DeepSeek V4.1 Flash 552B β H200 vs MI325X
Head-to-head AI inference benchmark comparison of H200 (NVIDIA Hopper) and MI325X (AMD CDNA 3) on DeepSeek V4.1 Flash 552B. 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.
AgentX replays real coding-agent sessions rather than fixed-length prompts, so context grows turn over turn and most of each request is served from cache instead of being recomputed. That turns the comparison into a systems question: KV transfer between nodes, prefix-aware routing, and cache capacity all move the curve alongside raw chip throughput. Fixed-sequence workloads stay the clean baseline for kernel and silicon performance, so the two scenarios answer different questions about the same hardware. Learn more about AgentX β
At 64 tok/s/user interactivity on DeepSeek V4.1 Flash 552B, H200 delivers 35258 tok/s/chip at $0.01 per million tokens; MI325X delivers 17029 tok/s/chip at $0.02. H200 is 87% cheaper per token; H200 delivers 107% more tok/s/chip at this point.
H200 posts 22212 tok/s/chip for $0.02 per million tokens at 112 tok/s/user on DeepSeek V4.1 Flash 552B; MI325X posts 8223 tok/s/chip for $0.04. H200 is 144% cheaper per token; H200 delivers 170% more tok/s/chip.
Throughput at 161 tok/s/user on DeepSeek V4.1 Flash 552B: H200 hits 12237 tok/s/chip, MI325X hits 4307. Per-million costs land at $0.03 and $0.07 respectively. H200 is 156% cheaper per token; H200 delivers 184% more tok/s/chip. (Numbers reflect the default agentic-traces Β· fp4 selection for this URL β table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | H200:35257.7MI325X:17029.4 | H200:22211.6MI325X:8223.3 | H200:12237.0MI325X:4306.7 |
| Cost ($/M tok) | H200:$0.010MI325X:$0.018 | H200:$0.015MI325X:$0.037 | H200:$0.028MI325X:$0.071 |
| tok/s/MW | H200:25735570MI325X:10076539 | H200:16212844MI325X:4865866 | H200:8932087MI325X:2548336 |
| Concurrency | H200:~32MI325X:~6 | H200:~12MI325X:~2 | H200:~7MI325X:~2 |
Inference Performance
Agentic inference metrics from the AgentX scenario and fixed-sequence inference metrics across models, hardware configurations, and serving parameters.