MiniMax M3 428B · Speculative decoding

MI300X FP8: M3 EAGLE vs Off speculative decoding

Speculative decoding comparison of M3 EAGLE versus Off on MI300X 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.

M3 EAGLE posts 1,132 tok/s/chip for $0.23 per million tokens at 30 tok/s/user on MiniMax M3 428B (MI300X FP8); Off posts 1,013 tok/s/chip for $0.26. M3 EAGLE delivers 12% more tok/s/chip on the same chip, which makes it 11% cheaper per token. Draft-token acceptance rates determine whether speculative decoding helps or hurts at a given concurrency level.

Throughput at 53 tok/s/user on MiniMax M3 428B (MI300X FP8): M3 EAGLE hits 676 tok/s/chip, Off hits 627. Costs per million tokens land at $0.39 and $0.42 respectively. M3 EAGLE delivers 8% more tok/s/chip on the same chip, which makes it 7% cheaper per token. Speculative decoding trades extra compute on draft tokens for fewer decoding steps — the payoff depends on sequence length and batch size.

Toward the upper edge of the 8–97 tok/s/user interactivity band, at 75 tok/s/user on MiniMax M3 428B (MI300X FP8): M3 EAGLE runs 631 tok/s/chip at $0.42 per million tokens, Off runs 407 at $0.65. M3 EAGLE delivers 55% more tok/s/chip on the same chip, which makes it 36% cheaper per token. Gains from speculative decoding vary by workload; short-output prompts tend to benefit less. (Numbers reflect this URL's pinned 8k/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 EAGLE1,132Off1,013
M3 EAGLE676Off627
M3 EAGLE631Off407
Cost ($/M tok)
M3 EAGLE$0.23Off$0.26
M3 EAGLE$0.39Off$0.42
M3 EAGLE$0.42Off$0.65
tok/s/MW
M3 EAGLE814,199Off728,533
M3 EAGLE486,038Off451,023
M3 EAGLE454,263Off292,539
Concurrency
M3 EAGLE~41Off~31
M3 EAGLE~13Off~11
M3 EAGLE~9Off~5
MiniMax M3 428B: MI300X FP8 M3 EAGLE versus Off speculative decoding comparison at matched interactivity levels
MI300X FP8 M3 EAGLE versus Off speculative decoding comparison for this page's canonical default workload
Open full-size image ↗