GB200 NVL72: FP4 vs FP8 precision comparison

At 125 tok/s/user on Qwen 3.5 397B-A17B (GB200 NVL72), FP4 delivers 11,554 tok/s/chip at $0.045 per million tokens; FP8 delivers 7,487 tok/s/chip at $0.069. FP4 delivers 54% more tok/s/chip on the same chip, which makes it 35% cheaper per token. Lower-precision quantization trades model accuracy for throughput — check the evaluation page for quality impact.

FP4 posts 4,680 tok/s/chip for $0.11 per million tokens at 208 tok/s/user on Qwen 3.5 397B-A17B (GB200 NVL72); FP8 posts 2,496 tok/s/chip for $0.21. FP4 delivers 87% more tok/s/chip on the same chip, which makes it 47% cheaper per token. Quantization-level accuracy differences are tracked on the evaluation tab.

Throughput at 291 tok/s/user on Qwen 3.5 397B-A17B (GB200 NVL72): FP4 hits 2,116 tok/s/chip, FP8 hits 1,160. Costs per million tokens land at $0.24 and $0.45 respectively. FP4 delivers 82% more tok/s/chip on the same chip, which makes it 45% cheaper per token. The cost-throughput tradeoff from lower precision is only part of the picture — see the evaluation page for accuracy data. (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.)