E4: Energy-Efficient DNN Inference for Edge Video Analytics via Early Exiting and DVFS

• DOI: 10.1609/aaai.v39i1.32104
• Publication Date: 2025-04-11T09:29:04Z
• Authors: Ziyang Zhang; Yang Zhao; Ming-Ching Chang; Changyao Lin; Jie Liu
• Keywords: FOS: Computer and information sciences; Artificial Intelligence (cs.AI); Computer Science - Artificial Intelligence; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition

Deep neural network (DNN) models are increasingly popular in edge video analytic applications. However, the computeintensive nature of DNN pose challenges for energyefficient inference on resource-constrained devices. Most existing solutions focus optimizing latency and accuracy, often overlooking energy efficiency. They also fail to account varying complexity frames, leading sub-optimal performance analytics. In this paper, we propose an EnergyEfficient Early-Exit (E4) framework that enhances efficiency analytics by integrating a novel early-exit mechanism with dynamic voltage frequency scaling (DVFS) governors. It employs attentionbased cascade module analyze frame diversity automatically determine optimal exit points. Additionally, E4 features just-in-time (JIT) profiler uses coordinate descent search co-optimize CPU GPU clock frequencies each layer before Extensive evaluations demonstrate outperforms current state-of-the-art methods, achieving up 2.8× speedup 26% average saving while maintaining high accuracy.