MiniMax M3 428B · Speculative decoding

MI325X FP8: M3 EAGLE vs Off speculative decoding

Speculative decoding comparison of M3 EAGLE versus Off on MI325X FP8 (AMD CDNA 3) running MiniMax M3 428B. 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 33 tok/s/user on MiniMax M3 428B (MI325X FP8), M3 EAGLE delivers 913 tok/s/chip at $0.33 per million tokens; Off delivers 405 tok/s/chip at $0.75. M3 EAGLE delivers 125% more tok/s/chip on the same chip, which makes it 56% cheaper per token. Speculative decoding accepts draft tokens to reduce per-token latency — gains vary by workload and prompt distribution.

M3 EAGLE posts 527 tok/s/chip for $0.58 per million tokens at 54 tok/s/user on MiniMax M3 428B (MI325X FP8); Off posts 245 tok/s/chip for $1.25. M3 EAGLE delivers 116% more tok/s/chip on the same chip, which makes it 54% cheaper per token. Draft-token acceptance rates determine whether speculative decoding helps or hurts at a given concurrency level.

Throughput at 74 tok/s/user on MiniMax M3 428B (MI325X FP8): M3 EAGLE hits 310 tok/s/chip, Off hits 137. Costs per million tokens land at $0.99 and $2.22 respectively. M3 EAGLE delivers 125% more tok/s/chip on the same chip, which makes it 56% 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 1k/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)
M3 EAGLE913Off405
M3 EAGLE527Off245
M3 EAGLE310Off137
Cost ($/M tok)
M3 EAGLE$0.33Off$0.75
M3 EAGLE$0.58Off$1.25
M3 EAGLE$0.99Off$2.22
tok/s/MW
M3 EAGLE540,424Off239,719
M3 EAGLE311,959Off144,739
M3 EAGLE183,142Off81,324
Concurrency
M3 EAGLE~120Off~50
M3 EAGLE~43Off~19
M3 EAGLE~18Off~8
MiniMax M3 428B: MI325X FP8 M3 EAGLE versus Off speculative decoding comparison at matched interactivity levels
MI325X FP8 M3 EAGLE versus Off speculative decoding comparison for this page's canonical default workload
Open full-size image ↗