Baseboard Management Controller PCB Design and Sourcing: HDI, Multilayer, and Heavy Copper Lessons from the Field

2026-10-20
sprintpcbgroup

Why the BMC PCB Deserves High-Speed Board Standards

After so many years in server hardware, what gives me the biggest headache is not the CPUs that draw hundreds of watts, but the palm-sized BMC PCB. You might say it is unremarkable, yet it is the only entry point for remote management of the whole machine, and once it misbehaves, operations will blow up your phone. Not long ago I was debugging a storage node. Because the PCIe Gen3 routing on the BMC side had no controlled impedance, the firmware upload always died halfway. After three days of investigation, the cause turned out to be that the multilayer pcb manufacturer had carelessly cut away the reference layer, and the differential pair impedance drifted straight to 110 ohms. So I learned my lesson: for any design with a BMC, board material and factory selection must follow high-speed board standards. You need a factory that has genuinely built as a high speed PCB manufacturer, not one you judge only by the certifications displayed on its website.

Dense Layouts That Call for an HDI Manufacturer

Last year there was a small edge computing box with extremely tight space. The entire BMC module was under forty millimeters square, and the DDR chips had to be stacked on the back of the BMC chip in a POP structure, with line width and spacing in the fan-out area pushed below three mils. An ordinary multilayer pcb supplier simply could not handle this. They could not even keep laser drilling alignment stable, let alone layer-to-layer registration. I went back and forth with purchasing for several weeks and finally switched to a team that specializes as an hdi pcb manufacturer, and we stacked a four-step any-layer interconnect to fan out all the signals. But HDI factories have a common weakness: the moment you mention thermal copper thickness, they shake their heads. The BMC itself does not consume much power, but once the heat radiated by the whole server board arrives, the temperature of the few MOSFETs at the edge of the module can soar to eighty degrees. Ordinary one-ounce copper cannot take it, and you have to go with heavy copper. So ever since, I have basically settled on one factory for long-term cooperation that can act both as an hdi pcb supplier and as a heavy copper pcb supplier, so I do not have to switch around and delay the schedule.

Management Buses: Slow Signals, Real Pitfalls

Routing for management buses such as I2C and SPI looks slow, but it has plenty of traps. Once I failed to give the I2C its own ground shielding, and it picked up crosstalk from the neighboring fan PWM signal. As a result, the BMC read a wrong fan speed pulse every dozen seconds or so, and the log was full of alarms even though the fans were actually spinning steadily. Software troubleshooting has no way into this kind of problem, and the only fix was to respin the PCB. So my habit now is that, however crowded the board is, the BMC’s low-speed signals such as I2C, SPI and GPIO are routed on inner layers of the BMC PCB as much as possible, leaving the outer layers for high-speed signals. A multilayer pcb manufacturer with real experience must also evaluate the stackup. You cannot hand it casually to a prototype shop that only cares about connectivity rates.

The BMC as the Server’s Steward

Put plainly, the essence of this small BMC board is that it is the server’s steward. A pile of sensors and buses hang off it, and signal quality cannot be compromised at all. Today’s new BMC chips, such as ASPEED’s AST2600 or Kunpeng’s Hi1711, often come with multiple gigabit network ports and a PCIe 2.0/3.0 uplink, and the substrate layout requirements are already about the same as the north and south bridges of ten years ago. If you still select parts with the old mindset, looking only at layer count and not at measured impedance, looking only at copper thickness and ignoring resin flow, then in the end you are the one who suffers. After this long in the industry, I feel the most valuable thing is to accumulate two or three truly reliable suppliers: ones that can handle high-density interconnect, handle local heavy copper, and are willing to tune stackup parameters together with you. That matters more than saving money.

BMC PCB Is Not as Simple as It Looks

Many people think the small BMC circuit board has little technical content, that it is just an ARM-based management chip with a few I2C lines. That was my attitude when I first drew a BMC module. When the board came back and was powered on, there were so many DDR read and write errors that the firmware engineer came to me for a life talk.

After several rounds of revisions I found that the BMC itself may look unremarkable, but that chip packs in a lot. The DDR memory controller is not at all forgiving about impedance and length matching. The space for BMC DDR routing is usually very narrow, so you cannot route it beautifully the way you can on a large board. You must control length matching and impedance within a small area, while a pile of low-speed management buses sits alongside. The I2C devices hang on one by one, often a dozen or so sensors, plus the power modules. If bus capacitance is slightly out of control, an ACK will occasionally be lost, and troubleshooting can take a whole day.

