GB300 NVL72: FP4 vs FP8 Precision Comparison

Near the low end of the 40–374 tok/s/user interactivity band, at 123 tok/s/user on Qwen 3.5 397B-A17B (GB300 NVL72): FP4 runs 29149 tok/s/chip at $0.02/M tokens, FP8 runs 6206 at $0.10/M. FP4 is 370% cheaper per token; FP4 delivers 370% more tok/s/chip. Precision changes affect both inference speed and model quality — consult the evaluation tab for accuracy benchmarks.

At 207 tok/s/user on Qwen 3.5 397B-A17B (GB300 NVL72), FP4 delivers 13668 tok/s/chip at $0.05 per million tokens; FP8 delivers 1960 tok/s/chip at $0.33. FP4 is 597% cheaper per token; FP4 delivers 597% more tok/s/chip. Lower-precision quantization trades model accuracy for throughput — check the evaluation page for quality impact.

FP4 posts 9482 tok/s/chip for $0.07 per million tokens at 291 tok/s/user on Qwen 3.5 397B-A17B (GB300 NVL72); FP8 posts 772 tok/s/chip for $0.83. FP4 is 1128% cheaper per token; FP4 delivers 1128% more tok/s/chip. 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.)