GLM 5/5.1 · Speculative Decoding

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 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.

Throughput at 50 tok/s/user on GLM 5/5.1 (GB200 NVL72 FP4): MTP hits 9,984 tok/s/chip, Off hits 2,207. Costs per million tokens land at $0.052 and $0.23 respectively. MTP delivers 4.5× the tok/s/chip on the same chip, which makes it 78% cheaper per token. 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 38–86 tok/s/user interactivity band, at 62 tok/s/user on GLM 5/5.1 (GB200 NVL72 FP4): MTP runs 9,414 tok/s/chip at $0.055 per million tokens, Off runs 1,209 at $0.43. MTP delivers 7.8× the tok/s/chip on the same chip, which makes it 87% cheaper per token. Gains from speculative decoding vary by workload; short-output prompts tend to benefit less.

At 74 tok/s/user on GLM 5/5.1 (GB200 NVL72 FP4), MTP delivers 8,514 tok/s/chip at $0.061 per million tokens; Off delivers 466 tok/s/chip at $1.11. MTP delivers 18× the tok/s/chip on the same chip, which makes it 95% cheaper per token. 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.)

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)
MTP9,984Off2,207
MTP9,414Off1,209
MTP8,514Off466
Cost ($/M tok)
MTP$0.052Off$0.23
MTP$0.055Off$0.43
MTP$0.061Off$1.11
tok/s/MW
MTP5,338,994Off1,180,133
MTP5,034,330Off646,310
MTP4,552,911Off249,075
Concurrency
MTP~1,340Off~227
MTP~818Off~105
MTP~618Off~74
GLM 5/5.1: 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 ↗