NVIDIA Blackwell PCB Materials: What Is Known About GB200 Boards and Laminates

NVIDIA Blackwell is NVIDIA’s GPU architecture for AI data centers, sold as the B200 GPU, the GB200 Grace Blackwell Superchip (one Grace CPU plus two Blackwell GPUs) and rack systems such as the GB200 NVL72, which links 72 GPUs at 1.8 TB/s each. NVIDIA does not publish the PCB materials or stackups of its boards. Boards that carry signals at these data rates generally need low-loss laminates in the class the industry calls M7 to M8, usually with smooth copper and backdrilled vias; that is an engineering expectation, not a confirmed NVIDIA specification. This guide separates what NVIDIA publishes from what analysts report, compares the laminate tiers on one test method, and lists what to specify for a board that has to work with Blackwell hardware. For laminate choices in general, see our PCB materials guide.

What Is NVIDIA Blackwell?

Blackwell is the GPU generation that followed Hopper in NVIDIA’s data-center line. NVIDIA packages it several ways, and the PCB questions differ for each:

• GB200 Grace Blackwell Superchip: one Grace CPU and two Blackwell GPUs connected by NVLink-C2C at 900 GB/s (NVIDIA multi-node tuning guide).
• GB200 NVL72: a liquid-cooled rack that connects 36 Grace CPUs and 72 Blackwell GPUs into one NVLink domain. Fifth-generation NVLink gives each GPU 1.8 TB/s, and the NVLink switch chips provide 130 TB/s across the domain.
• Blackwell Ultra (GB300): the follow-on version NVIDIA announced in March 2025, alongside its next architecture, Vera Rubin (CNBC report).

In the 72-GPU configuration, NVIDIA lists 18 one-rack-unit compute trays with 4 GPUs each and 9 NVLink switch trays (NVIDIA system guide).

What Is the Bianca Board?

“Bianca” is the name used in industry analysis for the GB200 compute board: one Grace CPU and two B200 GPUs on the same PCB, with two Bianca boards in each 1U compute tray of an NVL72 rack (SemiAnalysis GB200 architecture). Putting the CPU and GPUs on one board removes the need for PCIe switches or retimers between them in the reference design.

NVIDIA does not use the name in its public documentation, and it does not publish the board’s layer count, laminate or fabricator. Numbers you will see online for Bianca layer counts, material ratios and board prices come from analyst models and supply-chain reporting. They are useful for market context, not as a design input; if you are building hardware that interfaces with these systems, your requirements come from your own channel and the connector or module specification you design to.

Which Boards in a Blackwell System Need Low-Loss Laminate?

Any board that carries the fastest serial links over more than a short distance. In practice:

Many of these boards are hybrid builds: the premium laminate only on the layers that carry the fastest signals, with lower-cost material elsewhere. Whether a hybrid works depends on the two materials laminating together without warpage or delamination, which the fabricator confirms with the material makers’ data.

How Do M6, M7 and M8 Laminates Compare?

The “M” tiers come from Panasonic’s MEGTRON numbering and are now used loosely across the industry; they are not an IPC grade (see our M9 CCL guide). Panasonic’s own datasheets measure all three generations with the same method, the balanced-type circular disk resonator (BCDR), which makes a fair ladder:

Sources: Panasonic datasheets for R-5775(N), R-5785(N) and R-5795; 0.75 mm samples, typical values.

Each step roughly cuts the dielectric loss by a third, and the glass choice alone moves MEGTRON 8 from 0.0016 to 0.0012. Other makers sell laminates in the same class, such as EMC’s EM-892K2 and Doosan’s DS-7409DJN+; compare them on their own datasheets and test methods, as in our MEGTRON 8 guide.

What Else Besides the Laminate Controls Loss?

Copper and vias. At the frequencies these links run, rough copper adds conductor loss that a better laminate cannot recover, so high-speed layers use very-low-profile foil (the HVLP grades). Via stubs left below the last connected layer resonate and cut into the channel; backdrilling removes them, and the drawing should state which nets are backdrilled and the maximum stub you will accept. Glass weave matters for skew on differential pairs, which is why low-Dk and spread-glass styles are used on the fastest layers.

None of these is visible in a material name. If the drawing only says “MEGTRON 8,” a fabricator can still choose the glass style, copper profile and resin content, and each changes the result.

What Fabrication Features Do These Boards Need?

The same set appears on most AI server and switch boards:

• High layer counts and thick boards, where layer-to-layer registration decides whether fine-pitch BGA escapes and backdrills land correctly.
• Sequential lamination and microvias where the BGA pitch requires HDI escape routing.
• Backdrilling with depth control on high-speed nets.
• Hybrid lamination of different resin systems in one stack.
• Impedance and loss coupons built on the production panel and tested.
• Material lead time: the newest laminates can be allocated; check availability before freezing the stackup.

Similar process controls apply to other ultra-low-loss materials such as Tachyon 100G.

What Changes After Blackwell?

Loss budgets keep shrinking as lane rates rise, which pushes the fastest boards toward the next laminate tier, usually called M9, with quartz-based glass in some products and even smoother copper. Those materials are harder to drill and in shorter supply. For a board you are designing now, the practical step is to model the channel with the exact candidate materials and check whether an M8-class laminate with smooth copper and short stubs already closes the budget before paying for the next tier.

How Do You Specify a Board That Works with Blackwell Hardware?

How Can We Help?

QueenEMS builds multilayer boards up to 40 layers, including HDI builds, for switch, NIC and accelerator projects that sit alongside platforms like Blackwell. Send your stackup with the exact laminate, glass and copper you want; you get a bare-board quote within 2 hours that names the material we can source and the construction it assumes.

Conclusion

Blackwell is NVIDIA’s AI GPU architecture, shipped as GB200 superchips and NVL72 racks with 1.8 TB/s NVLink per GPU. NVIDIA does not publish its board materials; what is reliable is the laminate data from the material makers, where MEGTRON 6, 7 and 8 step Df from 0.0037 to 0.0012 on the same test. Choose the tier your channel model needs, specify the exact product, glass, copper and backdrill, and confirm material availability early. Request a quote by sending the stackup, laminate choice and Gerber files.

FAQ

What is an NVIDIA Blackwell GPU? NVIDIA’s data-center GPU generation after Hopper, sold as the B200 and in GB200 superchips and NVL72 racks.

What does GB200 mean? The Grace Blackwell Superchip: one Grace CPU and two Blackwell GPUs linked by NVLink-C2C.

What is the Bianca board? The industry name for the GB200 compute board, which carries one Grace CPU and two B200 GPUs; an NVL72 compute tray holds two.

What PCB material does NVIDIA use for Blackwell? NVIDIA does not publish it. High-speed boards in this class use ultra-low-loss laminates in the M7–M8 tier with smooth copper.

Is M8 laminate required for every Blackwell-related board? No. Power, management and short low-speed boards usually use standard high-Tg FR-4; the tier depends on each board’s channel loss budget.