Kimi K2.5/K2.6/K2.7-Code 1T — GB200 NVL72 vs MI325X
Head-to-head AI inference benchmark comparison of GB200 NVL72 (NVIDIA Blackwell) and MI325X (AMD CDNA 3) on Kimi K2.5/K2.6/K2.7-Code 1T. 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.
GB200 NVL72 hits 8980 tok/s/chip for $0.06 per million tokens at 64 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T. No MI325X data at this operating point.
GB200 NVL72: 1545 tok/s/chip, $0.33 per million tokens at 103 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T. MI325X is unmeasured here.
At 141 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T, GB200 NVL72 delivers 733 tok/s/chip at $0.69 per million tokens; MI325X hasn't been benchmarked at this target. (Numbers reflect the default 8k/1k · fp4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | GB200 NVL72:8979.7MI325X:— | GB200 NVL72:1545.4MI325X:— | GB200 NVL72:732.9MI325X:— |
| Cost ($/M tok) | GB200 NVL72:$0.057MI325X:— | GB200 NVL72:$0.334MI325X:— | GB200 NVL72:$0.688MI325X:— |
| tok/s/MW | GB200 NVL72:4801979MI325X:— | GB200 NVL72:826425MI325X:— | GB200 NVL72:391932MI325X:— |
| Concurrency | GB200 NVL72:~1203MI325X:— | GB200 NVL72:~167MI325X:— | GB200 NVL72:~27MI325X:— |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.