What is even more troublesome is that many BMC modules are soldered directly onto the motherboard using SODIMM gold fingers or high-density BGA, leaving really cramped space for the PCB. At that point, an ordinary double-sided board factory is definitely not an option. The hdi pcb manufacturer I later settled on for long-term cooperation handles this kind of high-density interconnect board smoothly, and their laser blind vias and fine-line yields are stable. Also, the Baseboard Management Controller PCB often sits in the server’s standby power domain. Even when the system is powered off it has to stay awake, so over time its heat output is not small. In a spot with a poor airflow path inside the chassis, the heat near the BMC chip cannot get out. At best the network port drops, and at worst the machine’s remote management stops working intermittently.

Thermal Planning with a Heavy Copper Partner

So for heat dissipation, I started paying special attention when dealing with a heavy copper pcb supplier: increase the copper thickness of the power layers, and connect the heat-spreading copper directly to the module’s mounting bracket. Some designs even need a thick copper board for independent heat dissipation, and the process of the heavy copper pcb manufacturer determines whether heat can be conducted down smoothly. Many people think it is just a matter of adding thick copper, but in practice, balancing thick copper against fine, dense traces is hard. If it is not done well, the etch factor makes your traces thinner, and high-speed signal quality drops with it.

I remember one revision very clearly. To save cost I switched to another multilayer pcb supplier. They were passable on ordinary 4-layer boards, but on the BMC’s 6-layer one-step HDI the impedance fluctuation was outrageous, and the eye diagram of the USB 3.0 differential pair could barely open. After I found a knowledgeable multilayer pcb manufacturer to redo the prototype, using high-speed material and carefully controlling the stubs of the differential vias, the problem finally calmed down. So choosing the right supplier is more important than drawing the right schematic, especially when you need to satisfy two seemingly contradictory demands, high speed and heavy copper. A reliable high speed PCB manufacturer can get the stackup design and process matching right, so that the BMC board does not become the most fragile board in the whole server.

Looking back, the BMC PCB is not a simple low-speed board, nor is it a pure extreme challenge on par with a high-speed backplane. It is more of a jack of all trades that has to take care of everything, with small area and demanding thermals. Before every layout, I first route the DDR groups well, then squeeze power and buses into the remaining space. When things truly cannot fit, I decisively push the board to an hdi pcb supplier and use blind and buried vias to claim space from the inner layers. Get this step right and the later debugging goes much more smoothly, and remote management will not raise an alarm and call you out in the middle of the night.

Form Factor Decisions Shape the Whole Supply Chain

Some time ago, a friend who does server integration complained that the BMC on one of their dual-socket motherboards kept dropping offline for no apparent reason. After a long investigation, it was traced to oxidation of the gold fingers on that tiny BMC module. Because of this trivial matter, a whole batch of machines sat idle in the data center for three extra days. It reminded me of all the BMC solutions I have handled over the years. Whether it is a BGA part soldered directly to the mainboard or a small card shaped like a SODIMM, the choice of physical form seems on the surface to be a matter for structural engineers, but it actually involves the entire manufacturing chain, especially PCB fabrication and assembly.

Many people think the Baseboard Management Controller PCB is just a low-speed mixed analog and digital board where you can route traces casually. Once you actually do it, you know it is not like that at all. Today’s BMC chips run at no low clock speeds. Parts at the level of ASPEED 26xx and above need external DDR4, some connect eMMC, and they must reserve NC-SI and a bunch of I2C lines routed off to the mainboard. If you make it an independent 30x45mm small board and also have to fit the PHY and several DCDC converters on it, the routing density heads straight into HDI territory. Last year, on a domestic-substitution server project using a self-developed BMC, I initially wanted to save effort and prototype the module on an ordinary through-hole board. The impedance could not be controlled, and the DDR eye diagram was a mess. In the end I honestly found an hdi pcb manufacturer to redo the stackup, added one step of laser vias, and squeezed the critical signals onto inner layers, which pulled signal quality back. So do not underestimate this small BMC PCB: once its form is chosen, the range of downstream manufacturers you can work with is basically decided. Not just any multilayer pcb supplier can take this kind of job.

