Qwen 3.5 397B-A17B · Chip comparison
Qwen 3.5 397B-A17B — GB300 NVL72 vs MI325X
Head-to-head AI inference benchmark comparison of GB300 NVL72 (NVIDIA Blackwell) and MI325X (AMD CDNA 3) 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.
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 →
At 146 tok/s/user interactivity on Qwen 3.5 397B-A17B, GB300 NVL72 delivers 58,116 tok/s/chip at $0.011 per million tokens; MI325X delivers 5,718 tok/s/chip at $0.053. GB300 NVL72 is 384% cheaper per token; GB300 NVL72 delivers 916% more tok/s/chip at this point.
GB300 NVL72 posts 24,563 tok/s/chip for $0.026 per million tokens at 291 tok/s/user on Qwen 3.5 397B-A17B; MI325X posts 5,718 tok/s/chip for $0.053. GB300 NVL72 is 105% cheaper per token; GB300 NVL72 delivers 330% more tok/s/chip.
Throughput at 436 tok/s/user on Qwen 3.5 397B-A17B: GB300 NVL72 hits 13,914 tok/s/chip, MI325X hits 5,718. Per-million costs land at $0.046 and $0.053 respectively. GB300 NVL72 is 16% cheaper per token; GB300 NVL72 delivers 143% more tok/s/chip. (Numbers reflect the default agentic-traces · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
Interpolated from real benchmark data. Edit target interactivity values below to compare at different operating points.
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | GB300 NVL7258,115.9MI325X5,717.9 | GB300 NVL7224,563.0MI325X5,717.9 | GB300 NVL7213,913.8MI325X5,717.9 |
| Cost ($/M tok) | GB300 NVL72$0.011MI325X$0.053 | GB300 NVL72$0.026MI325X$0.053 | GB300 NVL72$0.046MI325X$0.053 |
| tok/s/MW | GB300 NVL7227,413,165MI325X3,383,379 | GB300 NVL7211,586,311MI325X3,383,379 | GB300 NVL726,563,126MI325X3,383,379 |
| Concurrency | GB300 NVL72~370MI325X~12 | GB300 NVL72~112MI325X~12 | GB300 NVL72~53MI325X~12 |