Kimi K2.5/K2.6/K2.7-Code 1T — GB300 NVL72 vs MI355X
Head-to-head AI inference benchmark comparison of GB300 NVL72 (NVIDIA Blackwell) and MI355X (AMD CDNA 4) 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.
GB300 NVL72 posts 10326 tok/s/GPU for $0.07 per million tokens at 50 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T; MI355X posts 1099 tok/s/GPU for $0.37. GB300 NVL72 is 427% cheaper per token; GB300 NVL72 delivers 840% more tok/s/GPU.
Throughput at 76 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T: GB300 NVL72 hits 4074 tok/s/GPU, MI355X hits 695. Per-million costs land at $0.18 and $0.59 respectively. GB300 NVL72 is 226% cheaper per token; GB300 NVL72 delivers 486% more tok/s/GPU.
GB300 NVL72 / MI355X on Kimi K2.5/K2.6/K2.7-Code 1T at 102 tok/s/user: 621 / 482 tok/s/GPU, $1.20 / $0.83 per million tokens. MI355X is 44% cheaper per token; GB300 NVL72 delivers 29% more tok/s/GPU. (Numbers reflect the default 1k/1k · fp4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | GB300 NVL72:10326.0MI355X:1098.8 | GB300 NVL72:4073.5MI355X:695.5 | GB300 NVL72:621.5MI355X:482.3 |
| Cost ($/M tok) | GB300 NVL72:$0.071MI355X:$0.375 | GB300 NVL72:$0.181MI355X:$0.589 | GB300 NVL72:$1.199MI355X:$0.832 |
| tok/s/MW | GB300 NVL72:4870763MI355X:525730 | GB300 NVL72:1921477MI355X:332759 | GB300 NVL72:293151MI355X:230753 |
| Concurrency | GB300 NVL72:~4301MI355X:~44 | GB300 NVL72:~1242MI355X:~19 | GB300 NVL72:~135MI355X:~10 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.