Choosing Suppliers: Few Qualify for Dense, Mixed-Copper Boards

Speaking of suppliers, there are plenty of multilayer board factories in China, but when you simultaneously demand high-density routing, impedance control and mixed-material matching, there are really only a handful that qualify. I have also run into one that advertised itself as a high speed PCB manufacturer. When our engineer asked whether they could do 0.2mm BGA fan-out, they asked back why we were not using a 0.3mm pad. A conversation like that can really raise your blood pressure. Ultimately, although the BMC in a server does not have dozens of Gbps of SerDes the way a CPU or switch chip does, its packaging form determines the PCB manufacturing threshold. For example, a BGA part mounted directly on the board looks traditional, but to guarantee board-level reliability you have to haggle repeatedly with the board factory over pad surface finish and via-fill resin. If there is even a slight void, pads lift off when the BGA is reworked. Even more so, in some scenarios the small BMC board has to be combined with a heavy copper power zone, and then you have to deal with a heavy copper pcb supplier. Power layers are often 3oz or 4oz copper, while the BMC area needs fine-line routing. Heavy copper pcb manufacturers that can do a good job on this kind of mixed-lamination stackup are truly rare.

The word packaging is quite easy to confuse. At the chip level there are BGA and QFN, and at the board level there are module and on-board. My habit is to separate the chip package from the board-level integration method. Whatever package your BMC chip uses, in the end it has to exist on the server mainboard in some physical form. Once that form is decided, the supply chain is decided. If you choose a SODIMM module, you depend on an hdi pcb supplier that can make tiny multilayer boards and control outline tolerance with high precision, and you must also make sure the gold finger chamfer and plating thickness are held tight. If you choose LGA or BGA directly on the mainboard, the requirement on the multilayer pcb manufacturer for that main board goes up a level, because the mainboard has many layers and a large size and must also accommodate other high-speed and heavy copper areas, so producing a high-yield board tests overall capability. I have seen a project that, to save cost, changed a formerly independent BMC module back to on-board, and the mainboard had to add two layers because the BMC routing area was too dense. In the end the total cost was higher and delivery slipped by six weeks. This kind of pit can be avoided completely by evaluating the packaging scheme early.

Ultimately, what a server needs is long-term stable operation, and the BMC is the lifeline of out-of-band management. When it fails, you do not even get a chance to remotely view the logs. So when discussing the form of the BMC PCB, never look only at signals and space. From the manufacturability of prototype routing, to soldering yield in volume production, to the convenience of disassembly during later maintenance, every link puts the question to your original decision about how to build this board. When you have broken all these factors down and thought them through, you naturally know what kind of board factory and supplier to look for, rather than waiting until the SMT line is full of problems and then going back to change the design.

Why I Prefer an Independent BMC Module

A few years ago I was still torn between soldering the BMC directly on the mainboard and making a separate card. Now, if anyone recommends the former to me, I will probably just shake my head. High integration sounds like it saves space and cost, but once the chip has a problem, what you replace is the whole board. Having experienced the trouble that this depth of coupling brings once, you really learn your lesson. So in the projects I have handled in recent years, almost all go the independent module route. Even if it costs a few more PCB prototypes, it is worth letting the Baseboard Management Controller PCB stand on its own.

The most direct benefit of an independent BMC PCB is that switching and upgrading do not require touching the mainboard. A small board can be forty or fifty millimeters square with a thickness controlled within a millimeter or two, plugged into a standard slot or made as a clip-on board. I am not afraid of small size. I am afraid that density goes up and manufacturing cannot keep up. For this kind of module, a multilayer board is the baseline, and I generally start at 10 layers, with differential pairs and inner power planes all squeezed into that small area. Finding a multilayer pcb manufacturer that truly understands BMC signal timing is worth more than anything. In the past, to save effort I tried a factory that made consumer boards, and its line width compensation and layer registration deviation scrapped two batches of prototypes outright. After I learned, I only work with a multilayer pcb supplier that can show similar cases.

When density goes higher still, sometimes you simply must use HDI. For example, if the board has to fit the main BMC chip, DDR chips, eMMC and a pile of filter capacitors, and also leave a gold finger area, ordinary through holes cannot route it. When I meet this kind of design, I confirm in advance with the hdi pcb supplier whether they can do second-order or even third-order stacked vias, and the minimum laser drill diameter. One supplier claimed to be HDI but in fact could only do first-order, and I crossed it off after a single evaluation. When selecting an hdi pcb manufacturer, do not just read the introduction on their website. Look at the cross-section reports they have actually delivered. For inner-layer connection reliability, photos are more honest than sales talk.

