B200: FP4 vs FP8 Precision Comparison

At 76 tok/s/user on Llama 3.3 70B (B200), FP4 delivers 5971 tok/s/GPU at $0.09 per million tokens; FP8 delivers 4431 tok/s/GPU at $0.12. FP4 is 32% cheaper per token; FP4 delivers 35% more tok/s/GPU. Lower-precision quantization trades model accuracy for throughput — check the evaluation page for quality impact.

FP4 posts 3099 tok/s/GPU for $0.18 per million tokens at 108 tok/s/user on Llama 3.3 70B (B200); FP8 posts 2744 tok/s/GPU for $0.20. FP4 is 12% cheaper per token; FP4 delivers 13% more tok/s/GPU. Quantization-level accuracy differences are tracked on the evaluation tab.

Throughput at 139 tok/s/user on Llama 3.3 70B (B200): FP4 hits 1032 tok/s/GPU, FP8 hits 912. Per-million costs land at $0.53 and $0.58 respectively. FP4 is 10% cheaper per token; FP4 delivers 13% more tok/s/GPU. The cost-throughput tradeoff from lower precision is only part of the picture — see the evaluation page for accuracy data. (Numbers reflect the default 1k/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.)