Qwen 3.5 397B-A17B · Speculative decoding
H200 FP8: MTP vs Off speculative decoding
Speculative decoding comparison of MTP versus Off on H200 FP8 (NVIDIA Hopper) 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 55 tok/s/user on Qwen 3.5 397B-A17B (H200 FP8), MTP delivers 549 tok/s/chip at $0.62 per million tokens; Off delivers 342 tok/s/chip at $0.99. MTP delivers 61% more tok/s/chip on the same chip, which makes it 38% cheaper per token. Speculative decoding accepts draft tokens to reduce per-token latency — gains vary by workload and prompt distribution.
MTP posts 458 tok/s/chip for $0.74 per million tokens at 75 tok/s/user on Qwen 3.5 397B-A17B (H200 FP8); Off posts 247 tok/s/chip for $1.37. MTP delivers 85% more tok/s/chip on the same chip, which makes it 46% cheaper per token. Draft-token acceptance rates determine whether speculative decoding helps or hurts at a given concurrency level.
Throughput at 96 tok/s/user on Qwen 3.5 397B-A17B (H200 FP8): MTP hits 378 tok/s/chip, Off hits 197. Costs per million tokens land at $0.90 and $1.72 respectively. MTP delivers 92% more tok/s/chip on the same chip, which makes it 48% cheaper per token. Speculative decoding trades extra compute on draft tokens for fewer decoding steps — the payoff depends on sequence length and batch size. (Numbers reflect this URL's pinned 1k/1k · fp8 workload — changing sequence or model updates both the table and chart; the table stays pinned to this page's precision, so precision toggles in the controls affect the chart only.)
Interpolated from real benchmark data. Edit target interactivity values below to compare at different operating points.
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | MTP549Off342 | MTP458Off247 | MTP378Off197 |
| Cost ($/M tok) | MTP$0.62Off$0.99 | MTP$0.74Off$1.37 | MTP$0.90Off$1.72 |
| tok/s/MW | MTP401,050Off249,632 | MTP334,098Off180,402 | MTP276,271Off143,891 |
| Concurrency | MTP~40Off~25 | MTP~24Off~13 | MTP~16Off~9 |
