Llama 3.3 70B · Chip comparison
Llama 3.3 70B — H200 vs MI300X
Head-to-head AI inference benchmark comparison of H200 (NVIDIA Hopper) and MI300X (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 — same interactions as the main inference chart.
At 32 tok/s/user interactivity on Llama 3.3 70B, H200 delivers 3,631 tok/s/chip at $0.093 per million tokens; MI300X delivers 1,918 tok/s/chip at $0.14. H200 is 32% cheaper per token; H200 delivers 89% more tok/s/chip at this point.
H200 posts 2,611 tok/s/chip for $0.13 per million tokens at 55 tok/s/user on Llama 3.3 70B; MI300X posts 1,268 tok/s/chip for $0.21. H200 is 38% cheaper per token; H200 delivers 106% more tok/s/chip.
Throughput at 78 tok/s/user on Llama 3.3 70B: H200 hits 1,864 tok/s/chip, MI300X hits 688. Costs per million tokens land at $0.18 and $0.38 respectively. H200 is 53% cheaper per token; H200 delivers 171% more tok/s/chip. (Numbers reflect the default 8k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
Interpolated from real benchmark data. Edit target interactivity values below to compare at different operating points.
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | H2003,631MI300X1,918 | H2002,611MI300X1,268 | H2001,864MI300X688 |
| Cost ($/M tok) | H200$0.093MI300X$0.14 | H200$0.13MI300X$0.21 | H200$0.18MI300X$0.38 |
| tok/s/MW | H2002,650,223MI300X1,379,734 | H2001,905,520MI300X912,473 | H2001,360,591MI300X495,291 |
| Concurrency | H200~32MI300X~31 | H200~26MI300X~16 | H200~10MI300X~10 |