MI355X: FP4 vs INT4 Precision Comparison

Throughput at 39 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T (MI355X): FP4 hits 1432 tok/s/GPU, INT4 hits 527. Per-million costs land at $0.29 and $0.78 respectively. FP4 is 169% cheaper per token; FP4 delivers 172% more tok/s/GPU. The cost-throughput tradeoff from lower precision is only part of the picture — see the evaluation page for accuracy data.

Around the middle of the 24–85 tok/s/user interactivity band, at 54 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T (MI355X): FP4 runs 1010 tok/s/GPU at $0.41/M tokens, INT4 runs 340 at $1.20/M. FP4 is 194% cheaper per token; FP4 delivers 197% more tok/s/GPU. Precision changes affect both inference speed and model quality — consult the evaluation tab for accuracy benchmarks.

At 70 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T (MI355X), FP4 delivers 755 tok/s/GPU at $0.54 per million tokens; INT4 delivers 229 tok/s/GPU at $1.80. FP4 is 235% cheaper per token; FP4 delivers 230% more tok/s/GPU. Lower-precision quantization trades model accuracy for throughput — check the evaluation page for quality impact. (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.)