MiniMax M3 428B · Speculative decoding

MI355X FP8: M3 EAGLE vs Off speculative decoding

Speculative decoding comparison of M3 EAGLE versus Off on MI355X FP8 (AMD CDNA 4) 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.

Near the low end of the 8–148 tok/s/user interactivity band, at 43 tok/s/user on MiniMax M3 428B (MI355X FP8): M3 EAGLE runs 6,274 tok/s/chip at $0.066 per million tokens, Off runs 4,374 at $0.095. M3 EAGLE delivers 43% more tok/s/chip on the same chip, which makes it 30% cheaper per token. Gains from speculative decoding vary by workload; short-output prompts tend to benefit less.

At 78 tok/s/user on MiniMax M3 428B (MI355X FP8), M3 EAGLE delivers 4,409 tok/s/chip at $0.094 per million tokens; Off delivers 2,114 tok/s/chip at $0.20. M3 EAGLE delivers 109% more tok/s/chip on the same chip, which makes it 52% cheaper per token. Speculative decoding accepts draft tokens to reduce per-token latency — gains vary by workload and prompt distribution.

M3 EAGLE posts 3,359 tok/s/chip for $0.12 per million tokens at 114 tok/s/user on MiniMax M3 428B (MI355X FP8); Off posts 1,125 tok/s/chip for $0.37. M3 EAGLE delivers 199% more tok/s/chip on the same chip, which makes it 67% 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 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 EAGLE6,274Off4,374
M3 EAGLE4,409Off2,114
M3 EAGLE3,359Off1,125
Cost ($/M tok)
M3 EAGLE$0.066Off$0.095
M3 EAGLE$0.094Off$0.20
M3 EAGLE$0.12Off$0.37
tok/s/MW
M3 EAGLE3,002,063Off2,093,040
M3 EAGLE2,109,720Off1,011,611
M3 EAGLE1,607,312Off538,256
Concurrency
M3 EAGLE~78Off~95
M3 EAGLE~30Off~12
M3 EAGLE~16Off~5
MiniMax M3 428B: MI355X FP8 M3 EAGLE versus Off speculative decoding comparison at matched interactivity levels
MI355X FP8 M3 EAGLE versus Off speculative decoding comparison for this page's canonical default workload
Open full-size image ↗