High-speed signals demand even more care. Interfaces between the BMC and the CPU or PCH, such as eSPI and RGMII, are extremely sensitive to impedance continuity when running, and if the back-drilled stub is even slightly long, the eye diagram closes to the point of being unwatchable. My own lesson is that the high speed PCB manufacturer must take part in the stackup design stage, rather than being thrown the Gerbers after they are finished. One time I followed the vendor’s simulation throughout and adjusted the copper foil type of the differential pair’s reference layer, which pulled the insertion loss back into the protocol window. Since then I always keep two or three high speed PCB manufacturers on hand that can truly do pre- and post-layout signal integrity simulation. It costs more, but it saves countless boards that would have died in validation.

There is another easily overlooked case: BMC modules with high-current PoE power or drawing power externally. The power plane must carry a dozen or even tens of amps, and the 1oz copper on an ordinary board is nowhere near enough. In this case I only look for a Heavy copper PCB manufacturer, with outer layers at 3oz or even 4oz, combined with a thick copper filled via array, so that the temperature rise can be held down. One time, the Heavy copper PCB supplier I used balanced the etching of thick copper and inner fine lines quite well, and the trace corner flatness was clearly better than other vendors. Those boards later ran a thousand-hour burn-in without a single power dropout. If you have a design like this, I suggest asking about the factory’s thick copper process capability in advance, rather than being told to widen spacing at the DFM review.

Ultimately, although the BMC is not the most cutting-edge topic in high-speed digital circuits, it is the lifeline of server manageability, and if this little board is botched, it directly raises on-site operations costs. When choosing PCB partners now, I do not care much about brand size. I look at their obsession with detail, such as whether the nickel-gold thickness of the gold fingers can be held between 0.3 and 0.5 microns. If the deviation exceeds 0.05 microns, insertion life drops. There is also the edge chamfer angle. Being off by a single degree makes the slot spring contacts bear force unevenly. All of these were settled by repeated fussing at the millimeter scale. Carrying a pair of calipers and measuring outline tolerance right at the incoming inspection area is a habit I have developed over the past two years. It is not necessarily dignified, but it is very effective.

Don’t Worship Impedance Numbers: Find the Right HDI Manufacturer

After working on BMC modules for so many years, my biggest takeaway is this: do not put too much faith in the differential impedance numbers in the design specs. Finding an hdi pcb manufacturer that can truly make a Baseboard Management Controller PCB is what matters. The ball pitch of that BMC chip’s BGA is mostly 0.8mm, and when fanning out, every trace must pass between pads. An ordinary multilayer board factory using mechanical drilling simply cannot fit it, and yield collapses halfway through via plugging. Earlier we wanted to save effort, finished the BMC PCB and handed it to a regular multilayer pcb supplier, figuring it was just low-speed DDR3 and that controlling the length matching of a dozen or so differential pairs would be enough. When the prototypes came back, the impedance of the DDR clock differential pair had drifted to 110 ohms. Powered on, it occasionally reached the operating system and occasionally stalled directly at the memory self-test. After two weeks of debugging we found that the factory had thinned the reference layer copper by a third.

Later we switched to a manufacturer that could genuinely do HDI. Right away they proposed changing the PCIe vias to a blind and buried via structure, cutting the stub length from 12mil to under 3mil, and the eye opening improved by two full grades. When they make high-layer-count boards, layer registration accuracy can be within 2mil, and they control the length difference between P and N of a differential pair to within 3mil in their own process, so I do not need to route a pile of serpentine lines in the layout. This made me fully understand that a reliable hdi pcb supplier can digest much of the anxiety on the design side, such as reference plane integrity and the glass-weave effect. Their process capability is there, so you do not have to treat the specifications you copied as iron commands.

The power section is even more obvious. A pile of decoupling capacitors must be laid around the BMC. The current is not large, but ripple is picky about layout, especially the 1V core supply. We tried having an ordinary Heavy copper PCB manufacturer make a four-layer board with 2oz copper. The inner layers had many lamination voids, and the transient voltage drop was so large that the chip reset. Afterwards we moved to a factory that specializes in heavy copper. They used 3oz copper directly on the inner layers and even reshaped my power islands, and ripple dropped from 80mV to under 30mV. My habit now is that, however simple the BMC board, as long as it carries high-speed ports such as PCIe or USB 3.0, I first ask the factory about its multi-step HDI capacity and copper thickness tolerance. That is more useful than agonizing over 85 ohms versus 100 ohms. Differential pairs are designed, but even more they are manufactured.

Material Matters More Than a Few Mils of Length Mismatch

After so many years working on BMC-related boards, I increasingly feel that everyone falls into a pit from the start: we keep agonizing over a few mils of error in USB differential length matching, yet nobody talks about the board material itself.

