| Precision | Peak (PFLOPS) | Per GPU (TFLOPS) | TFLOPS/W |
|---|---|---|---|
| FP64 | 2.88 | 40 | 27.00 |
| FP32 | 5.76 | 80 | 54.00 |
| TF32 | 90.0 | 1,250 | 1500.00 |
| FP16 | 180 | 2,500 | 2700.00 |
| BF16 | 180 | 2,500 | 2700.00 |
| FP8 | 360 | 5,000 | 5400.00 |
| FP4 | 720 | 10,000 | 10800.00 |
| INT8 | 360 | 5,000 | 5400.00 |
| FP6 | 360 | 5,000 | — |
All figures are dense. Vendors commonly headline the sparse number, which is twice the dense one.
NVLink
36 GB200 superchips (72 B200 GPUs + 36 Grace CPUs) in a liquid-cooled rack-scale system