A Throughput-Optimized Channel-Oriented Processing Element Array for Convolutional Neural Networks

摘要

Over the past decade, significant developments have taken place in the field of deep learning. State-of-the-art convolutional neural networks (CNNs), a branch of deep learning, have been increasingly applied in various fields such as image classification, speech recognition, and natural language processing. Due to the high computational complexity of CNNs, lots of works have proposed their CNN accelerators to address this issue. Besides, a processing element (PE) array has been recently further focused and discussed since it is responsible for the entire computations as the core of CNN accelerators. Therefore, the specialized design of a PE array becomes one of the main researches on CNN accelerators for energy efficiency and high throughput. In this brief, a throughput-optimized PE array for CNNs based on the channel-oriented data pattern is proposed. The proposed PE array features fully PE interconnection which achieves scalability. Besides, any sized convolution can be processed in the PE array while maximizing the utilization of PEs by exploiting the channel-oriented data pattern. Compared to previous works, this brief achieves 1.22x and 1.25x improvement in the throughput density on AlexNet and VGG-16 respectively.