MI355X: FP4 vs FP8 Precision Comparison

FP4 posts 1772 tok/s/GPU for $0.23 per million tokens at 56 tok/s/user on Qwen 3.5 397B-A17B (MI355X); FP8 posts 2363 tok/s/GPU for $0.17. FP8 is 34% cheaper per token; FP8 delivers 33% more tok/s/GPU. Quantization-level accuracy differences are tracked on the evaluation tab.

Throughput at 98 tok/s/user on Qwen 3.5 397B-A17B (MI355X): FP4 hits 1084 tok/s/GPU, FP8 hits 1258. Per-million costs land at $0.38 and $0.32 respectively. FP8 is 18% cheaper per token; FP8 delivers 16% 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 16–180 tok/s/user interactivity band, at 139 tok/s/user on Qwen 3.5 397B-A17B (MI355X): FP4 runs 675 tok/s/GPU at $0.59/M tokens, FP8 runs 853 at $0.49/M. FP8 is 22% cheaper per token; FP8 delivers 26% 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.)