DeepSeek R1 · 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 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.

Near the low end of the 19–175 tok/s/user interactivity band, at 57 tok/s/user on DeepSeek R1 (GB200 NVL72 FP4): MTP runs 11,333 tok/s/chip at $0.046 per million tokens, Off runs 6,442 at $0.080. MTP delivers 76% more tok/s/chip on the same chip, which makes it 43% cheaper per token. Gains from speculative decoding vary by workload; short-output prompts tend to benefit less.

At 97 tok/s/user on DeepSeek R1 (GB200 NVL72 FP4), MTP delivers 7,092 tok/s/chip at $0.073 per million tokens; Off delivers 549 tok/s/chip at $0.94. MTP delivers 13× the tok/s/chip on the same chip, which makes it 92% cheaper per token. Speculative decoding accepts draft tokens to reduce per-token latency — gains vary by workload and prompt distribution.

MTP posts 3,059 tok/s/chip for $0.17 per million tokens at 136 tok/s/user on DeepSeek R1 (GB200 NVL72 FP4); Off posts 187 tok/s/chip for $2.76. MTP delivers 16× the tok/s/chip on the same chip, which makes it 94% cheaper per token. Draft-token acceptance rates determine whether speculative decoding helps or hurts at a given concurrency level. (Numbers reflect this URL's pinned 1k/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)
MTP11,333Off6,442
MTP7,092Off549
MTP3,059Off187
Cost ($/M tok)
MTP$0.046Off$0.080
MTP$0.073Off$0.94
MTP$0.17Off$2.76
tok/s/MW
MTP6,060,616Off3,444,905
MTP3,792,765Off293,833
MTP1,635,620Off100,184
Concurrency
MTP~3,397Off~3,248
MTP~2,028Off~130
MTP~686Off~11
DeepSeek R1: 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 ↗