GB200 NVL72 FP8: MTP vs Off Speculative Decoding
Speculative decoding comparison of MTP versus Off on GB200 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.
Throughput at 59 tok/s/user on GLM 5/5.1 (GB200 NVL72 FP8): MTP hits 3865 tok/s/GPU, Off hits 59. Per-million costs land at $0.16 and $10.4 respectively. MTP is 6522% cheaper per token; MTP delivers 6402% more tok/s/GPU. 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 58–62 tok/s/user interactivity band, at 60 tok/s/user on GLM 5/5.1 (GB200 NVL72 FP8): MTP runs 3675 tok/s/GPU at $0.16/M tokens, Off runs 49 at $12.0/M. MTP is 7188% cheaper per token; MTP delivers 7381% more tok/s/GPU. Gains from speculative decoding vary by workload; short-output prompts tend to benefit less.
At 61 tok/s/user on GLM 5/5.1 (GB200 NVL72 FP8), MTP delivers 3470 tok/s/GPU at $0.18 per million tokens; Off delivers 43 tok/s/GPU at $13.7. MTP is 7725% cheaper per token; MTP delivers 7962% more tok/s/GPU. Speculative decoding accepts draft tokens to reduce per-token latency — gains vary by workload and prompt distribution. (Numbers reflect this URL's pinned 8k/1k · fp8 workload — changing sequence or model updates both the table and chart; the table stays pinned to this page's precision, so precision toggles in the controls affect the chart only.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | MTP:3864.7Off:59.4 | MTP:3675.1Off:49.1 | MTP:3469.5Off:43.0 |
| Cost ($/M tok) | MTP:$0.157Off:$10.406 | MTP:$0.165Off:$11.999 | MTP:$0.175Off:$13.725 |
| tok/s/MW | MTP:2066700Off:31788 | MTP:1965277Off:26271 | MTP:1855350Off:23014 |
| Concurrency | MTP:~1069Off:~9 | MTP:~882Off:~9 | MTP:~712Off:~8 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.