NVIDIA HGX B200

HGX B200 is the first generation Blackwell eight GPU baseboard and remains the volume platform for LLM training and inference clusters. Each GPU has 180 GB of HBM3e at 8 TB/s, and NVLink with NVSwitch links all eight at 1.8 TB/s per GPU. NVIDIA publishes 144 PFLOPS of sparse FP4 and 72 PFLOPS of sparse FP8 per baseboard, with a per GPU power envelope configurable up to 1,000 W.

Compared with HGX B300, B200 carries about 37 percent less GPU memory (180 GB against 288 GB per GPU) and lower dense FP4 throughput, but it is available to order without allocation and runs in the same 8U and 10U air cooled chassis classes at a slightly lower power draw. It is a sound choice for training clusters that are already tuned for Blackwell, and for inference fleets where memory per GPU is not the constraint. We quote it with 400 Gb/s ConnectX-7 or 800 Gb/s ConnectX-8 east west networking.