A Scalable and Adaptive Convolutional Neural Network Accelerator

Abstract

Machine learning has become ubiquitous and penetrated every field of technology, medicine, and finance. Convolutional Neural Network (CNN) is one of the most commonly used class of machine learning algorithms that is being used in video and image processing, big data processing, natural language processing, robotics, and a variety of pattern matching and recognition tasks. Depending on the end application, CNNs are being employed on different scales ranging from tiny motion sensors and smartphones to automobiles and server farms. Although existing CNN accelerators are adaptive for different types of CNN models, they are generally suited for a particular scale of operation. In this paper, we describe a scalable and adaptive CNN accelerator. The same hardware-cum-software stack can be configured by a system-level parameter to be synthesized for different scales of operation. This makes the accelerator highly portable across systems of different scales. Furthermore, one single synthesized hardware can run inference for multiple CNN models because of the flexible software stack and hardware control unit making the system highly adaptive. We demonstrate the working of the system at different scales by implementing it on the Xilinx Virtex 7 FPGA and by running multiple CNN models at each scale.