GLM 5/5.1 · Speculative Decoding
GB300 NVL72 FP8: MTP vs Off Speculative Decoding
Speculative decoding comparison of MTP versus Off on GB300 NVL72 FP8 (NVIDIA Blackwell) 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 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 6,418 tok/s/chip for $0.100 per million tokens at 42 tok/s/user on GLM 5/5.1 (GB300 NVL72 FP8); Off posts 3,237 tok/s/chip for $0.20. MTP delivers 98% more tok/s/chip on the same chip, which makes it 50% cheaper per token. Draft-token acceptance rates determine whether speculative decoding helps or hurts at a given concurrency level.
Throughput at 50 tok/s/user on GLM 5/5.1 (GB300 NVL72 FP8): MTP hits 5,729 tok/s/chip, Off hits 510. Costs per million tokens land at $0.11 and $1.26 respectively. MTP delivers 11× the tok/s/chip on the same chip, which makes it 91% 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 35–65 tok/s/user interactivity band, at 57 tok/s/user on GLM 5/5.1 (GB300 NVL72 FP8): MTP runs 3,576 tok/s/chip at $0.18 per million tokens, Off runs 233 at $2.75. MTP delivers 15× the tok/s/chip on the same chip, which makes it 93% cheaper per token. Gains from speculative decoding vary by workload; short-output prompts tend to benefit less. (Numbers reflect this URL's pinned 8k/1k · fp8 workload — changing sequence or model updates both the table and chart; the table stays pinned to this page's precision, so precision toggles in the controls affect the chart only.)
Interpolated from real benchmark data. Edit target interactivity values below to compare at different operating points.
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | MTP6,418Off3,237 | MTP5,729Off510 | MTP3,576Off233 |
| Cost ($/M tok) | MTP$0.100Off$0.20 | MTP$0.11Off$1.26 | MTP$0.18Off$2.75 |
| tok/s/MW | MTP3,027,312Off1,527,004 | MTP2,702,536Off240,530 | MTP1,686,798Off109,954 |
| Concurrency | MTP~2,312Off~799 | MTP~1,721Off~132 | MTP~867Off~36 |
