MI355X: FP4 vs FP8 Precision Comparison

At 82 tok/s/user on MiniMax M3 428B (MI355X), FP4 delivers 6248 tok/s/GPU at $0.07 per million tokens; FP8 delivers 4257 tok/s/GPU at $0.10. FP4 is 46% cheaper per token; FP4 delivers 47% more tok/s/GPU. Lower-precision quantization trades model accuracy for throughput — check the evaluation page for quality impact.

FP4 posts 3868 tok/s/GPU for $0.11 per million tokens at 157 tok/s/user on MiniMax M3 428B (MI355X); FP8 posts 2254 tok/s/GPU for $0.18. FP4 is 70% cheaper per token; FP4 delivers 72% more tok/s/GPU. Quantization-level accuracy differences are tracked on the evaluation tab.

Throughput at 232 tok/s/user on MiniMax M3 428B (MI355X): FP4 hits 2229 tok/s/GPU, FP8 hits 1259. Per-million costs land at $0.18 and $0.32 respectively. FP4 is 77% cheaper per token; FP4 delivers 77% more tok/s/GPU. The cost-throughput tradeoff from lower precision is only part of the picture — see the evaluation page for accuracy data. (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.)