Kimi K2.5/K2.6 1T — H200 vs MI300X
Head-to-head AI inference benchmark comparison of H200 (NVIDIA Hopper) and MI300X (AMD CDNA 3) 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.
H200 / MI300X on Kimi K2.5/K2.6 1T at 37 tok/s/user: 464 / 80 tok/s/GPU, $0.86 / $3.90 per million tokens. H200 is 354% cheaper per token; H200 delivers 484% more tok/s/GPU.
Around the middle of the 34–48 tok/s/user interactivity band, at 41 tok/s/user on Kimi K2.5/K2.6 1T: H200 runs 412 tok/s/GPU at $0.97/M tokens, MI300X runs 68 at $4.65/M. H200 is 381% cheaper per token; H200 delivers 509% more tok/s/GPU.
Setting 45 tok/s/user as the target on Kimi K2.5/K2.6 1T, H200 produces 366 tok/s/GPU ($1.07 per million tokens) and MI300X produces 57 ($5.57). H200 is 418% cheaper per token; H200 delivers 540% more tok/s/GPU. (Numbers reflect the default 1k/1k · int4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | H200:464.3MI300X:79.6 | H200:411.8MI300X:67.6 | H200:365.6MI300X:57.1 |
| Cost ($/M tok) | H200:$0.860MI300X:$3.904 | H200:$0.967MI300X:$4.646 | H200:$1.074MI300X:$5.566 |
| tok/s/MW | H200:268380MI300X:44450 | H200:238033MI300X:37789 | H200:211356MI300X:31923 |
| Concurrency | H200:~53MI300X:~9 | H200:~42MI300X:~7 | H200:~34MI300X:~5 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.