GLM 5/5.1 · Speculative decoding

H200 FP8: MTP vs Off speculative decoding

Speculative decoding comparison of MTP versus Off on H200 FP8 (NVIDIA Hopper) 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.

Throughput at 41 tok/s/user on GLM 5/5.1 (H200 FP8): MTP hits 329 tok/s/chip, Off hits 145. Costs per million tokens land at $1.03 and $2.34 respectively. MTP delivers 128% more tok/s/chip on the same chip, which makes it 56% 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 34–60 tok/s/user interactivity band, at 47 tok/s/user on GLM 5/5.1 (H200 FP8): MTP runs 286 tok/s/chip at $1.19 per million tokens, Off runs 113 at $3.00. MTP delivers 153% 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.

At 54 tok/s/user on GLM 5/5.1 (H200 FP8), MTP delivers 247 tok/s/chip at $1.37 per million tokens; Off delivers 82 tok/s/chip at $4.14. MTP delivers 3.0× the tok/s/chip on the same chip, which makes it 67% 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 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)
MTP329Off145
MTP286Off113
MTP247Off82
Cost ($/M tok)
MTP$1.03Off$2.34
MTP$1.19Off$3.00
MTP$1.37Off$4.14
tok/s/MW
MTP240,498Off105,518
MTP208,666Off82,381
MTP180,539Off59,680
Concurrency
MTP~34Off~14
MTP~24Off~10
MTP~19Off~6
GLM 5/5.1: H200 FP8 MTP versus Off speculative decoding comparison at matched interactivity levels
H200 FP8 MTP versus Off speculative decoding comparison for this page's canonical default workload
Open full-size image ↗