MI325X: FP8 vs BF16 Precision Comparison

Near the low end of the 41–72 tok/s/user interactivity band, at 49 tok/s/user on Qwen 3.5 397B-A17B (MI325X): FP8 runs 347 tok/s/GPU at $1.04/M tokens, BF16 runs 462 at $0.77/M. BF16 is 34% cheaper per token; BF16 delivers 33% more tok/s/GPU. Precision changes affect both inference speed and model quality — consult the evaluation tab for accuracy benchmarks.

At 57 tok/s/user on Qwen 3.5 397B-A17B (MI325X), FP8 delivers 211 tok/s/GPU at $1.69 per million tokens; BF16 delivers 382 tok/s/GPU at $0.94. BF16 is 80% cheaper per token; BF16 delivers 81% more tok/s/GPU. Lower-precision quantization trades model accuracy for throughput — check the evaluation page for quality impact.

FP8 posts 124 tok/s/GPU for $2.86 per million tokens at 65 tok/s/user on Qwen 3.5 397B-A17B (MI325X); BF16 posts 307 tok/s/GPU for $1.17. BF16 is 144% cheaper per token; BF16 delivers 147% more tok/s/GPU. Quantization-level accuracy differences are tracked on the evaluation tab. (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.)