Convolutional Neural Network Accelerator with Reconfigurable Dataflow

Abstract

Convolutional-Neural-Network (CNN) is used in broad applications. There are dataflows for convolutional layers in CNN such as row-stationary and weight-stationary. However, these dataflows have strengths and weaknesses. This paper analyzed two representative dataflows and introduce the dataflow-reconfigurable CNN accelerator that takes advantage of both dataflows.