DSSA: Dual-Side Sparse Systolic Array Architecture for Accelerating Convolutional Neural Network Training.

Ever-growing CNN size incurs a significant amount of redundancy in model parameters, which in turn, puts considerable burden on hardware. Unstructured pruning is widely used to reduce model sparsity. While, the irregularity introduced by unstructured pruning makes it difficult to accelerate sparse CNNs on systolic array. To address this issue, a variety of accelerators have been proposed. SIGMA, the state-of-the-art sparse GEMM accelerator, achieves significant speedup over systolic array. However, SIGMA suffers from two disadvantages: 1) it only supports one-side sparsity, leaving potential for further performance gains; 2) SIGMA improves utilization of large-sized systolic arrays at the cost of extra overhead. In this paper, we propose DSSA, a dual-side sparse systolic array, to accelerate CNN training. DSSA bases its designs on a small-sized systolic array, which naturally achieves higher cell utilization without additional overhead. To facilitate dual-side sparsity processing, DSSA utilizes a cross-cycle reduction module to accumulate partial sum that belongs to the same column but being processed in different cycles. A comprehensive design space exploration is performed to seek the local optimal configurations for DSSA. We implement the logic design of DSSA using Verilog in RTL and evaluate its performance using a C++-based cycle-accurate performance simulator we built. Experimental results show that DSSA delivers, on average, a speedup of 2.13x and 13.81x over SIGMA and a basic systolic array with the same number of cells. Compared to SIGMA, DSSA incurs 16.59% area overhead and 25.49% power overhead when sparse filter is excluded, as SIGMA did.