Paper Title
EnergyBench++: A Comprehensive Framework for Dynamic Energy–Accuracy Trade-Offs in Edge Deep Learning

Abstract
Energy efficiency is a critical concern for deploying deep neural networks on resource-constrained edge devices. However, existing benchmarks (e.g., MLPerf) focus on inference speed and throughput, failing to capture dynamic trade-offs between energy and accuracy. We introduce EnergyBench++, a unified evaluation framework that systematically profiles energy, latency, and accuracy across varied operating modes. EnergyBench++ features a new Dynamic Energy-Accuracy Efficiency Ratio (D-EAER) metric (accuracy per Joule), rigorous statistical evaluation (with confidence intervals), per-layer energy breakdowns, and built-in support for early-exit networks. Our results on CIFAR-10 and TinyImageNet, as well as audio classification tasks, show, for instance, that SqueezeNet achieves the best energy-efficiency (D-EAER ≈ 54.1) under batch=1 on the RTX 3090, while on the Jetson Nano, it maintains high efficiency with significantly lower energy consumption. Additionally, using an early-exit threshold of τ = 0.9 yields a sweet spot in the accuracy-energy trade-off across devices. This framework enables principled comparison of models and strategies under realistic edge constraints. To facilitate reproducibility, we will release the code upon acceptance. Keywords - Energy Efficiency, Edge Deep Learning, Power-Aware Inference, Early-Exit Networks, Resource-Constrained Devices