I have run into this too many times. The layout engineer spaces the USB 3.0 TX and RX far enough apart and the simulation looks beautiful, but when the board comes back, link training simply will not pass. Where is the problem? It is not your routing. It is that the multilayer PCB supplier gave you the most ordinary FR4, whose loss cannot be controlled at high frequency. Although the BMC itself does not run fast, once the high-speed interfaces on the board reach 5Gbps or 10Gbps, the board’s Dk and Df matter an order of magnitude more than your length matching. This is especially so for the stacked scheme where DDR is placed directly on top of the BMC chip, where the heat is trapped in the middle. If the factory you picked has no capability for high-layer-count mixed lamination boards, even slightly uneven inner copper thickness causes impedance jumps that cost you half your life.

Later I simply shifted the focus of cooperation to finding the right people. It is not enough to find any factory that calls itself a high-speed PCB maker. You have to see whether they have really made management boards like the BMC. The HDI PCB manufacturer I now work with permanently tuned the parameters three times for the laser blind via copper-filling process of BMC chip fan-out alone. This kind of thing cannot be learned from spec sheets, only from the factory’s case library. Then there is heavy copper. Many people think the BMC is just a low-power chip and does not need thick copper. But if your board carries 48V PoE or a high-current path for a GPU and your ground layer does not use more than 2oz thick copper, you should honestly find a supplier that specializes in heavy copper PCBs. You cannot imagine how strange the timing errors that ground bounce noise produces on your I2C bus can be.

On the USB side I also do not bother with whether D+ and D- traces are parallel. In a compact BMC layout, forcing every signal to be neat and orderly is simply unrealistic. My approach now is that for a low-speed differential pair like USB 2.0, as long as it does not run side by side with LC traces of a switching power supply and the impedance is near the nominal 90 ohms, it basically works stably. Where you really should put effort is the RGMII data and clock lines coming out of the BMC’s gigabit PHY, which should not be treated as slow interfaces. One time our board was dropping packets, everyone stared at the software, and in the end we found the multilayer PCB supplier had created a tiny dent in the reference layer when making inner copper thickness. Ground bounce plus signal skew ruined long frames. We switched to a factory that really understands high-speed link integrity and remade one revision, changing nothing else, and it was fixed.

So BMC PCB work looks like a pile of design rules but is actually a stack of process choices and supply chain trust. It is true that stacked packaging can save you a quarter of the board area, but you must confirm that your HDI PCB supplier can calmly handle the thermal vias under the BMC and the stress relief on both sides. This kind of work cannot be written into design guides, but it can be written into the number of boards you scrap.

The Hidden Costs of Stacking DDR on the BMC

In the years I have designed server boards, I have increasingly felt that stacking the BMC’s DDR memory chips directly on the chip is an idea that looks beautiful but runs into a pile of problems in execution. When calculating area at the design stage, it really is pleasing: the area occupied by BMC plus DRAM can shrink to the size of a palm, leaving room for other components. But when prototypes come back for debugging, firmware flashing and stress testing, all the problems surface.

Heat is the most direct pit. The BMC itself has low power, just a few watts, but with the DDR chip pressed on top it is like putting a quilt over it. Junction temperature that is more than ten degrees higher than when placed alone is common, and over long operation the reliability of the solder joints between the BMC and memory drops severely. The first time we built this stacked solution, we did not add underfill, and in the temperature cycling test, after a few hundred cycles there were signal bit errors. When we took it apart, the solder balls had cracked. Even after adding underfill, rework is a nightmare: if the memory fails you cannot replace it alone, the whole BMC module has to be scrapped, and cost goes up immediately.

I feel stacking should be approached very cautiously. Unless your structural design leaves no choice and the height limit is extremely harsh, I prefer to spread things apart and use HDI process to compress routing. After all, what a reliable hdi pcb manufacturer can do today in layer registration and microvia filling is already enough to keep BMC-to-DRAM high-speed routing within a reasonable length while preserving signal integrity. A multilayer pcb manufacturer I have worked with, which specializes in this kind of high-density server board, suggested I use multi-step blind and buried vias to put the BMC and memory routing on inner layers and leave a complete ground plane on the outer layers. That way EMC is easier to manage and heat can also dissipate.

