Energy-efficient FFT-based CNN accelerator through approximate computing

Abstract

We suggested the FFT-based CNN accelerator to reduce the number of multiplication of the spatial-domain convolution. However, the precision of multiplication increases in spectral-domain convolution as much as the processing-gain of the FFT. Thus, the property of the image in the spectral-domain is used to reduce precision. The energy consumed by convolution operation is saved with the minimal penalty of classification accuracy. Also, the wireline communication bottleneck the energy of the acceleration system. We investigated the reduced bit-error-rate requirement for wireline interface with significant energy-saving and negligible accuracy-loss.