GLM 5/5.1 · Speculative Decoding

GB300 NVL72 FP4: MTP vs Off Speculative Decoding

Speculative decoding comparison of MTP versus Off on GB300 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.

Near the low end of the 21–104 tok/s/user interactivity band, at 42 tok/s/user on GLM 5/5.1 (GB300 NVL72 FP4): MTP runs 10,234 tok/s/chip at $0.063 per million tokens, Off runs 5,653 at $0.11. MTP delivers 81% more tok/s/chip on the same chip, which makes it 45% cheaper per token. Gains from speculative decoding vary by workload; short-output prompts tend to benefit less.

At 63 tok/s/user on GLM 5/5.1 (GB300 NVL72 FP4), MTP delivers 7,527 tok/s/chip at $0.085 per million tokens; Off delivers 429 tok/s/chip at $1.50. MTP delivers 18× the tok/s/chip on the same chip, which makes it 94% cheaper per token. Speculative decoding accepts draft tokens to reduce per-token latency — gains vary by workload and prompt distribution.

MTP posts 4,517 tok/s/chip for $0.14 per million tokens at 84 tok/s/user on GLM 5/5.1 (GB300 NVL72 FP4); Off posts 92 tok/s/chip for $7.01. MTP delivers 49× the tok/s/chip on the same chip, which makes it 98% 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)
MTP10,234Off5,653
MTP7,527Off429
MTP4,517Off92
Cost ($/M tok)
MTP$0.063Off$0.11
MTP$0.085Off$1.50
MTP$0.14Off$7.01
tok/s/MW
MTP4,827,317Off2,666,668
MTP3,550,446Off202,127
MTP2,130,446Off43,199
Concurrency
MTP~3,770Off~2,253
MTP~1,486Off~104
MTP~721Off~35
GLM 5/5.1: GB300 NVL72 FP4 MTP versus Off speculative decoding comparison at matched interactivity levels
GB300 NVL72 FP4 MTP versus Off speculative decoding comparison for this page's canonical default workload
Open full-size image ↗