Synchronized Analog Capacitor Arrays for Parallel Convolutional Neural Network Training

Abstract

We report a novel Synchronized Analog Capacitor Arrays (SACA) to accelerate Convolution Neural Network (CNN) training. The synchronized cross-point capacitor arrays, functioning as replicated weights kernels, train on image patches in parallel. Parallel CNN training is challenging in analog arrays because of weight divergence in the replicated kernel. Capacitor arrays solve this problem by charge sharing between correlated capacitor in the kernel replicas to keep them synchronized. Using SACA, we show we can accelerate CNN training by >100x compared to other analog accelerators.