| Precision | Peak (PFLOPS) | Per GPU (TFLOPS) | TFLOPS/W |
|---|---|---|---|
| FP8 | 1,450 | 20,133 | — |
| FP4 | 2,899 | 40,265 | — |
All figures are dense. Vendors commonly headline the sparse number, which is twice the dense one.
UALink over Ethernet (UALoE) scale-up; Ultra Ethernet (UEC) scale-out; AMD Pensando DPUs