Another easily overlooked place is the layout of the I2C management bus. On a BMC PCB, I2C often has to carry a dozen or so devices: temperature sensors, voltage monitors, power modules and fan control, all squeezed onto one or two buses. If bus capacitance is slightly large, the rising edge slows and communication drops packets occasionally. My habit is to place a dedicated I2C buffer near the BMC to cut the bus into several segments, and during routing to have the multilayer pcb supplier help with impedance control. Although I2C frequency is low, on long traces impedance continuity can avoid reflections. Once, on a board in mass production, the BMC could not read data from a voltage monitor chip, and it took two days to find that the bus branch was too long and the waveform was distorted. After adding a buffer in the next revision, the world was quiet again.

Power layer handling is also critical. Around the BMC there is a pile of low-speed chips, but the core supply requires very small ripple, especially the VREF supply for the DDR interface. I generally ask the PCB factory to give the power layer thick copper treatment: if 1 ounce is not enough, go to 2 ounces. Find a Heavy copper PCB supplier with a stable process that can guarantee uniform inner copper thickness without wrinkling during lamination. Some heavy copper PCB manufacturers can do local thick copper on the same board, handling the high-current section under the BMC while other areas keep normal copper thickness. That way the board is not too heavy, and voltage drop and heat dissipation on the critical paths are covered. After all, once the BMC supply is unstable, the entire system management is paralyzed, and seeing all sensors as question marks remotely from the data center is a helplessness I do not want to experience again.

Thinking back, BMC circuit design looks simple, but in fact it ties together the chip, PCB process and thermal design, and none of the links can fail. I am now more willing to spend extra time on planning at the initial layout and repeatedly discuss the stackup with the multilayer board supplier. I would rather spend more on prototype fees than force aggressive schemes that look like they save space. After all, a product that runs for three to five years without problems is worth more than anything.

I2C Bus Pitfalls and Why Factory Capability Decides Everything

When designing the Baseboard Management Controller PCB, the first thought in my mind is always that the I2C bus looks simple but is actually full of pits. Two wires, one for data and one for clock, in theory a graceful one-master multi-slave arrangement. But after you throw seven or eight temperature sensors onto the BMC board, add an EEPROM, and bring one more run to the power module, with traces stretching a dozen or so centimeters, the oscilloscope shows the waveform edges turned into slopes. Do not mention the 400pF bus capacitance upper limit to me here. Measured without a buffer, it easily exceeds 600. Many newcomers directly use multiplexers to switch channels. I prefer to break the problem down by physical location, moving devices with low response speed requirements into a separate I2C domain, with a level translator and isolation in between, keeping the main segment lightly loaded. Address conflicts are even more annoying. Some sensors have fixed addresses from the factory, and two of the same model on the bus fight with each other. Then you have to rely on BMC firmware to pull low an address pin on one branch during initialization before powering up and switching addresses. What you are playing with is the timing difference.

These ordeals made me deeply appreciate that a BMC PCB is not finished once the circuit is merely connected. The factory’s process capability directly determines how much room you have to tinker. In the past, to save effort, I used a multilayer pcb supplier with an attractive price. It was a six-layer board and promised trace impedance control of plus or minus 10%. When it came back, the measured I2C rise time had huge dispersion, one board passing the test and another failing. After a long investigation it turned out inner copper thickness was uneven and the common-mode noise of the differential pair had crept onto the bus. Later I switched to a manufacturer that began with high-layer-count boards. Although the market mainly regards them as a high speed PCB manufacturer, their control over batch stability of the substrate dielectric is exactly that strict. Even for a snail-paced signal like I2C, once the routing environment is clean, it is clean.

There is another case that is easily overlooked. A BMC board does not only run low-speed control. It often has to power the whole machine’s fan array, with instantaneous current of a dozen amps being common. If the multilayer pcb manufacturer you found cannot do thick copper, you have to pile on solder on the outer layer or add jumper wires, which is ugly, and at high current density it also heats up. A good heavy copper pcb manufacturer can make inner layers of 3 or even 4 ounces copper and laminate them together with high-speed trace layers without delamination, which depends heavily on accumulated process experience. Some factories claim to be a heavy copper pcb supplier, but when you ask, they will not take the order if HDI stacked vias are combined with heavy copper. BMC chips are mostly BGA packages, and fan-out must use laser blind vias, so without HDI capability you simply cannot play. My tendency now is to look for partners that have both hdi pcb supplier and heavy copper volume production experience, ideally with an engineering team that understands the meaning of pseudo-differential pairs and length-matching serpentines in BMC routing, rather than one that mechanically processes Gerbers. In the end, the BMC looks simple, but it is a role standing at the intersection of digital and analog, low speed and heavy current in a server. Get the board right, and half of those I2C bus troubles really go away.

