Qwen 3.5 397B-A17B · Speculative decoding

B300 FP8: MTP vs Off speculative decoding

Speculative decoding comparison of MTP versus Off on B300 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.

MTP posts 2,963 tok/s/chip for $0.21 per million tokens at 67 tok/s/user on Qwen 3.5 397B-A17B (B300 FP8); Off posts 1,169 tok/s/chip for $0.54. MTP delivers 154% more tok/s/chip on the same chip, which makes it 61% cheaper per token. Draft-token acceptance rates determine whether speculative decoding helps or hurts at a given concurrency level.

Throughput at 94 tok/s/user on Qwen 3.5 397B-A17B (B300 FP8): MTP hits 1,913 tok/s/chip, Off hits 731. Costs per million tokens land at $0.33 and $0.86 respectively. MTP delivers 162% more tok/s/chip on the same chip, which makes it 62% cheaper per token. Speculative decoding trades extra compute on draft tokens for fewer decoding steps — the payoff depends on sequence length and batch size.

Toward the upper edge of the 40–147 tok/s/user interactivity band, at 120 tok/s/user on Qwen 3.5 397B-A17B (B300 FP8): MTP runs 1,299 tok/s/chip at $0.48 per million tokens, Off runs 484 at $1.30. MTP delivers 168% more tok/s/chip on the same chip, which makes it 63% cheaper per token. Gains from speculative decoding vary by workload; short-output prompts tend to benefit less. (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)
MTP2,963Off1,169
MTP1,913Off731
MTP1,299Off484
Cost ($/M tok)
MTP$0.21Off$0.54
MTP$0.33Off$0.86
MTP$0.48Off$1.30
tok/s/MW
MTP1,559,227Off615,024
MTP1,006,906Off384,914
MTP683,932Off254,921
Concurrency
MTP~90Off~36
MTP~42Off~16
MTP~23Off~8
Qwen 3.5 397B-A17B: B300 FP8 MTP versus Off speculative decoding comparison at matched interactivity levels
B300 FP8 MTP versus Off speculative decoding comparison for this page's canonical default workload
Open full-size image ↗