An Inference Accelerator Design for Sparse Convolution Neural Network

Convolution neural networks (CNN) is a widely used deep learning method at present. The weight and activation in the inference process of CNN have high sparsity. Making full use of the sparsity of CNN to accelerate is a research hotspot. This paper analyzes various classical accelerator structures, sparsity and network compression methods and sparsity accelerator structures. The weight pre compression storage method is used to explore the sparsity of weight in the inference process, and the aggregation-diffusion method is used to explore the sparsity of activation. A new sparse CNN inference accelerator structure is proposed, the complete design and implementation are carried out, and the analysis model of performance and overhead is established. The experimental results show that in terms of sparsity utilization, the proposed method can make use of 100% weight sparsity and more than 60% activation sparsity. In terms of circuit overhead, the proposed structure has strong scalability.