An Accelerator for Sparse Convolutional Neural Networks Leveraging Systolic General Matrix-matrix Multiplication

This article proposes a novel hardware accelerator for the inference task with sparse convolutional neural networks (CNNs) by building unit to perform Image Column ( Im2Col ) transformation of input feature map coupled systolic-array-based general matrix-matrix multiplication (GEMM) unit. Our design carefully overlaps GEMM computation maximize parallelism. We propose that uses set distributed local memories connected ring network, which improves energy efficiency and latency streaming only once. The in can be dynamically configured as multiple units square-shaped systolic arrays or single tall array. dynamic reconfigurability enables effective pipelining operations attains high processing element utilization wide range CNNs. Further, our is sparsity aware, improving performance effectively mapping maps weights elements, skipping ineffectual unnecessary data movements involving zeros. prototype, SPOTS, on average 2.16 \( \times \) , 1.74 1.63 faster than Gemmini, Eyeriss, Sparse-PE, are prior accelerators dense CNNs, respectively. SPOTS also 78 12 more energy-efficient when compared CPU GPU implementations,