Qwen 3.5 397B-A17B · Speculative decoding
MI325X FP8: MTP vs Off speculative decoding
Speculative decoding comparison of MTP versus Off on MI325X FP8 (AMD CDNA 3) running Qwen 3.5 397B-A17B. Throughput, cost, and interactivity differences across LLM workloads. Use the chart controls below to switch sequences and metrics — same interactions as the main inference chart.
MTP acceptance-rate implementations differ across inference engines. Points from different engines are not directly comparable on the same curve — throughput and cost at matched interactivity may reflect engine-level differences rather than pure speculative decoding gains. Interpret cross-engine comparisons with caution.
At 46 tok/s/user on Qwen 3.5 397B-A17B (MI325X FP8), MTP delivers 402 tok/s/chip at $0.76 per million tokens; Off delivers 400 tok/s/chip at $0.76. Cost per token is essentially tied; throughput per chip is essentially tied. Speculative decoding accepts draft tokens to reduce per-token latency — gains vary by workload and prompt distribution.
MTP posts 230 tok/s/chip for $1.33 per million tokens at 55 tok/s/user on Qwen 3.5 397B-A17B (MI325X FP8); Off posts 229 tok/s/chip for $1.33. Cost per token is essentially tied; throughput per chip is essentially tied. Draft-token acceptance rates determine whether speculative decoding helps or hurts at a given concurrency level.
Throughput at 64 tok/s/user on Qwen 3.5 397B-A17B (MI325X FP8): MTP hits 130 tok/s/chip, Off hits 132. Costs per million tokens land at $2.36 and $2.31 respectively. Off delivers 2% more tok/s/chip on the same chip, which makes it 2% cheaper per token. Speculative decoding trades extra compute on draft tokens for fewer decoding steps — the payoff depends on sequence length and batch size. (Numbers reflect this URL's pinned 1k/1k · fp8 workload — changing sequence or model updates both the table and chart; the table stays pinned to this page's precision, so precision toggles in the controls affect the chart only.)
Interpolated from real benchmark data. Edit target interactivity values below to compare at different operating points.
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | MTP402Off400 | MTP230Off229 | MTP130Off132 |
| Cost ($/M tok) | MTP$0.76Off$0.76 | MTP$1.33Off$1.33 | MTP$2.36Off$2.31 |
| tok/s/MW | MTP238,121Off236,764 | MTP136,137Off135,575 | MTP76,674Off78,371 |
| Concurrency | MTP~37Off~37 | MTP~18Off~18 | MTP~9Off~9 |
