GB200 NVL72: FP4 vs FP8 Precision Comparison

FP4 posts 12503 tok/s/GPU for $0.05 per million tokens at 75 tok/s/user on DeepSeek R1 (GB200 NVL72); FP8 posts 5790 tok/s/GPU for $0.11. FP4 is 116% cheaper per token; FP4 delivers 116% more tok/s/GPU. Quantization-level accuracy differences are tracked on the evaluation tab.

Throughput at 120 tok/s/user on DeepSeek R1 (GB200 NVL72): FP4 hits 9356 tok/s/GPU, FP8 hits 3988. Per-million costs land at $0.07 and $0.15 respectively. FP4 is 134% cheaper per token; FP4 delivers 135% more tok/s/GPU. The cost-throughput tradeoff from lower precision is only part of the picture — see the evaluation page for accuracy data.

Toward the upper edge of the 31–208 tok/s/user interactivity band, at 164 tok/s/user on DeepSeek R1 (GB200 NVL72): FP4 runs 3995 tok/s/GPU at $0.16/M tokens, FP8 runs 818 at $0.81/M. FP4 is 419% cheaper per token; FP4 delivers 389% more tok/s/GPU. Precision changes affect both inference speed and model quality — consult the evaluation tab for accuracy benchmarks. (Numbers reflect the default 8k/1k selection for this URL — table and chart below update if you change sequence or model in the controls. Each side uses the best available serving configuration for that precision, which may include speculative decoding such as MTP where recipes exist — the same convention as the other comparison pages.)