SPOTS: An Accelerator for Sparse CNNs Leveraging General Matrix-Matrix Multiplication

This paper proposes a new hardware accelerator for sparse convolutional neural networks (CNNs) by building a hardware unit to perform the Image to Column (IM2COL) transformation of the input feature map coupled with a systolic array-based general matrix-matrix multiplication (GEMM) unit. Our design carefully overlaps the IM2COL transformation with the GEMM computation to maximize parallelism. We propose a novel design for the IM2COL unit that uses a set of distributed local memories connected by a ring network, which improves energy efficiency and latency by streaming the input feature map only once. We propose a tall systolic array for the GEMM unit while also providing the ability to organize it as multiple small GEMM units, which enables our design to handle a wide range of CNNs and their parameters. Further, our design improves performance by effectively mapping the sparse data to the hardware units by utilizing sparsity in both input feature maps and weights. Our prototype, SPOTS, is on average 1.74X faster than Eyeriss. It is also 78X, and 12X more energy-efficient when compared to CPU and GPU implementations, respectively.