DeepSeek R1 · Speculative decoding

GB300 NVL72 FP8: MTP vs Off speculative decoding

Speculative decoding comparison of MTP versus Off on GB300 NVL72 FP8 (NVIDIA Blackwell) running DeepSeek R1. 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.

At 60 tok/s/user on DeepSeek R1 (GB300 NVL72 FP8), MTP delivers 8,373 tok/s/chip at $0.077 per million tokens; Off delivers 3,004 tok/s/chip at $0.21. MTP delivers 179% more tok/s/chip on the same chip, which makes it 64% cheaper per token. Speculative decoding accepts draft tokens to reduce per-token latency — gains vary by workload and prompt distribution.

MTP posts 6,186 tok/s/chip for $0.10 per million tokens at 84 tok/s/user on DeepSeek R1 (GB300 NVL72 FP8); Off posts 634 tok/s/chip for $1.01. MTP delivers 9.8× the tok/s/chip on the same chip, which makes it 90% cheaper per token. Draft-token acceptance rates determine whether speculative decoding helps or hurts at a given concurrency level.

Throughput at 107 tok/s/user on DeepSeek R1 (GB300 NVL72 FP8): MTP hits 5,055 tok/s/chip, Off hits 357. Costs per million tokens land at $0.13 and $1.80 respectively. MTP delivers 14× the tok/s/chip on the same chip, which makes it 93% cheaper per token. Speculative decoding trades extra compute on draft tokens for fewer decoding steps — the payoff depends on sequence length and batch size. (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)
MTP8,373Off3,004
MTP6,186Off634
MTP5,055Off357
Cost ($/M tok)
MTP$0.077Off$0.21
MTP$0.10Off$1.01
MTP$0.13Off$1.80
tok/s/MW
MTP3,949,326Off1,417,211
MTP2,918,029Off299,207
MTP2,384,349Off168,194
Concurrency
MTP~1,171Off~511
MTP~535Off~30
MTP~383Off~23
DeepSeek R1: GB300 NVL72 FP8 MTP versus Off speculative decoding comparison at matched interactivity levels
GB300 NVL72 FP8 MTP versus Off speculative decoding comparison for this page's canonical default workload
Open full-size image ↗