Kimi K2.5/K2.6 1T — GB200 NVL72 vs MI355X
Head-to-head AI inference benchmark comparison of GB200 NVL72 (NVIDIA Blackwell) and MI355X (AMD CDNA 4) on Kimi K2.5/K2.6 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.
Setting 46 tok/s/user as the target on Kimi K2.5/K2.6 1T, GB200 NVL72 produces 5003 tok/s/GPU ($0.12 per million tokens) and MI355X produces 726 ($0.56). GB200 NVL72 is 363% cheaper per token; GB200 NVL72 delivers 589% more tok/s/GPU.
At 68 tok/s/user interactivity on Kimi K2.5/K2.6 1T, GB200 NVL72 delivers 1418 tok/s/GPU at $0.43 per million tokens; MI355X delivers 554 tok/s/GPU at $0.74. GB200 NVL72 is 71% cheaper per token; GB200 NVL72 delivers 156% more tok/s/GPU at this point.
GB200 NVL72 posts 728 tok/s/GPU for $0.85 per million tokens at 91 tok/s/user on Kimi K2.5/K2.6 1T; MI355X posts 419 tok/s/GPU for $0.98. GB200 NVL72 is 15% cheaper per token; GB200 NVL72 delivers 74% more tok/s/GPU. (Numbers reflect the default 1k/1k · fp4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | GB200 NVL72:5003.3MI355X:726.2 | GB200 NVL72:1417.9MI355X:553.9 | GB200 NVL72:728.1MI355X:418.8 |
| Cost ($/M tok) | GB200 NVL72:$0.122MI355X:$0.565 | GB200 NVL72:$0.433MI355X:$0.741 | GB200 NVL72:$0.854MI355X:$0.982 |
| tok/s/MW | GB200 NVL72:2382538MI355X:274039 | GB200 NVL72:675195MI355X:209003 | GB200 NVL72:346731MI355X:158047 |
| Concurrency | GB200 NVL72:~1337MI355X:~35 | GB200 NVL72:~511MI355X:~17 | GB200 NVL72:~155MI355X:~9 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.