Inference Energy Consumption on Qwen3-MoE 30B (inference)
3.8Energy per Run (J)NPUMoE
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| NPUMoEMode=Ours, Hardware=M2 Ultra2026.04 | 3.8 | 1.23 | 1.2 | 1.37 | 1 | |
| anemllMode=anemll, Hardware=M2 Ultra2026.04 | 12.13 | 8.87 | 0.26 | 2.99 | 3.19 | |
| coreml-onlyMode=coreml-only, Hardware=M2 Ultra2026.04 | 12.67 | 9.66 | 0 | 3.01 | 3.33 |