B200 vs B300

HGX B200 and HGX B300 are NVIDIA's two eight-GPU Blackwell baseboards. B200 carries Blackwell SXM GPUs and B300 carries Blackwell Ultra SXM GPUs. Both link their GPUs with fifth-generation NVLink through NVLink 5 switches at 1.8 TB/s per GPU and 14.4 TB/s per board. Server makers build both into their own systems.

NVIDIA's reference architecture lists 180 GB of HBM3e per B200 GPU and 288 GB per B300 GPU, both at up to 8 TB/s, with eight GPUs per board. NVIDIA's HGX page lists FP4 at 144 PFLOPS sparse on both boards, but 108 PFLOPS dense on B300 against 72 PFLOPS on B200, and 2x attention performance on B300. FP64 falls from 296 TFLOPS on B200 to 10 TFLOPS on B300. NVIDIA lists networking bandwidth at 1.6 TB/s for B300 against 0.8 TB/s for B200, and pairs B300 with ConnectX-8 SuperNICs at up to 800 Gb/s per GPU.

Pick B300 for FP4 serving and large models that need the extra memory. Pick B200 for double-precision HPC. Take rack power and cooling from the server maker's system rating, since NVIDIA's HGX page lists no board power.

Sources

Last verified October 30, 2026.