A Digital SRAM Computing-in-Memory Design Utilizing Activation Unstructured Sparsity for High-Efficient DNN Inference

Abstract

The Computing-in-Memory (CIM) architecture has emerged as a promising approach for designing energy-efficient DNN processors. While previous CIM designs have explored the use of DNN weight sparsity, these approaches often involve pruning the weight matrix in a specific manner. This process may increase the new complexity of the calculation and negatively impact DNN accuracy. However, there are barely any digital CIM circuits that leverage the sparsity in activation which is naturally sparse in many scenarios due to the ReLU activation functions. In order to fully utilize activation unstructured sparsity, we proposed a digital SRAM CIM. This circuit is designed using the booth encoding scheme and adopts the circuit structure of an accumulator-based multiply-accumulate (MAC) calculation. It utilizes SRAM bit-line (BL) computing to obtain matrix sparse information and employs an allocator to allocate data calculation for SRAM-CIM. The proposed design is implemented and evaluated at 40 nm CMOS process. Our evaluation results show that the proposed circuit can achieve a clock frequency of 1 GHz at 1.1 V, with a peak performance of 819.2 GOPS, and in the case of 50%-90% sparsity, SRAM-CIM achieves $1.12 \times 3.32 \times$ speedup, and energy savings of 48.2% to 90.57% over dense mode. When performing an 8-bit matrix multiplication with 90% sparsity, the energy efficiency is 10.57 TOPS/W.