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 91 tok/s/user: 10437 / 4006 tok/s/chip, $0.06 / $0.10 per million tokens. GB300 NVL72 is 70% cheaper per token; GB300 NVL72 delivers 161% more tok/s/chip.
Around the middle of the 48–222 tok/s/user interactivity band, at 135 tok/s/user on Qwen 3.5 397B-A17B: GB300 NVL72 runs 4862 tok/s/chip at $0.13/M tokens, MI355X runs 2862 at $0.14/M. GB300 NVL72 is 8% cheaper per token; GB300 NVL72 delivers 70% more tok/s/chip.
Setting 179 tok/s/user as the target on Qwen 3.5 397B-A17B, GB300 NVL72 produces 2817 tok/s/chip ($0.23 per million tokens) and MI355X produces 2153 ($0.20). MI355X is 16% cheaper per token; GB300 NVL72 delivers 31% more tok/s/chip. (Numbers reflect the default 8k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | GB300 NVL72:10437.1MI355X:4006.0 | GB300 NVL72:4861.9MI355X:2861.6 | GB300 NVL72:2817.4MI355X:2153.5 |
| Cost ($/M tok) | GB300 NVL72:$0.061MI355X:$0.104 | GB300 NVL72:$0.134MI355X:$0.144 | GB300 NVL72:$0.227MI355X:$0.197 |
| tok/s/MW | GB300 NVL72:4923177MI355X:1916739 | GB300 NVL72:2293329MI355X:1369182 | GB300 NVL72:1328953MI355X:1030367 |
| Concurrency | GB300 NVL72:~972MI355X:~21 | GB300 NVL72:~95MI355X:~10 | GB300 NVL72:~39MI355X:~5 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.