Efficient Processing of Deep Neural Networks: A Tutorial and Survey

Deep neural networks are widely used across artificial intelligence applications such as computer vision, speech recognition, and robotics, but their state-of-the-art accuracy comes at the cost of high computational complexity. This article provides a comprehensive tutorial and survey on techniques that enable efficient processing of DNNs, aiming to improve energy efficiency and throughput without sacrificing application accuracy or increasing hardware cost. It gives an overview of DNNs, discusses various hardware platforms and architectures that support them, and highlights key trends in reducing computation cost either solely through hardware design changes or through joint hardware design and DNN algorithm codesign.

Beyond surveying methods, the article summarizes development resources that help researchers and practitioners get started, and highlights important benchmarking metrics and design considerations for evaluating the rapidly growing number of proposed DNN hardware designs. Readers come away able to understand key DNN design considerations, evaluate different hardware implementations using benchmarks and comparison metrics, weigh tradeoffs among hardware architectures and platforms, and assess the utility of various efficiency techniques, which mattered for guiding practical, energy-efficient DNN deployment.