Qwen 3.5 397B-A17B · Speculative decoding
MI355X FP8: MTP vs Off speculative decoding
Speculative decoding comparison of MTP versus Off on MI355X FP8 (AMD CDNA 4) 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.
Near the low end of the 25–131 tok/s/user interactivity band, at 51 tok/s/user on Qwen 3.5 397B-A17B (MI355X FP8): MTP runs 2,517 tok/s/chip at $0.17 per million tokens, Off runs 1,330 at $0.31. MTP delivers 89% more tok/s/chip on the same chip, which makes it 47% cheaper per token. Gains from speculative decoding vary by workload; short-output prompts tend to benefit less.
At 78 tok/s/user on Qwen 3.5 397B-A17B (MI355X FP8), MTP delivers 1,711 tok/s/chip at $0.24 per million tokens; Off delivers 683 tok/s/chip at $0.61. MTP delivers 150% more tok/s/chip on the same chip, which makes it 60% cheaper per token. Speculative decoding accepts draft tokens to reduce per-token latency — gains vary by workload and prompt distribution.
MTP posts 1,183 tok/s/chip for $0.35 per million tokens at 104 tok/s/user on Qwen 3.5 397B-A17B (MI355X FP8); Off posts 422 tok/s/chip for $0.99. MTP delivers 180% more tok/s/chip on the same chip, which makes it 64% cheaper per token. Draft-token acceptance rates determine whether speculative decoding helps or hurts at a given concurrency level. (Numbers reflect this URL's pinned 1k/1k · fp8 workload — changing sequence or model updates both the table and chart; the table stays pinned to this page's precision, so precision toggles in the controls affect the chart only.)
Interpolated from real benchmark data. Edit target interactivity values below to compare at different operating points.
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | MTP2,517Off1,330 | MTP1,711Off683 | MTP1,183Off422 |
| Cost ($/M tok) | MTP$0.17Off$0.31 | MTP$0.24Off$0.61 | MTP$0.35Off$0.99 |
| tok/s/MW | MTP1,204,353Off636,477 | MTP818,518Off326,961 | MTP565,905Off201,765 |
| Concurrency | MTP~96Off~32 | MTP~44Off~32 | MTP~23Off~9 |
