GLM 5/5.1 — H200 vs MI325X
Head-to-head AI inference benchmark comparison of H200 (NVIDIA Hopper) and MI325X (AMD CDNA 3) on GLM 5/5.1. 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 posts 396 tok/s/chip for $0.86 per million tokens at 24 tok/s/user on GLM 5/5.1; MI325X posts 148 tok/s/chip for $2.12. H200 is 147% cheaper per token; H200 delivers 168% more tok/s/chip.
Throughput at 28 tok/s/user on GLM 5/5.1: H200 hits 396 tok/s/chip, MI325X hits 75. Per-million costs land at $0.86 and $4.14 respectively. H200 is 384% cheaper per token; H200 delivers 431% more tok/s/chip.
H200 / MI325X on GLM 5/5.1 at 31 tok/s/user: 396 / 54 tok/s/chip, $0.86 / $5.63 per million tokens. H200 is 558% cheaper per token; H200 delivers 627% more tok/s/chip. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | H200:395.7MI325X:147.5 | H200:395.7MI325X:74.5 | H200:395.7MI325X:54.5 |
| Cost ($/M tok) | H200:$0.856MI325X:$2.118 | H200:$0.856MI325X:$4.145 | H200:$0.856MI325X:$5.632 |
| tok/s/MW | H200:288810MI325X:87303 | H200:288810MI325X:44089 | H200:288810MI325X:32220 |
| Concurrency | H200:~64MI325X:~25 | H200:~64MI325X:~14 | H200:~64MI325X:~8 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.