Thermal Paths: When the Heatsink Is Not the Problem

What I fear most in this kind of board is thinking the problem is simple. I once followed a rack server case where the BMC module sat on an independent six-layer board, and the whole power budget had been calculated clearly. When the prototype came back and ran stress tests, the temperature curve shot straight up, and even full fan speed did not help. My first reaction was that the heatsink was not big enough, so I nearly doubled the fin height with no real change. Later I took it apart and measured the PCB temperature distribution on the back of the chip, and found that heat was never conducted out of the substrate at all and was all trapped around the few vias where the copper changed layers. After that I rethought the whole thermal path. It was not that the heatsink was not powerful enough. It was that the board’s own thermal channel had never been opened.

Looking at many BMC PCB designs now, they depend too much on the heatsink’s thermal resistance parameter. However low the junction-to-case thermal resistance, as long as there are extra interfaces on the path from PCB to heatsink, heat accumulates. My habit now is to put effort first into connecting the copper well, and not to be stingy with cost where thick copper should be used. Sometimes ordering a board with 2-ounce inner copper from a Heavy copper PCB supplier is far more effective than piling fins onto the heatsink later. This is especially true for a device like the BMC chip that is always online, with low power but absolute stability required. You cannot treat it casually just because it is only three or four watts. The direct consequence of heat that cannot be removed is not that it fails itself, but that the analog sampling circuits next to it get baked into drift, and the voltage and temperature readings of the whole machine float. When that happens, finding the problem takes much longer.

Choosing a manufacturer is also an unavoidable hurdle. A Baseboard Management Controller PCB may not be large, but its density is usually not low, with small BGA ball pitch and several differential pairs to route to the management network port. Some factories force the ordinary multilayer process onto an HDI design, and the expansion and contraction of mixed laser and mechanical holes goes uncontrolled. Once inner layer registration shifts widely, the impedance of the whole batch is scattered all over the place. Later I went straight to an experienced multilayer pcb manufacturer and had them join the stackup design early, putting several critical signal groups on adjacent layers to save a lot of back-drilling trouble. If routing really cannot fit and blind and buried vias are needed, you need an even more reliable hdi pcb supplier. Not every factory that claims to do HDI can control build-up layer uniformity to an acceptable level.

Thermal management, put plainly, is no deep theory. It is just a straight path from chip junction to air. Many people spend their energy choosing the thermal conductivity of the thermal pad, but the gap between 2.8 and 3.5, under uneven assembly pressure in practice, matters far less than whether the pad is cut properly and whether the mounting clip force is even. My habit is to first spread a thin layer of phase change material on the chip lid, then attach the thermal pad, then lock the heatsink with backplate screws directly onto the reinforcement plate reserved on the mainboard. A push pin method is not unusable, but over time it tends to loosen, especially after transport vibration, when clamping force drops to less than half of what the datasheet states. Heat cannot be transferred out, and the chip surface does not look hot, but the internal junction temperature has long since reached the throttling threshold.

Now when I meet a heavily loaded BMC design, I even consider treating the BMC chip directly as a low-speed device that needs active cooling. Rather than struggling over the heatsink form, it is better to put the resources of the high-speed PCB manufacturer to the right use. For example, have the high speed PCB manufacturer help control the board material’s glass transition temperature to ensure that under long-term operation the substrate will not deform from temperature cycling and crack the BGA solder joints under stress. If that cracks, adding a ten-watt heatsink is useless, and the signal is simply cut. In the end, the stability of a BMC PCB depends on board-level thermal and mechanical coordination, not on the few theta values in the chip datasheet alone.

Selecting a Supplier for a Stacked, Heavy-Copper BMC Module

I have recently been selecting a supplier for a BMC module, and the matter is much more troublesome than it looks. That small board, less than a palm in size, which we call the Baseboard Management Controller PCB, has to hold the chip, DDR chips and power, packed airtight. At first I thought any multilayer board factory would do, since multilayer pcb suppliers are everywhere these days. But when the prototype came back and ran DDR, there were problems, and signal timing was all over the place. Later I realized this thing must go through HDI process, with laser blind vias plus buried vias, otherwise the several dozen lines from the BMC to the DDR simply cannot be length-matched. In the end I went directly to a factory that works both as an hdi pcb manufacturer and specializes in high-speed boards, and had them control everything from the stackup to impedance, otherwise the crosstalk brought by stacked packaging could not be suppressed at all.

