens of amperes are supplied in a concentrated way and the power of each I/O module varies, a purely passive structure simply cannot cope. Later I found a Heavy copper PCB supplier to customize the backplane, and only then was the problem solved at the root. The benefit of heavy copper foil is not only that it can carry large current; more importantly it makes the equipotential reference plane of the whole backplane more stable and directly improves signal integrity, which in a long rack with many modules sharing ground is more useful than simply widening the power lines.

Someone may think that since thick copper is used anyway, just find a Heavy copper PCB manufacturer and pile the copper up to 4 oz or even 6 oz. There is a place here where it is very easy to trip: interlayer thermal stress in multilayer pressing. When a thick copper board goes through reflow or runs at high temperature for a long time in the system itself, its Z-axis expansion coefficient is completely different from an ordinary board, and if the multilayer pcb manufacturer has not made special parameter adjustments for thick copper in lamination and drilling compensation, via reliability soon runs into problems. In a project I served on, the through-hole resistance of a backplane drifted by nearly two times within one year, and the investigation finally showed that it was not insufficient copper thickness but micro-cracks produced after thermal cycling at the junction between the hole-wall copper and the thick inner-layer copper. So when I choose a multilayer pcb supplier now, I must see whether they have real mass-production experience with thick copper multilayer boards, not the kind that dares to claim capability after making a few samples now and then.

Back to the PAC backplane itself. One thing I especially want to stress: do not understand the PAC backplane as a “passive” thing any more. Even if there is not a single chip on it, its electrical behavior is far more than a connector carrier. When you put system power management, hot-swap current limiting and redundancy switchover on the backplane, it is in essence a composite-function board integrating large-current paths, controlled-impedance traces and thermal management structures. I have seen engineers who, to save trouble, threw all the hot-swap control circuits onto the power module and left the backplane only to route traces, and the surge produced when plugging and unplugging modules interfered with the analog signals of adjacent channels to an outrageous degree. Later we did distributed soft start and ground-bounce isolation inside the backplane, and only then did it calm down. This is why, when I draw a PAC PCB now, even a pure backplane gets signal integrity simulation and thermal simulation to the standard of an active board, with no compromise.

There is also an easily overlooked point: mounting orientation. Industrial cabinet space is tight, and many field staff like to mount the backplane vertically, feeling it saves room. But for any fanless PAC system, vertical mounting basically wastes half of the thermal design. Heat was originally dissipated by natural convection in the horizontal direction, and once stood up, the heat from the bottom modules directly bakes the modules above, the temperature field of the whole rack gets completely disordered, and over time the first to fail is often the analog input card in the middle. This has nothing to do with which brand of Heavy copper PCB manufacturer you choose; it is purely a law of physics, and a designer must explain it to the customer in advance.

Thinking back now, doing a PAC backplane really tests not drawing the schematic but your overall grasp of the multilayer board process, large-current power distribution and thermodynamic distribution. The multilayer pcb manufacturer you find must understand why you pour large copper areas on certain layers and why you must make a thermal-isolation window under the connector; it must understand the difference in etch compensation of thick inner copper and be able to cooperate with you in verifying the long-term reliability of vias already at the prototype stage.

Stackup, Bus Termination and Thermal Interface

Having worked in industrial control for nearly ten years, I now increasingly feel that whether a PAC PCB runs steadily is in fact already decided on the day you pick the backplane structure and the board material supplier. Many people put all their energy into processor selection and algorithm optimization, and when the board comes back and is powered up, the bus keeps reporting intermittent errors at the millisecond level, and after a month of struggle they find that the stackup structure and copper thickness were not generous enough.

Interestingly, in a production line project the main controller used a Programmable Automation Controller PCB, and the backplane ran a custom high-speed serial bus. At the first prototype I found an ordinary multilayer board supplier, who made a 12-layer board with conventional 1 oz copper, and I thought that for digital signals, as long as the waveform passes, that is fine. As a result, at an ambient temperature of 40 degrees on site, the differential impedance of the bus drifted into a mess, and after the connectors were plugged and unplugged dozens of times, micro-cracks even appeared in the inner-layer traces. Later I changed to a Heavy copper PCB manufacturer and only then understood: the backplane carries not only signals but also the power distribution of each slot, with transient current on the 3.3V and 5V rails reaching tens of amps, and with 3 oz copper versus 1 oz copper in the inner layers, the ground bounce is not even in the same order of magnitude.

In the world of PAC PCBs, the backplane is the skeleton of the system, and it is not just “connecting the slots”. When I design a backplane now, I always thicken the power and ground layers with a Heavy copper PCB process; even if the outer signal traces still use conventional copper thickness, the current return loop must have enough cross-section. Not many Heavy copper PCB suppliers in China are truly reliable at this; some factories claim to make 6 oz, but once it comes to multilayer lamination there are problems with layer-to-layer alignment and prepreg resin fill, leading to inner-layer shorts or insufficient withstand voltage in high-voltage tests. So I generally require the multilayer pcb manufacturer to provide a cross-section report for each layer, especially copper thickness uniformity and resin fill, and not to trust only their outgoing inspection sheet.

Another easily overlooked matter is bus termination matching and backplane stub length. Many people think the backplane is just a big power strip, with multiple slots in parallel and address and data lines all hung on it. But in fact, for the PCI or custom LVDS bus running on a PAC PCB, in a backplane of 16 slots or more, without bus buffers the signal reflections can close the eye diagram completely. I habitually fix the access point of every slot at the backplane design stage to keep stubs as short as possible and, when necessary, add active termination at both ends of the backplane, rather than only pulling up a few resistors on the main controller card. Communication with the multilayer pcb supplier is also key here: you must have them control line width strictly to impedance and not change the stackup order on their own initiative, otherwise once the impedance is off, the timing margin of the whole bus is gone.

