Qwen 3.5 397B-A17B · Speculative decoding
B200 FP4: MTP vs Off speculative decoding
Speculative decoding comparison of MTP versus Off on B200 FP4 (NVIDIA Blackwell) 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.
Throughput at 77 tok/s/user on Qwen 3.5 397B-A17B (B200 FP4): MTP hits 7,266 tok/s/chip, Off hits 2,670. Costs per million tokens land at $0.066 and $0.18 respectively. MTP delivers 172% more tok/s/chip on the same chip, which makes it 63% cheaper per token. Speculative decoding trades extra compute on draft tokens for fewer decoding steps — the payoff depends on sequence length and batch size.
Around the middle of the 41–186 tok/s/user interactivity band, at 114 tok/s/user on Qwen 3.5 397B-A17B (B200 FP4): MTP runs 4,613 tok/s/chip at $0.10 per million tokens, Off runs 998 at $0.48. MTP delivers 4.6× the tok/s/chip on the same chip, which makes it 78% cheaper per token. Gains from speculative decoding vary by workload; short-output prompts tend to benefit less.
At 150 tok/s/user on Qwen 3.5 397B-A17B (B200 FP4), MTP delivers 2,640 tok/s/chip at $0.18 per million tokens; Off delivers 515 tok/s/chip at $0.93. MTP delivers 5.1× the tok/s/chip on the same chip, which makes it 80% cheaper per token. Speculative decoding accepts draft tokens to reduce per-token latency — gains vary by workload and prompt distribution. (Numbers reflect this URL's pinned 1k/1k · fp4 workload — changing sequence or model updates both the table and chart; the table stays pinned to this page's precision, so precision toggles in the controls affect the chart only.)
Interpolated from real benchmark data. Edit target interactivity values below to compare at different operating points.
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | MTP7,266Off2,670 | MTP4,613Off998 | MTP2,640Off515 |
| Cost ($/M tok) | MTP$0.066Off$0.18 | MTP$0.10Off$0.48 | MTP$0.18Off$0.93 |
| tok/s/MW | MTP4,248,854Off1,561,622 | MTP2,697,719Off583,789 | MTP1,543,823Off301,339 |
| Concurrency | MTP~406Off~67 | MTP~191Off~26 | MTP~92Off~8 |
