GB300 NVL72: FP4 vs FP8 Precision Comparison

Near the low end of the 44–206 tok/s/user interactivity band, at 85 tok/s/user on Qwen 3.5 397B-A17B (GB300 NVL72): FP4 runs 6560 tok/s/GPU at $0.11/M tokens, FP8 runs 3329 at $0.22/M. FP4 is 101% cheaper per token; FP4 delivers 97% more tok/s/GPU. Precision changes affect both inference speed and model quality — consult the evaluation tab for accuracy benchmarks.

At 125 tok/s/user on Qwen 3.5 397B-A17B (GB300 NVL72), FP4 delivers 3234 tok/s/GPU at $0.22 per million tokens; FP8 delivers 1703 tok/s/GPU at $0.43. FP4 is 94% cheaper per token; FP4 delivers 90% more tok/s/GPU. Lower-precision quantization trades model accuracy for throughput — check the evaluation page for quality impact.

FP4 posts 1841 tok/s/GPU for $0.39 per million tokens at 166 tok/s/user on Qwen 3.5 397B-A17B (GB300 NVL72); FP8 posts 775 tok/s/GPU for $0.94. FP4 is 143% cheaper per token; FP4 delivers 138% more tok/s/GPU. Quantization-level accuracy differences are tracked on the evaluation tab. (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.)