Efficient FPGA-Based Convolutional Neural Network Implementation for Edge Computing

In recent years, accelerating convolutional neural networks on Field Programmable Gate Array (FPGA) to improve the performance of the inference phase of artificial intelligent edge computing applications is a promising approach. This paper presents our proposed architecture for building a convolution neural network acceleration core on FPGA. The proposed FPGA-based core targets edge computing platforms where hardware resources and power efficiency are essential requirements. We use the MobileNet neural network model for image classification as a case study to evaluate our proposed system. We compare our work with a quad-core ARM Cortex processor at 1.2GHz and achieve speed-ups by up to 14.77x convolution operators. Although our system is worse than a 6-core Intel Core i7 processor, it is more energy-efficiency than the Intel processor. Our proposed system is the best fit for edge computing.