Quantitative Analysis of Various 2D CNN Structures based on Dataflow

Abstract

Convolutional Neural Networks (CNNs) are used in a wide range of fields due to their excellent accuracy. Previous researchers have proposed convolution architectures for the typical convolutional layer in CNN whose input is larger than the kernel size. However, neural networks are continuously evolving and developing, and there is a deep convolutional layer unlike classic CNNs demands a larger kernel size than the input. In this paper, we conduct a quantitative analysis based on dataflow for various CNN structures. A total of eight 2D CNN structures are described and compared in terms of processing time, total area, and energy efficiency, based on different dataflow graphs. As a result, the comparison provides advantages and disadvantages of the different CNN structures and aids in determining the optimal hardware structure solution for various neural network types.