| Precision | Peak (PFLOPS) | Per GPU (TFLOPS) | TFLOPS/W |
|---|---|---|---|
| FP32 | 6.0 | 83.3 | — |
| TF32 | 90.0 | 1250.0 | — |
| FP16 | 180.0 | 2500.0 | — |
| BF16 | 180.0 | 2500.0 | — |
| FP8 | 360.0 | 5000.0 | — |
| FP4 | 1080.0 | 15000.0 | — |
All figures are dense. Vendors commonly headline the sparse number, which is twice the dense one.
NVLink + NVSwitch
36 GB300 superchips (72 B300 GPUs + 36 Grace CPUs) in liquid-cooled rack