HGX

HGX is NVIDIA's GPU baseboard platform, a reference design that server makers build into their own rack servers rather than a finished system NVIDIA sells directly. NVIDIA's HGX page lists a single baseboard design that joins eight Rubin, Blackwell, or Blackwell Ultra SXM GPUs over NVLink and NVSwitch. Rubin boards pair with an NVIDIA Vera CPU or an x86 CPU baseboard, and for x86 the server maker sets CPU and memory specs.

The HGX B200 baseboard holds eight Blackwell GPUs with 180 GB of HBM3e each, for 1,440 GB per board. The HGX B300 holds eight Blackwell Ultra GPUs with up to 288 GB of HBM3e each. Both give each GPU 1.8 TB/s of bidirectional NVLink bandwidth, 14.4 TB/s per board. HGX Rubin NVL8 holds eight Rubin GPUs with 288 GB of HBM4 each. NVIDIA's reference architecture for HGX H100, H200, and B200 systems also allows 4-GPU designs.

A quote naming HGX B200, HGX B300, or HGX Rubin NVL8 describes the GPU baseboard inside a server maker's system, not a standalone NVIDIA product. Compare those quotes on the CPU, system memory, storage, and cooling built around the same board. For double-precision (FP64) work, check the board first, since NVIDIA lists 296 TFLOPS on HGX B200 and 10 TFLOPS on HGX B300.

Sources

Last verified October 7, 2026.