GB200 NVL72: FP4 vs FP8 Precision Comparison

Throughput at 60 tok/s/user on MiniMax M2.5/M2.7 (GB200 NVL72): FP4 hits 10256 tok/s/GPU, FP8 hits 3263. Per-million costs land at $0.06 and $0.19 respectively. FP4 is 212% cheaper per token; FP4 delivers 214% more tok/s/GPU. The cost-throughput tradeoff from lower precision is only part of the picture — see the evaluation page for accuracy data.

Around the middle of the 31–147 tok/s/user interactivity band, at 89 tok/s/user on MiniMax M2.5/M2.7 (GB200 NVL72): FP4 runs 5504 tok/s/GPU at $0.11/M tokens, FP8 runs 831 at $0.75/M. FP4 is 567% cheaper per token; FP4 delivers 563% more tok/s/GPU. Precision changes affect both inference speed and model quality — consult the evaluation tab for accuracy benchmarks.

At 118 tok/s/user on MiniMax M2.5/M2.7 (GB200 NVL72), FP4 delivers 2019 tok/s/GPU at $0.30 per million tokens; FP8 delivers 331 tok/s/GPU at $1.83. FP4 is 505% cheaper per token; FP4 delivers 510% more tok/s/GPU. Lower-precision quantization trades model accuracy for throughput — check the evaluation page for quality impact. (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.)