MI355X: FP4 vs FP8 Precision Comparison

FP4 posts 2158 tok/s/GPU for $0.19 per million tokens at 68 tok/s/user on Qwen 3.5 397B-A17B (MI355X); FP8 posts 2012 tok/s/GPU for $0.20. FP4 is 8% cheaper per token; FP4 delivers 7% more tok/s/GPU. Quantization-level accuracy differences are tracked on the evaluation tab.

Throughput at 119 tok/s/user on Qwen 3.5 397B-A17B (MI355X): FP4 hits 1229 tok/s/GPU, FP8 hits 1050. Per-million costs land at $0.33 and $0.39 respectively. FP4 is 18% cheaper per token; FP4 delivers 17% more tok/s/GPU. The cost-throughput tradeoff from lower precision is only part of the picture — see the evaluation page for accuracy data.

Toward the upper edge of the 17–220 tok/s/user interactivity band, at 170 tok/s/user on Qwen 3.5 397B-A17B (MI355X): FP4 runs 757 tok/s/GPU at $0.54/M tokens, FP8 runs 590 at $0.69/M. FP4 is 28% cheaper per token; FP4 delivers 28% more tok/s/GPU. Precision changes affect both inference speed and model quality — consult the evaluation tab for accuracy benchmarks. (Numbers reflect the default 1k/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.)