Kimi K2.5/K2.6/K2.7-Code 1T — MI300X vs MI355X
Head-to-head AI inference benchmark comparison of MI300X (AMD CDNA 3) 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.
MI300X posts 112 tok/s/chip for $2.34 per million tokens at 27 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T; MI355X posts 825 tok/s/chip for $0.50. MI355X is 363% cheaper per token; MI355X delivers 634% more tok/s/chip.
Throughput at 34 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T: MI300X hits 90 tok/s/chip, MI355X hits 628. Per-million costs land at $2.94 and $0.66 respectively. MI355X is 344% cheaper per token; MI355X delivers 598% more tok/s/chip.
MI300X / MI355X on Kimi K2.5/K2.6/K2.7-Code 1T at 42 tok/s/user: 65 / 480 tok/s/chip, $4.12 / $0.86 per million tokens. MI355X is 377% cheaper per token; MI355X delivers 639% more tok/s/chip. (Numbers reflect the default 1k/1k · int4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | MI300X:112.4MI355X:825.2 | MI300X:90.1MI355X:628.5 | MI300X:64.9MI355X:479.5 |
| Cost ($/M tok) | MI300X:$2.336MI355X:$0.504 | MI300X:$2.941MI355X:$0.663 | MI300X:$4.125MI355X:$0.864 |
| tok/s/MW | MI300X:80834MI355X:394829 | MI300X:64804MI355X:300702 | MI300X:46707MI355X:229430 |
| Concurrency | MI300X:~17MI355X:~62 | MI300X:~11MI355X:~38 | MI300X:~6MI355X:~23 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.