Computational Storage for an Energy-Efficient Deep Neural Network Training System

Shiju Li, Kevin Tang, Jin Lim, Chul-Ho Lee, Jongryool Kim

Abstract: Near-storage data processing and computational storage have recently received considerable attention from the industry as energy- and cost-efficient ways to improve system performance. This paper introduces a computational-storage solution to enhance the performance and energy efficiency of an AI training system, especially for training a deep learning model with large datasets or high-dimensional data. Our system leverages dimensionality reduction effectively by offloading its operations to computational storage in a systematic manner. Our experiment results show that it can reduce the training time of a deep learning model by over 40.3%, while lowering energy consumption by 38.2%.