Qwen 3.5 397B-A17B · Speculative decoding
B300 BF16: MTP vs Off speculative decoding
Speculative decoding comparison of MTP versus Off on B300 BF16 (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.
At 97 tok/s/user on Qwen 3.5 397B-A17B (B300 BF16), MTP delivers 1,475 tok/s/chip at $0.43 per million tokens; Off delivers 506 tok/s/chip at $1.24. MTP delivers 191% more tok/s/chip on the same chip, which makes it 66% cheaper per token. Speculative decoding accepts draft tokens to reduce per-token latency — gains vary by workload and prompt distribution.
MTP posts 1,053 tok/s/chip for $0.60 per million tokens at 122 tok/s/user on Qwen 3.5 397B-A17B (B300 BF16); Off posts 376 tok/s/chip for $1.67. 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.
Throughput at 146 tok/s/user on Qwen 3.5 397B-A17B (B300 BF16): MTP hits 803 tok/s/chip, Off hits 256. Costs per million tokens land at $0.78 and $2.45 respectively. MTP delivers 3.1× the tok/s/chip on the same chip, which makes it 68% cheaper per token. Speculative decoding trades extra compute on draft tokens for fewer decoding steps — the payoff depends on sequence length and batch size. (Numbers reflect this URL's pinned 1k/1k · bf16 workload — changing sequence or model updates both the table and chart; the table stays pinned to this page's precision, so precision toggles in the controls affect the chart only.)
Interpolated from real benchmark data. Edit target interactivity values below to compare at different operating points.
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | MTP1,475Off506 | MTP1,053Off376 | MTP803Off256 |
| Cost ($/M tok) | MTP$0.43Off$1.24 | MTP$0.60Off$1.67 | MTP$0.78Off$2.45 |
| tok/s/MW | MTP776,371Off266,417 | MTP554,433Off198,075 | MTP422,419Off134,641 |
| Concurrency | MTP~64Off~11 | MTP~41Off~6 | MTP~24Off~4 |