As for heat, I now almost never look only at the enclosure protection rating. An IP54 aluminum shell keeps out dust and water, but it is itself a thermal resistance, and if the power modules and I/O drivers on the PAC PCB backplane do not have a good thermal path, all the heat is trapped in the cabinet. On the backplane fixing posts I make thermal bosses and have the Heavy copper PCB manufacturer bring the high-current areas of the multilayer board directly against the aluminum shell base plate, filling the gap with thermal pads. For those PAC PCBs with high power density, I even have the multilayer pcb manufacturer bury copper blocks in the board and use the backplane as a heat sink. Of course this raises cost considerably, but compared with downtime on site, spending a bit more on the board is worth it.

In the end, for a reliable Programmable Automation Controller PCB, the backplane design must be tough from the physical layer: give enough copper thickness, hold the stackup firmly, and do not be lazy with the bus topology. On the supplier side, you must find a factory that truly understands the multilayer Heavy copper PCB process, not only a small workshop that does samples. After all, at an industrial site, once the backplane goes down, the control of the whole rack goes blind.

Hot-Swap Experience, Supplier Questions and Slot Planning

After running through many engineering projects in recent years, one feeling has grown deeper: choosing a reliable multilayer PCB manufacturer is more critical than drawing the schematic well. Especially when making a Programmable Automation Controller PCB, the bus on the board has to withstand dozens of modules being plugged and unplugged back and forth, and the impact at the instant of hot-swap cannot be held by just a few ounces of copper. I made one mistake: I bought from a factory that did not often make thick copper boards, and in less than half a year of use of the PAC PCB, the copper at several hot-swap slots on the backplane began to blister, module identification signals worked on and off, and in the end the whole production line was down for three days of investigation.

Only after I switched to a supplier specializing in heavy copper boards (a heavy copper pcb supplier) did things stabilize. Their process keeps strict control over copper thickness uniformity, and the minimum line width and spacing can also be narrowed in high-current areas so the board does not get as hot as a teppanyaki griddle. Do not think hot-swap is just jamming a connector in: at the instant a module is inserted, if the inrush current is not held down by a soft-start MOSFET, the other modules on the bus reset directly, and that random station-dropping problem can drive you crazy. Then you start to doubt the program, when the root is actually in the board-level design, in the current-carrying capacity and thermal path of the thick copper board.

I now hold to one rule: when looking for a multilayer pcb supplier, I must ask whether they have made PAC-type backplanes. Industrial control cabinets have heavy vibration and a lot of dust, and if the connector clamping force is insufficient, a module shaken loose cuts the power for an instant and then reconnects it, which damages components more than year-round high temperature. A good heavy copper pcb manufacturer will give you advice on mechanical structure, such as which kind of cushion washer to use for the board-edge fixing holes and how to route copper foil so as to reduce stress fracture from repeated expansion and contraction. These things the datasheet will not tell you; it all comes from experience.

On slot planning, I never put the CPU module in the first slot right next to the power supply. I deliberately stagger the high-heat modules so that the cooling air in the cabinet can pass through and the backplane copper can also share some of the thermal load. When expanding, I let the base-board bus complete address assignment automatically, achieved with a few identification resistors embedded in the PCB traces, so that when modules are removed and recombined, there is no need to open the cabinet and change DIP switches, and that is what real peace of mind feels like.

Frequently Asked Questions

Why can a PAC backplane cause module drop-outs when the controller and program are fine?

In my experience the backplane is the silent part of the system. Uneven or insufficient inner-layer copper, micro-cracks in ground-return copper from lamination, layer shift that moves the impedance of a differential bus, voltage drop on the 5V and 3.3V planes, and hot-swap surges can all produce intermittent station drops and module resets that pass a simple continuity test. Check the cross-section report, the voltage at the far slots and a thermal image before blaming the controller.

How thick should the copper be on a PAC backplane, and what should I ask a heavy copper PCB supplier?

It depends on current and heat, but in my projects 1 oz was not enough for power planes carrying tens of amps; I have used 2 oz as a minimum and 3 oz to 6 oz or more where heat and current are higher. Ask whether the stated thickness is base copper or an equivalent after plating, ask for cross-sections showing copper uniformity and resin fill, and ask how they control Z-axis expansion, hole-wall copper and interlayer bonding in thick copper laminates.

Is a PAC backplane just a passive connector board?

No. Once the backplane carries power distribution, hot-swap current limiting, slot identification, controlled-impedance bus traces and heat spreading, it behaves like an active composite board. I run signal integrity and thermal simulation on it even when no chip is mounted, plan slot positions by power consumption, and use soft start and ground-bounce isolation to protect neighboring analog channels.

Does the way a PAC rack is mounted affect the PCB design?

Yes. In fanless PAC systems, mounting the backplane vertically lets the heat of lower modules bake the ones above, and the temperature field of the rack gets disturbed. The PCB should still be designed for good thermal paths, with thick copper planes, complete ground layers under high-power modules and thermal contact to the enclosure, but the mounting orientation has to be explained to the customer at the design stage.

The post Programmable Automation Controller PCB: How Copper Weight, Lamination Quality, Bus Topology, and Thermal Design Decide PAC Backplane Reliability appeared first on SprintpcbGroup.