Qwen 3.5 397B-A17B · Speculative decoding

B200 FP8: MTP vs Off speculative decoding

Speculative decoding comparison of MTP versus Off on B200 FP8 (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 comparability

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 74 tok/s/user on Qwen 3.5 397B-A17B (B200 FP8): MTP hits 8,147 tok/s/chip, Off hits 4,542. Costs per million tokens land at $0.059 and $0.11 respectively. MTP delivers 79% more tok/s/chip on the same chip, which makes it 44% 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 23–226 tok/s/user interactivity band, at 124 tok/s/user on Qwen 3.5 397B-A17B (B200 FP8): MTP runs 5,867 tok/s/chip at $0.082 per million tokens, Off runs 2,631 at $0.18. MTP delivers 123% more tok/s/chip on the same chip, which makes it 55% cheaper per token. Gains from speculative decoding vary by workload; short-output prompts tend to benefit less.

At 175 tok/s/user on Qwen 3.5 397B-A17B (B200 FP8), MTP delivers 4,342 tok/s/chip at $0.11 per million tokens; Off delivers 1,338 tok/s/chip at $0.36. MTP delivers 3.2× the tok/s/chip on the same chip, which makes it 69% 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 8k/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)
MTP8,147Off4,542
MTP5,867Off2,631
MTP4,342Off1,338
Cost ($/M tok)
MTP$0.059Off$0.11
MTP$0.082Off$0.18
MTP$0.11Off$0.36
tok/s/MW
MTP4,764,309Off2,656,104
MTP3,430,887Off1,538,371
MTP2,539,215Off782,445
Concurrency
MTP~52Off~29
MTP~23Off~10
MTP~12Off~4
Qwen 3.5 397B-A17B: B200 FP8 MTP versus Off speculative decoding comparison at matched interactivity levels
B200 FP8 MTP versus Off speculative decoding comparison for this page's canonical default workload
Open full-size image ↗