NVIDIA HGX B200 8-GPU Baseboard, 1,536GB HBM3e (Umbriel)

No. This is the HGX B200 baseboard, the 8-GPU SXM carrier board only. It has no host CPUs, system memory, storage, chassis, power supplies, or fans. To run, it must be integrated into a system with a host platform, substantial power delivery, and appropriate cooling. It is a component for builders and integrators, not a turnkey machine. For a finished server, MillionMiner offers complete B200 systems and the DGX B200.

Umbriel is NVIDIA's codename for this 8-GPU HGX B200 baseboard design. It is not a separate product line, it refers to the standard current-generation Blackwell 8-GPU HGX B200 carrier board, the assembly described here (model 935-26287-00A0-000). It is the same class of baseboard that goes inside the major B200 systems and the DGX B200.

Generation and memory. The B200 is NVIDIA's current-generation Blackwell flagship, a generation beyond the Hopper H100 and H200, with substantially higher AI performance and FP4/FP8 acceleration. On memory, this baseboard has 192GB per GPU and 1,536GB total, the most in the lineup, above the H200's 141GB per GPU (1,128GB) and the H100's 80GB (640GB). For the largest current models, the B200 baseboard is the most capable assembly here; the Hopper baseboards remain capable at their generation and a different price point.

System builders and integrators assembling current-generation Blackwell nodes. If you have the chassis, host platform, power, and cooling, or are building them, the baseboard is the core 8-GPU assembly. If you want a server that runs out of the box, the complete B200 systems or the DGX B200 are the right path instead.

By NVLink and NVSwitch, the high-bandwidth fabric that interconnects all eight GPUs for fast all-to-all communication, so they work as a single GPU complex for large distributed training and inference. This full NVLink fabric is what distinguishes a true HGX baseboard from separate GPUs, and is essential to scaling the largest models across the eight GPUs efficiently.

It is the current generation. Blackwell is NVIDIA's newest data center architecture, a step beyond Hopper (H100, H200), with substantially higher AI throughput and the FP4 and FP8 low-precision formats that accelerate modern large-model training and inference. Combined with 192GB of HBM3e per GPU, the B200 is built for the largest current models and the most demanding training and inference.

An 8-GPU B200 assembly draws substantial power, well over ten kilowatts under load, and requires data center power delivery and cooling appropriate to a current-generation Blackwell baseboard, supplied by the host system it is integrated into. This is data center infrastructure, not a workstation part. MillionMiner confirms the exact power and cooling requirements, including whether your build is air- or liquid-cooled, in the quote.

Yes, unreservedly. This is current-generation flagship data center AI silicon: eight Blackwell B200 GPUs with full HBM3e and full NVLink fabric, the configuration used to train and serve the largest current AI models. Unlike the workstation and display cards elsewhere in this catalog, there is no overstatement to qualify here, only the scope to note that it is supplied as a baseboard for integration.

Eight B200 Blackwell SXM GPUs on one HGX baseboard (935-26287-00A0-000, Umbriel; Supermicro GPU-NVHGX-B200-8192), 1,536GB (1.5TB) of HBM3e total (192GB per GPU), full NVLink and NVSwitch interconnect, Blackwell architecture, with a 24-month warranty. It is currently available for pre-order. MillionMiner confirms the exact baseboard revision, the per-GPU and aggregate performance specs, the cooling and power requirements, and the pre-order timing in the quote.

B200 is a Blackwell-class accelerator subject to US export controls that depend on destination, and MillionMiner confirms destination eligibility and the compliance position before sale. As current high-value flagship hardware, condition and provenance matter, and each baseboard is verified before shipment with its condition stated clearly. Every unit is shipped worldwide DDP with duties handled, with hosting in MillionMiner's own data centers available.