GB200 NVL72: FP4 vs FP8 Precision Comparison

Throughput at 45 tok/s/user on GLM 5/5.1 (GB200 NVL72): FP4 hits 3344 tok/s/GPU, FP8 hits 635. Per-million costs land at $0.19 and $0.96 respectively. FP4 is 417% cheaper per token; FP4 delivers 426% 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 40–62 tok/s/user interactivity band, at 51 tok/s/user on GLM 5/5.1 (GB200 NVL72): FP4 runs 1781 tok/s/GPU at $0.35/M tokens, FP8 runs 270 at $2.44/M. FP4 is 606% cheaper per token; FP4 delivers 559% more tok/s/GPU. Precision changes affect both inference speed and model quality — consult the evaluation tab for accuracy benchmarks.

At 56 tok/s/user on GLM 5/5.1 (GB200 NVL72), FP4 delivers 1456 tok/s/GPU at $0.42 per million tokens; FP8 delivers 108 tok/s/GPU at $6.48. FP4 is 1433% cheaper per token; FP4 delivers 1245% more tok/s/GPU. Lower-precision quantization trades model accuracy for throughput — check the evaluation page for quality impact. (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.)