A Design of Sparse Neural Network Accelerator for Memory Access Optimization

At present, convolutional neural networks have developed rapidly. How to improve computational efficiency and optimize memory access has become an important research direction in neural network accelerators. In this paper, we proposed a new sparse convolutional neural network accelerator. The new accelerator improves the efficiency of memory access. What's more, it can also effectively utilize the neuron sparsity in the sparse convolutional neural network, thus improving the computation speed of the accelerator. Compared with the existing sparse convolutional neural network accelerators, the experiments show that the memory access time of the accelerator is reduced by 60%, the average speedup of overall neural networks and the convolutional layer is 1.81x and 2.38x respectively.