Depends on how big the NPU is and how much power/memory the inference model needs.