DeepSeek V4 Pro 0813 1.6T · Speculative decoding

GB200 NVL72 FP4: MTP vs Off speculative decoding

Speculative decoding comparison of MTP versus Off on GB200 NVL72 FP4 (NVIDIA Blackwell) running DeepSeek V4 Pro 0813 1.6T. 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 38 tok/s/user on DeepSeek V4 Pro 0813 1.6T (GB200 NVL72 FP4), MTP delivers 9,479 tok/s/chip at $0.055 per million tokens; Off delivers 5,722 tok/s/chip at $0.090. MTP delivers 66% more tok/s/chip on the same chip, which makes it 40% cheaper per token. Speculative decoding accepts draft tokens to reduce per-token latency — gains vary by workload and prompt distribution.

MTP posts 9,479 tok/s/chip for $0.055 per million tokens at 56 tok/s/user on DeepSeek V4 Pro 0813 1.6T (GB200 NVL72 FP4); Off posts 482 tok/s/chip for $1.07. MTP delivers 20× the tok/s/chip on the same chip, which makes it 95% cheaper per token. Draft-token acceptance rates determine whether speculative decoding helps or hurts at a given concurrency level.

Throughput at 75 tok/s/user on DeepSeek V4 Pro 0813 1.6T (GB200 NVL72 FP4): MTP hits 8,616 tok/s/chip, Off hits 188. Costs per million tokens land at $0.060 and $2.75 respectively. MTP delivers 46× the tok/s/chip on the same chip, which makes it 98% 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 · 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.)

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)
MTP9,479Off5,722
MTP9,479Off482
MTP8,616Off188
Cost ($/M tok)
MTP$0.055Off$0.090
MTP$0.055Off$1.07
MTP$0.060Off$2.75
tok/s/MW
MTP5,068,910Off3,059,750
MTP5,068,910Off257,871
MTP4,607,427Off100,642
Concurrency
MTP~16,384Off~1,110
MTP~16,384Off~54
MTP~6,302Off~18
DeepSeek V4 Pro 0813 1.6T: GB200 NVL72 FP4 MTP versus Off speculative decoding comparison at matched interactivity levels
GB200 NVL72 FP4 MTP versus Off speculative decoding comparison for this page's canonical default workload
Open full-size image ↗