Qwen 3.5 397B-A17B — H100 vs MI325X
Head-to-head AI inference benchmark comparison of H100 (NVIDIA Hopper) and MI325X (AMD CDNA 3) on Qwen 3.5 397B-A17B. 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.
Throughput at 46 tok/s/user on Qwen 3.5 397B-A17B: H100 hits 974 tok/s/chip, MI325X hits 403. Per-million costs land at $0.33 and $0.76 respectively. H100 is 127% cheaper per token; H100 delivers 142% more tok/s/chip.
H100 / MI325X on Qwen 3.5 397B-A17B at 55 tok/s/user: 678 / 219 tok/s/chip, $0.48 / $1.39 per million tokens. H100 is 190% cheaper per token; H100 delivers 209% more tok/s/chip.
Toward the upper edge of the 37–72 tok/s/user interactivity band, at 64 tok/s/user on Qwen 3.5 397B-A17B: H100 runs 433 tok/s/chip at $0.75/M tokens, MI325X runs 134 at $2.28/M. H100 is 203% cheaper per token; H100 delivers 223% more tok/s/chip. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | H100:973.7MI325X:402.9 | H100:677.7MI325X:219.4 | H100:433.3MI325X:134.3 |
| Cost ($/M tok) | H100:$0.334MI325X:$0.758 | H100:$0.480MI325X:$1.393 | H100:$0.750MI325X:$2.276 |
| tok/s/MW | H100:710725MI325X:238420 | H100:494657MI325X:129810 | H100:316265MI325X:79448 |
| Concurrency | H100:~93MI325X:~37 | H100:~55MI325X:~16 | H100:~28MI325X:~9 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.