Evaluating the Energy Efficiency of Deep Convolutional Neural Networks on CPUs and GPUs

Abstract

In recent years convolutional neural networks (CNNs) have been successfully applied to various applications that are appropriate for deep learning, from image and video processing to speech recognition. The advancements in both hardware (e.g. more powerful GPUs) and software (e.g. deep learning models, open-source frameworks and supporting libraries) have significantly improved the accuracy and training time of CNNs. However, the high speed and accuracy are at the cost of energy consumption, which has been largely ignored in previous CNN design. With the size of data sets grows exponentially, the energy demand for training such data sets increases rapidly. It is highly desirable to design deep learning frameworks and algorithms that are both accurate and energy efficient. In this paper, we conduct a comprehensive study on the power behavior and energy efficiency of numerous well-known CNNs and training frameworks on CPUs and GPUs, and we provide a detailed workload characterization to facilitate the design of energy efficient deep learning solutions.