A Review of Convolutional Neural Networks Hardware Accelerators for AIoT Edge Computing

Convolutional neural networks (hereafter CNN) have been widely used in the Artificial Intelligence & Internet of Things (AIoT) applications, due to its powerful feature extraction and classification capabilities. However, with the scale of CNN models increasing, the large number of network parameters and operations pose a great challenge on the design of low latency and low power CNN hardware accelerator for AIoT edge computing. This paper reviews the design techniques of state-of-the-art CNN hardware accelerators, including the design methods of dense and sparse CNN hardware accelerators, optimization methods of CNN hardware accelerators through software and hardware co-design, and the advanced design methods of computing in memory. These methods may help understand how to design CNN hardware accelerators with high performance, low power, and small area by improving the utilization of processing elements (PEs), reducing invalid data movement and increasing hardware flexibility. Besides, this paper also discusses the future direction of the design of CNN hardware accelerators.