NVIDIA HGX B300
Overview
Performance Metrics
| Precision | Peak (PFLOPS) | Per GPU (TFLOPS) | TFLOPS/W |
|---|---|---|---|
| FP64 | 0.0 | 1.3 | — |
| FP32 | 0.6 | 75.0 | — |
| TF32 | 9.0 | 1125.0 | — |
| FP16 | 18.0 | 2250.0 | — |
| BF16 | 18.0 | 2250.0 | — |
| FP8 | 36.0 | 4500.0 | — |
| FP4 | 108.0 | 13500.0 | — |
| FP6 | 36.0 | 4500.0 | — |
All figures are dense. Vendors commonly headline the sparse number, which is twice the dense one.
Configuration
- Accelerator
- B300 SXM 288GB
- Count
- 8
- Memory per GPU
- 288 GB
- Bandwidth per GPU
- 8.0 TB/s
Power & Efficiency
- Aggregate Power
- —
- TDP per GPU
- 1400 W
- FP8 Efficiency
- —
- Aggregate Bandwidth
- 64.0 TB/s
System Details
Interconnect
NVLink 5.0 / NVSwitch 4.0
Notes
The eight-GPU Blackwell Ultra baseboard OEMs build around. NVIDIA publishes both a dense and a sparse figure for every tensor precision here, and for FP4 the two are not a factor of two apart: dense rose by half over the B200, from 72 to 108 PFLOPS, while sparse stayed at 144. Total memory is NVIDIA's own 2.1 TB rather than the 2.304 TB that eight 288 GB GPUs would suggest. Memory bandwidth of 64 TB/s is the one derived number, being eight times the B300's published 8 TB/s, because NVIDIA gives no board-level HBM figure. No INT8 row is recorded: NVIDIA's table gives 3 POPS sparse for this board while our B300 chip entry implies 36 POPS dense, the two cannot both be right, and the question is open rather than guessed at. NVIDIA publishes no baseboard power figure, so total power is blank.