B200: FP8 vs BF16 Precision Comparison

FP8 posts 1429 tok/s/GPU for $0.37 per million tokens at 112 tok/s/user on Qwen 3.5 397B-A17B (B200); BF16 posts 1216 tok/s/GPU for $0.46. FP8 is 23% cheaper per token; FP8 delivers 18% more tok/s/GPU. Quantization-level accuracy differences are tracked on the evaluation tab.

Throughput at 166 tok/s/user on Qwen 3.5 397B-A17B (B200): FP8 hits 828 tok/s/GPU, BF16 hits 660. Per-million costs land at $0.65 and $0.82 respectively. FP8 is 27% cheaper per token; FP8 delivers 25% 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 59–272 tok/s/user interactivity band, at 219 tok/s/user on Qwen 3.5 397B-A17B (B200): FP8 runs 552 tok/s/GPU at $0.95/M tokens, BF16 runs 299 at $1.88/M. FP8 is 97% cheaper per token; FP8 delivers 84% 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.)