Estimating energy consumption of neural networks with joint Structure–Device encoding

Abstract

The surge in IoT devices has led to an increase in energy consumption, necessitating the optimization of neural networks deployed on these energy-constrained devices to reduce power usage. Although various techniques, such as pruning and quantization, can reduce the size and computational requirements of neural networks, the resulting energy savings still need to be verified through resource-intensive inference processes, which require cumbersome adjustments to measurement devices and neural network deployment. To address these challenges, we propose SDEnergy, a novel approach that combines Structure–Device encoding to quickly and accurately predict the Energy consumption of neural networks across various devices. SDEnergy utilizes graph neural networks to extract structural features of neural networks and employs fully connected networks to extract device features, using their fusion for energy consumption prediction. Experimental validation demonstrates that SDEnergy has established state-of-the-art results on our dataset based on NAS-Bench-101 and various IoT device parameter scenarios, with a mean absolute percentage error of 5.35%.