Optimizing Power Efficiency in Embedded AI A CNN Performance Study on Custom SoC Architectures

Efficient power management is critical for embedded AI systems operating in constrained environments. This study investigates the power-performance trade-offs of deep learning models on a specialized heterogeneous processing platform. By leveraging dynamic workload distribution and frequency scaling techniques, I analyze the energy efficiency of multiple CNN architectures under varying computational loads. The findings provide key insights into optimizing inference performance while minimizing power consumption, offering practical guidance for deploying AI models on embedded hardware.