Qwen 3.5 397B-A17B — GB300 NVL72 vs MI355X
Head-to-head AI inference benchmark comparison of GB300 NVL72 (NVIDIA Blackwell) and MI355X (AMD CDNA 4) on Qwen 3.5 397B-A17B. 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.
GB300 NVL72 / MI355X on Qwen 3.5 397B-A17B at 75 tok/s/user: 1084 / 1799 tok/s/GPU, $0.68 / $0.23 per million tokens. MI355X is 192% cheaper per token; MI355X delivers 66% more tok/s/GPU.
Around the middle of the 32–206 tok/s/user interactivity band, at 119 tok/s/user on Qwen 3.5 397B-A17B: GB300 NVL72 runs 447 tok/s/GPU at $1.65/M tokens, MI355X runs 1050 at $0.39/M. MI355X is 325% cheaper per token; MI355X delivers 135% more tok/s/GPU.
Setting 163 tok/s/user as the target on Qwen 3.5 397B-A17B, GB300 NVL72 produces 208 tok/s/GPU ($3.46 per million tokens) and MI355X produces 639 ($0.64). MI355X is 438% cheaper per token; MI355X delivers 208% more tok/s/GPU. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | GB300 NVL72:1083.9MI355X:1799.4 | GB300 NVL72:446.9MI355X:1049.5 | GB300 NVL72:207.7MI355X:639.1 |
| Cost ($/M tok) | GB300 NVL72:$0.677MI355X:$0.232 | GB300 NVL72:$1.648MI355X:$0.388 | GB300 NVL72:$3.463MI355X:$0.644 |
| tok/s/MW | GB300 NVL72:511290MI355X:860962 | GB300 NVL72:210784MI355X:502164 | GB300 NVL72:97959MI355X:305794 |
| Concurrency | GB300 NVL72:~63MI355X:~48 | GB300 NVL72:~16MI355X:~18 | GB300 NVL72:~5MI355X:~8 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.