An efficient ReRAM-based inference accelerator for convolutional neural networks via activation reuse

Abstract

In this paper, a novel resistive random access memory (ReRAM) based accelerator is proposed for convolution neural network (CNN) inference accelerations. In ReRAM-based CNN computation, weight parameters can be pre-programmed in ReRAM crossbar arrays, and activations are generated by processing the multiplication-and-accumulation (MAC) operations in the ReRAM crossbar arrays. However, prior works cannot reuse activations in computation, in which the activation dominates the data movements and raises significant energy cost. To deal with this dilemma, a tiling-based dataflow is proposed to enable activation reuse among adjacent ReRAM crossbar arrays to reduce the activation movements. We then develop a ReRAM-based CNN accelerator that can well suit the dataflow to reduce the cost of ReRAM access. Evaluation results show that the proposed design achieves 1.8× energy saving and 2.8× bandwidth saving compared with a state-of-the-art PipeLayer accelerator.