ReShare: A Resource-Efficient Weight Pattern Sharing Scheme for Memristive DNN Accelerators

Memristor crossbar-based computing-in-memory (CIM) has garnered significant attention for accelerating deep neural networks (DNNs). Although practical operation unit (OU)-based designs incur abundant repetitive computations, recent studies propose sharing mechanisms, where redundant computations are replaced with data transfers. However, existing sharing approaches carry a heavy burden in terms of storage and energy in practical implementation. To alleviate this problem, this paper proposes a resource-efficient scheme for weight pattern sharing (ReShare). A pattern reorder algorithm is introduced to facilitate the asynchronous running of computations and sharing, which circumvents the bottleneck posed by the output buffer in a conventional sharing scheme. In addition, ReShare enables block write operations to reduce write costs. Our evaluation across four DNN models demonstrates that the proposed ReShare has a lighter overhead in energy and storage while simultaneously enhancing performance. Specifically, ReShare can achieve 1.66 × speedup, 30.7% energy saving, and 49.5% storage reduction compared to the state-of-the-art weight sharing method.