A Reconfigurable Process Engine for Flexible Convolutional Neural Network Acceleration

Abstract

Convolutional neural network (CNN) is the most powerful artificial intelligence algorithm widely used in computer vision due to its state-of-the-art performance. There are many accelerators proposed for CNN to handle its huge computation and communication cost. In this paper we proposed a reconfigurable process engine which can support different data flows, bit-widths, and parallelism strategies for CNNs. The process engine was implemented on Xilinx ZC706 FPGA board, with high flexibility to support all popular CNNs, and better energy efficiency compared to other state-of-the-art designs.