Qwen 3.5 397B-A17B · Speculative decoding

MI355X FP4: MTP vs Off speculative decoding

Speculative decoding comparison of MTP versus Off on MI355X FP4 (AMD CDNA 4) 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 2,351 tok/s/chip for $0.18 per million tokens at 53 tok/s/user on Qwen 3.5 397B-A17B (MI355X FP4); Off posts 1,505 tok/s/chip for $0.28. MTP delivers 56% more tok/s/chip on the same chip, which makes it 36% cheaper per token. Draft-token acceptance rates determine whether speculative decoding helps or hurts at a given concurrency level.

Throughput at 77 tok/s/user on Qwen 3.5 397B-A17B (MI355X FP4): MTP hits 1,963 tok/s/chip, Off hits 909. Costs per million tokens land at $0.21 and $0.46 respectively. MTP delivers 116% more tok/s/chip on the same chip, which makes it 54% 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 30–123 tok/s/user interactivity band, at 100 tok/s/user on Qwen 3.5 397B-A17B (MI355X FP4): MTP runs 1,459 tok/s/chip at $0.29 per million tokens, Off runs 563 at $0.74. MTP delivers 159% more tok/s/chip on the same chip, which makes it 61% 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 · 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)
MTP2,351Off1,505
MTP1,963Off909
MTP1,459Off563
Cost ($/M tok)
MTP$0.18Off$0.28
MTP$0.21Off$0.46
MTP$0.29Off$0.74
tok/s/MW
MTP1,124,716Off720,132
MTP939,462Off434,913
MTP697,900Off269,311
Concurrency
MTP~45Off~29
MTP~26Off~11
MTP~15Off~8
Qwen 3.5 397B-A17B: MI355X FP4 MTP versus Off speculative decoding comparison at matched interactivity levels
MI355X FP4 MTP versus Off speculative decoding comparison for this page's canonical default workload
Open full-size image ↗