GB200 NVL72 FP4: MTP vs Off Speculative Decoding
Speculative decoding comparison of MTP versus Off on GB200 NVL72 FP4 (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 54 tok/s/user on GLM 5/5.1 (GB200 NVL72 FP4): MTP hits 7883 tok/s/GPU, Off hits 1566. Per-million costs land at $0.08 and $0.39 respectively. MTP is 397% cheaper per token; MTP delivers 403% 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 45–84 tok/s/user interactivity band, at 64 tok/s/user on GLM 5/5.1 (GB200 NVL72 FP4): MTP runs 6816 tok/s/GPU at $0.09/M tokens, Off runs 1193 at $0.52/M. MTP is 468% cheaper per token; MTP delivers 471% more tok/s/GPU. Gains from speculative decoding vary by workload; short-output prompts tend to benefit less.
At 75 tok/s/user on GLM 5/5.1 (GB200 NVL72 FP4), MTP delivers 4214 tok/s/GPU at $0.15 per million tokens; Off delivers 487 tok/s/GPU at $1.46. MTP is 899% cheaper per token; MTP delivers 765% more tok/s/GPU. Speculative decoding accepts draft tokens to reduce per-token latency — gains vary by workload and prompt distribution. (Numbers reflect this URL's pinned 8k/1k · fp4 workload — changing sequence or model updates both the table and chart; the table stays pinned to this page's precision, so precision toggles in the controls affect the chart only.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | MTP:7883.0Off:1566.1 | MTP:6816.2Off:1192.9 | MTP:4213.7Off:487.2 |
| Cost ($/M tok) | MTP:$0.079Off:$0.392 | MTP:$0.091Off:$0.516 | MTP:$0.147Off:$1.463 |
| tok/s/MW | MTP:3753800Off:745764 | MTP:3245831Off:568045 | MTP:2006535Off:231983 |
| Concurrency | MTP:~1529Off:~131 | MTP:~949Off:~129 | MTP:~368Off:~100 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.