| Precision | Peak (PFLOPS) | Per GPU (TFLOPS) | TFLOPS/W |
|---|---|---|---|
| FP64 | 0.10 | 1.4 | — |
| FP32 | 6.00 | 83 | — |
| TF32 | 90.0 | 1,250 | — |
| FP16 | 180 | 2,500 | — |
| BF16 | 180 | 2,500 | — |
| FP8 | 360 | 5,000 | — |
| FP4 | 1,080 | 15,000 | — |
| INT8 | 12.0 | 167 | — |
| FP6 | 360 | 5,000 | — |
All figures are dense. Vendors commonly headline the sparse number, which is twice the dense one.
NVLink + NVSwitch
NVIDIA's own build of the GB300 NVL72 reference rack: 36 GB300 superchips, 72 Blackwell Ultra GPUs and 36 Grace CPUs, joined into one NVLink domain in a liquid-cooled rack.