An OpenCL-Based FPGA Accelerator with the Winograd’s Minimal Filtering Algorithm for Convolution Neuron Networks

Convolutional neural network has been extensively used in image classification, video processing and semantic recognition. Generally speaking, convolutional neural network is accelerated by high power GPU, but this work uses the FPGA to speed up the convolutional neural network because of its configurability and low power consumption advantage over GPU. Especially, OpenCL-based high-level synthesis tools can provide fast verification and implementation flows. This work mainly uses the Winograd’s minimal filtering algorithm to accelerate VGG-16 network on Intel Arria 10 GX FPGA board and has a good effect on 3x3 convolution kernel. The peak performance of 227 GOP/S has been achieved with 544 DSP.