Another easy pit to step into is copper thickness. The power conversion section on the module draws considerable current, and ordinary thin copper simply cannot bear the voltage drop. But the board is thin and small, and heat dissipation is already tight. In the second revision I changed the power area directly to thick copper, which meant finding a reliable heavy copper pcb manufacturer that could make local copper thickness differences on one board without affecting drilling and pad flatness. That factory happened to also have heavy copper pcb supplier capability, and during prototyping they proactively reminded me that the thick copper area must not be too close to the BGA pads, otherwise it would absorb too much heat during reflow.

Heat dissipation was what gave me the biggest headache. The BMC chip and DDR are stacked together and the heat is trapped in the middle, so a heatsink on the surface cannot conduct it out. One time, at 55 degree inlet air, the DDR junction temperature almost broke 100. Scared, I quickly modified the board by embedding a copper block under the BMC, and had the high speed pcb manufacturer re-tune the stackup to make the thermal channel smooth. That taught me that for a high-density module like this, supplier selection cannot look only at whether they can do the work of an hdi pcb supplier. You also have to see whether they have thermal management engineering experience, and even whether they can make reasonable heavy copper zoning inside the board. Now when I see the market full of factories claiming to make BMC PCBs, the first thing I do is ask whether they have done projects with stacked DDR and local heavy copper. If they cannot answer, I turn and walk away.

Power-Up Sequencing, Clocks and Power Integrity

In BMC board work, many people start by staring at the dozens of pages of timing diagrams in the chip manual, as if soldering to the recommended values makes everything fine. I tend to think that what is truly hard has never been the waveform diagram, but whether the PCB factory in your hands can route the traces well. For example, the clock that provides a reference for the BMC: the manual says the 25MHz crystal load capacitance must match precisely. The problem is that if your HDI PCB manufacturer miscalculated a few picofarads of parasitics in the stackup, no amount of back-end tuning helps. So I later simply drew the clock routing as an island, with length controlled more tightly than 5mil length matching. It is not that I fear delay, but that I worry a jitter at the neighboring power switching node will make the whole BMC reset for no reason.

In earlier years I debugged a server mainboard that was stuck at the BMC startup stage for two full days. In the end I found that the 1.05V auxiliary supply came up less than 20 milliseconds before the 3.3V. At that moment I truly understood that so-called power-up sequencing is not a few rigid numbers in a design document but the first hurdle for whether a Baseboard Management Controller PCB can work properly. Many engineers treat sequencing as a process of ticking boxes against the module manual, but in reality, the same BMC chip placed on boards with different stackups and copper thicknesses has completely different power ramp slopes, and the window given in the manual cannot cover real-world tolerances.

Now whenever I handle a BMC PCB project, I have the layout engineer simulate the actual impedance of the power planes before prototyping, especially inconspicuous details such as trace width and via count for 3.3V and 1.8V. Some multilayer pcb manufacturers by default lower the inner copper thickness in their process parameters for yield. But once you meet a module that needs sequential power-up, that small difference is enough to make the reset signal and core voltage collide, producing that half-dead latch-up state, and even the emulator cannot catch the root cause. So I increasingly lean toward finding a supplier that can make Heavy copper PCB to handle the separate small BMC board, even if it is only a four-layer board, requiring 2oz copper on inner layers to minimize voltage drop in the power channel. This is not over-design. A data center operations site simply has no time budget for tuning sequencing board by board.

Speaking of board material selection, many people think the BMC is only a low-speed management chip and any cheap multilayer pcb supplier will do. But as soon as you have used the high-resolution VGA virtual media function, or run the long-distance routing of PCIe sideband signals from the BMC to the PCH, you will know that high-speed signal integrity can also overturn. I now basically work with one fixed hdi pcb manufacturer, having them use second-order stacked vias to shorten the routing from the BMC to the DDR4 chips to an acceptable length-matching range, while still shrinking the overall board size. This is actually more meaningful than finding a so-called high-speed PCB manufacturer, because the truly high-speed part on a BMC PCB is just a few differential pairs, and the key is impedance continuity, not how extreme the board material’s loss tangent is.

Returning to power, my approach is to change the sequencing control of the Vcore, VDDIO and VMON rails from a simple RC delay circuit to a small programmable sequencer chip, and to require the Heavy copper PCB supplier, when making the board, to route the feedback sense points of these rails directly to the load side rather than near the output of the power chip.