This academic paper discusses the potential of N:M fine-grained structured sparsity in reducing the computational cost of Deep Neural Networks (DNNs) while maintaining accuracy. It proposes a bidirectional weight pruning method (BDWP), a sparse accelerator for DNN training (SAT), and multiple optimization methods to boost computational efficiency. Experiments on various models and datasets show an average speedup of 1.75 with a negligible accuracy loss of 0.56%, significantly improving training throughput and energy efficiency over prior FPGA-based accelerators.
Publication date: 25 Sep 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2309.13015
