Refresh Power Reduction of DRAMs in DNN Systems Using Hybrid Voting and ECC Method

Abstract

A deep neural network (DNN) system typically needs dynamic random access memories (DRAMs) for the data buffering. In this paper, an error-correction-code (ECC)-based technique is proposed to reduce the refresh power of DRAMs in the DNN system by extending the refresh period. By taking advantage of the characteristics of weight data of DNNs, a hybrid voting and ECC (VECC) method is used to protect the weight data from data retention fault caused by the refresh period extension. Analysis results show that the VECC method can achieve about 93% refresh power saving with about 0.5% accuracy loss and smaller than 0.5% check bit overhead on AlexNet, ResNet, and VGG19 convolutional neural network CNN models trained by ImageNet data set.