| Precision | Peak (TFLOPS) | Per GPU (TFLOPS) | TFLOPS/W |
|---|---|---|---|
| FP64 | 296 | 37 | — |
| FP32 | 600 | 75 | — |
| TF32 | 9,000 | 1,125 | — |
| FP16 | 18,000 | 2,250 | — |
| BF16 | 18,000 | 2,250 | — |
| FP8 | 36,000 | 4,500 | — |
| FP4 | 72,000 | 9,000 | — |
| INT8 | 36,000 | 4,500 | — |
All figures are dense. Vendors commonly headline the sparse number, which is twice the dense one.
NVLink 5.0 / 2x NVSwitch, 14.4 TB/s aggregate
Unveiled at GTC March 2024. 10 RU air-cooled chassis with 2x Intel Xeon Platinum 8570 (112 cores) and 2 TB system memory, configurable to 4 TB. Networking is 4x OSFP serving 8x single-port ConnectX-7 plus 2x dual-port QSFP112 BlueField-3 DPUs, each up to 400 Gb/s. Same eight B200 SXM parts as HGX B200; the 14.3 kW figure is the whole system, against 10.2 kW for the bare baseboard.