GB300 NVL72: FP4 vs FP8 Precision Comparison

Near the low end of the 44–228 tok/s/user interactivity band, at 90 tok/s/user on Qwen 3.5 397B-A17B (GB300 NVL72): FP4 runs 6025 tok/s/GPU at $0.12/M tokens, FP8 runs 10539 at $0.07/M. FP8 is 75% cheaper per token; FP8 delivers 75% more tok/s/GPU. Precision changes affect both inference speed and model quality — consult the evaluation tab for accuracy benchmarks.

At 136 tok/s/user on Qwen 3.5 397B-A17B (GB300 NVL72), FP4 delivers 2801 tok/s/GPU at $0.26 per million tokens; FP8 delivers 4769 tok/s/GPU at $0.16. FP8 is 64% cheaper per token; FP8 delivers 70% more tok/s/GPU. Lower-precision quantization trades model accuracy for throughput — check the evaluation page for quality impact.

FP4 posts 1338 tok/s/GPU for $0.55 per million tokens at 183 tok/s/user on Qwen 3.5 397B-A17B (GB300 NVL72); FP8 posts 2682 tok/s/GPU for $0.27. FP8 is 102% cheaper per token; FP8 delivers 100% 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.)