Energy efficient techniques using FFT for deep convolutional neural networks

Abstract

Deep convolutional neural networks (CNNs) has been developed for a wide range of applications such as image recognition, nature language processing, etc. However, the deployment of deep CNNs in home and mobile devices remains challenging due to substantial requirements for computing resources and energy needed for the computation of high-dimensional convolutions. In this paper, we propose a novel approach designed to minimize energy consumption in the computation of convolutions in deep CNNs. The proposed solution includes (i) an optimal selection method for Fast Fourier Transform (FFT) configuration associated with splitting input feature maps, (ii) a reconfigurable hardware architecture for computing high-dimensional convolutions based on 2D-FFT, and (iii) an optimal pipeline data movement scheduling. The FFT size selecting method enables us to determine the optimal length of the split input for the lowest energy consumption. The hardware architecture contains a processing engine (PE) array, whose PEs are connected to form parallel flexible-length Radix-2 single-delay feedback lines, enabling the computation of variable-size 2D-FFT. The pipeline data movement scheduling optimizes the transition between row-wise FFT and column-wise FFT in a 2D-FFT process and minimizes the required data access for the element-wise accumulation across input channels. Using simulations, we demonstrated that the proposed framework improves the energy consumption by 89.7% in the inference case.