DeepSeek V4 Pro 1.6T — GB200 NVL72 vs MI355X
Head-to-head AI inference benchmark comparison of GB200 NVL72 (NVIDIA Blackwell) and MI355X (AMD CDNA 4) on DeepSeek V4 Pro 1.6T. 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 DeepSeek V4 Pro 1.6T: GB200 NVL72 hits 9562 tok/s/chip, MI355X hits 3027. Per-million costs land at $0.05 and $0.14 respectively. GB200 NVL72 is 156% cheaper per token; GB200 NVL72 delivers 216% more tok/s/chip.
GB200 NVL72 / MI355X on DeepSeek V4 Pro 1.6T at 75 tok/s/user: 8616 / 1613 tok/s/chip, $0.06 / $0.26 per million tokens. GB200 NVL72 is 328% cheaper per token; GB200 NVL72 delivers 434% more tok/s/chip.
Toward the upper edge of the 18–133 tok/s/user interactivity band, at 104 tok/s/user on DeepSeek V4 Pro 1.6T: GB200 NVL72 runs 6714 tok/s/chip at $0.08/M tokens, MI355X runs 777 at $0.53/M. GB200 NVL72 is 588% cheaper per token; GB200 NVL72 delivers 764% more tok/s/chip. (Numbers reflect the default 8k/1k · fp4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | GB200 NVL72:9562.2MI355X:3027.2 | GB200 NVL72:8615.9MI355X:1613.3 | GB200 NVL72:6714.0MI355X:777.4 |
| Cost ($/M tok) | GB200 NVL72:$0.054MI355X:$0.138 | GB200 NVL72:$0.060MI355X:$0.257 | GB200 NVL72:$0.077MI355X:$0.531 |
| tok/s/MW | GB200 NVL72:5113476MI355X:1448401 | GB200 NVL72:4607427MI355X:771914 | GB200 NVL72:3590381MI355X:371948 |
| Concurrency | GB200 NVL72:~15421MI355X:~100 | GB200 NVL72:~6302MI355X:~27 | GB200 NVL72:~2196MI355X:~7 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.