MI300X: FP8 vs BF16 Precision Comparison

At 45 tok/s/user on Qwen 3.5 397B-A17B (MI300X), FP8 delivers 312 tok/s/GPU at $1.01 per million tokens; BF16 delivers 414 tok/s/GPU at $0.76. BF16 is 32% cheaper per token; BF16 delivers 32% more tok/s/GPU. Lower-precision quantization trades model accuracy for throughput — check the evaluation page for quality impact.

FP8 posts 179 tok/s/GPU for $1.71 per million tokens at 54 tok/s/user on Qwen 3.5 397B-A17B (MI300X); BF16 posts 315 tok/s/GPU for $0.97. BF16 is 77% cheaper per token; BF16 delivers 76% more tok/s/GPU. Quantization-level accuracy differences are tracked on the evaluation tab.

Throughput at 63 tok/s/user on Qwen 3.5 397B-A17B (MI300X): FP8 hits 116 tok/s/GPU, BF16 hits 258. Per-million costs land at $2.68 and $1.21 respectively. BF16 is 122% cheaper per token; BF16 delivers 121% 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 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.)