B300: FP4 vs FP8 Precision Comparison

At 78 tok/s/user on DeepSeek R1 (B300), FP4 delivers 6579 tok/s/GPU at $0.10 per million tokens; FP8 delivers 662 tok/s/GPU at $1.00. FP4 is 900% cheaper per token; FP4 delivers 894% more tok/s/GPU. Lower-precision quantization trades model accuracy for throughput — check the evaluation page for quality impact.

FP4 posts 688 tok/s/GPU for $0.93 per million tokens at 143 tok/s/user on DeepSeek R1 (B300); FP8 posts 275 tok/s/GPU for $2.37. FP4 is 154% cheaper per token; FP4 delivers 151% more tok/s/GPU. Quantization-level accuracy differences are tracked on the evaluation tab.

Throughput at 207 tok/s/user on DeepSeek R1 (B300): FP4 hits 314 tok/s/GPU, FP8 hits 125. Per-million costs land at $2.08 and $5.25 respectively. FP4 is 152% cheaper per token; FP4 delivers 152% 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.)