Head-to-head AI inference benchmark comparison of H100 (NVIDIA Hopper) and MI325X (AMD CDNA 3) on Llama 3.3 70B. Latency, throughput, and cost across LLM workloads. Use the chart controls below to switch sequences, precisions, and metrics β same interactions as the main inference chart.
At 38 tok/s/user interactivity on Llama 3.3 70B, H100 delivers 1848 tok/s/chip at $0.18 per million tokens; MI325X delivers 1696 tok/s/chip at $0.18. H100 is 2% cheaper per token; H100 delivers 9% more tok/s/chip at this point.
H100 posts 1171 tok/s/chip for $0.28 per million tokens at 58 tok/s/user on Llama 3.3 70B; MI325X posts 1354 tok/s/chip for $0.23. MI325X is 23% cheaper per token; MI325X delivers 16% more tok/s/chip.
Throughput at 78 tok/s/user on Llama 3.3 70B: H100 hits 877 tok/s/chip, MI325X hits 887. Per-million costs land at $0.37 and $0.34 respectively. MI325X is 8% cheaper per token; MI325X delivers 1% more tok/s/chip. (Numbers reflect the default 8k/1k Β· fp8 selection for this URL β table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | H100:1847.6MI325X:1696.0 | H100:1171.4MI325X:1354.4 | H100:877.4MI325X:887.2 |
| Cost ($/M tok) | H100:$0.176MI325X:$0.180 | H100:$0.277MI325X:$0.226 | H100:$0.370MI325X:$0.344 |
| tok/s/MW | H100:1348638MI325X:1003526 | H100:855073MI325X:801393 | H100:640410MI325X:524995 |
| Concurrency | H100:~20MI325X:~25 | H100:~16MI325X:~25 | H100:~8MI325X:~11 |
Agentic inference metrics from the AgentX scenario and fixed-sequence inference metrics across models, hardware configurations, and serving parameters.
TCO $/chip/hr:
Source: SemiAnalysis Market July 2026 Pricing Surveys & AI Cloud TCO Model
Note: Disaggregated inference configurations (e.g., MoRI SGLang, Dynamo TRTLLM) calculate cost per million tokens per decode chip or per prefill chip, rather than per total chip count. This makes direct token cost comparison with aggregated configs not an apples-to-apples comparison.
Note: Disaggregated inference configurations (e.g., MoRI SGLang, Dynamo TRTLLM) calculate input throughput per decode chip or per prefill chip, rather than per total chip count. This makes direct input throughput comparison with aggregated configs not an apples-to-apples comparison.
Note: Disaggregated inference configurations (e.g., MoRI SGLang, Dynamo TRTLLM) calculate output throughput per decode chip or per prefill chip, rather than per total chip count. This makes direct output throughput comparison with aggregated configs not an apples-to-apples comparison.
Note: Disaggregated inference configurations (e.g., MoRI SGLang, Dynamo TRTLLM) calculate power per decode chip or per prefill chip, rather than per total chip count. This makes direct power comparison with aggregated configs not an apples-to-apples comparison.
Note: Disaggregated inference configurations (e.g., MoRI SGLang, Dynamo TRTLLM) calculate Joules per decode chip or per prefill chip, rather than per total chip count. This makes direct Joules per token comparison with aggregated configs not an apples-to-apples comparison.
No measurements to plot for this selection. Review the benchmark controls above or adjust quick filters.
Shift+Scroll to zoom β’ Drag to pan β’ Double-click to reset β’ Click a point to pin tooltip
Llama 3.3 70B Instruct 8K / 1K Total Tokens per $1 TCO vs. Interactivity