GLM 5/5.1 · Speculative decoding

MI355X FP8: MTP vs Off speculative decoding

Speculative decoding comparison of MTP versus Off on MI355X FP8 (AMD CDNA 4) running GLM 5/5.1. 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 726 tok/s/chip for $0.57 per million tokens at 42 tok/s/user on GLM 5/5.1 (MI355X FP8); Off posts 193 tok/s/chip for $2.15. MTP delivers 3.8× the tok/s/chip on the same chip, which makes it 73% cheaper per token. Draft-token acceptance rates determine whether speculative decoding helps or hurts at a given concurrency level.

Throughput at 47 tok/s/user on GLM 5/5.1 (MI355X FP8): MTP hits 595 tok/s/chip, Off hits 169. Costs per million tokens land at $0.70 and $2.46 respectively. MTP delivers 3.5× the tok/s/chip on the same chip, which makes it 72% 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 37–56 tok/s/user interactivity band, at 52 tok/s/user on GLM 5/5.1 (MI355X FP8): MTP runs 467 tok/s/chip at $0.89 per million tokens, Off runs 90 at $4.65. MTP delivers 5.2× the tok/s/chip on the same chip, which makes it 81% 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)
MTP726Off193
MTP595Off169
MTP467Off90
Cost ($/M tok)
MTP$0.57Off$2.15
MTP$0.70Off$2.46
MTP$0.89Off$4.65
tok/s/MW
MTP347,245Off92,570
MTP284,615Off81,053
MTP223,584Off42,900
Concurrency
MTP~47Off~16
MTP~25Off~6
MTP~18Off~4
GLM 5/5.1: MI355X FP8 MTP versus Off speculative decoding comparison at matched interactivity levels
MI355X FP8 MTP versus Off speculative decoding comparison for this page's canonical default workload
Open full-size image ↗