MI325X FP8: MTP vs Off Speculative Decoding

At 9 tok/s/user on DeepSeekv4 Pro 0813 1.6T (MI325X FP8), MTP delivers 503 tok/s/chip at $0.61 per million tokens; Off delivers 365 tok/s/chip at $0.84. MTP is 38% cheaper per token; MTP delivers 38% more tok/s/chip. Speculative decoding accepts draft tokens to reduce per-token latency — gains vary by workload and prompt distribution.

MTP posts 356 tok/s/chip for $0.86 per million tokens at 13 tok/s/user on DeepSeekv4 Pro 0813 1.6T (MI325X FP8); Off posts 215 tok/s/chip for $1.42. MTP is 66% cheaper per token; MTP delivers 66% more tok/s/chip. Draft-token acceptance rates determine whether speculative decoding helps or hurts at a given concurrency level.

Throughput at 17 tok/s/user on DeepSeekv4 Pro 0813 1.6T (MI325X FP8): MTP hits 267 tok/s/chip, Off hits 99. Per-million costs land at $1.14 and $3.10 respectively. MTP is 171% cheaper per token; MTP delivers 171% more tok/s/chip. Speculative decoding trades extra compute on draft tokens for fewer decoding steps — the payoff depends on sequence length and batch size. (Numbers reflect this URL's pinned 1k/1k · fp8 workload — changing sequence or model updates both the table and chart; the table stays pinned to this page's precision, so precision toggles in the controls affect the chart only.)