IOFusion: A Sparse Accelerator based on Hybrid Inner-Outer Product Dataflow for DNN Training

Output sparsity is a widely observed characteristic during the training of Deep Neural Networks (DNNs), yet it has not been fully exploited. This paper presents IOFusion, a sparse accelerator based on a hybrid inner-outer product (IP-OP) dataflow. By partitioning the output matrix with known sparsity into dense and sparse regions, IOFusion applies a compute-intensive outer product dataflow for the dense regions and a memory-intensive inner product dataflow for the sparse regions. This hybrid dataflow minimizes redundant computations and unnecessary memory access during the convolution process. Additionally, we introduce a Z-shaped storage format and a fusion adder tree to improve the performance of the accelerator, along with a data reuse mechanism specifically designed for the inner product phase. Experiments on three classic DNN networks demonstrate that IOFusion achieves an average speedup of 3.25× and 2.91× compared to traditional outer/inner product dataflow accelerators, while reducing on-chip memory access by at least 48%.