MiniMax M3 428B — GB300 NVL72 vs MI300X
Head-to-head AI inference benchmark comparison of GB300 NVL72 (NVIDIA Blackwell) and MI300X (AMD CDNA 3) on MiniMax M3 428B. Latency, throughput, and cost across LLM workloads. Use the chart controls below to switch sequences, precisions, and metrics — same interactions as the main inference chart.
Throughput at 47 tok/s/user on MiniMax M3 428B: GB300 NVL72 hits 2518 tok/s/GPU, MI300X hits 517. Per-million costs land at $0.28 and $0.59 respectively. GB300 NVL72 is 110% cheaper per token; GB300 NVL72 delivers 387% more tok/s/GPU.
GB300 NVL72 / MI300X on MiniMax M3 428B at 81 tok/s/user: 808 / 202 tok/s/GPU, $0.92 / $1.53 per million tokens. GB300 NVL72 is 67% cheaper per token; GB300 NVL72 delivers 299% more tok/s/GPU.
Toward the upper edge of the 13–149 tok/s/user interactivity band, at 115 tok/s/user on MiniMax M3 428B: GB300 NVL72 runs 312 tok/s/GPU at $2.34/M tokens, MI300X runs 142 at $2.14/M. MI300X is 9% cheaper per token; GB300 NVL72 delivers 119% more tok/s/GPU. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | GB300 NVL72:2517.9MI300X:517.4 | GB300 NVL72:807.5MI300X:202.2 | GB300 NVL72:312.3MI300X:142.4 |
| Cost ($/M tok) | GB300 NVL72:$0.280MI300X:$0.590 | GB300 NVL72:$0.921MI300X:$1.534 | GB300 NVL72:$2.342MI300X:$2.145 |
| tok/s/MW | GB300 NVL72:1187700MI300X:372210 | GB300 NVL72:380902MI300X:145436 | GB300 NVL72:147310MI300X:102431 |
| Concurrency | GB300 NVL72:~649MI300X:~47 | GB300 NVL72:~96MI300X:~11 | GB300 NVL72:~29MI300X:~5 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.