H-RIS: Hybrid Computing-in-Memory Architecture Exploring Repetitive Input Sharing

Yu-Chen Chen, Cheng-Yang Chang, An-Yeu Wu

Abstract: Computing-in-memory (CIM) has become a potential trend for accelerating convolutional neural networks (CNNs). Ongoing research, e.g., Repetitive Input Sharing (RIS), focuses on removing redundant matrix-vector multiplication (MVM) by exploiting computational reuse for higher energy efficiency. However, we argue that the RIS neglects the extra overheads of the computation reuse scheme. Moreover, analog CIM is inherently vulnerable to noise. Consequently, reusing the noisy MVM results may lead to severe accuracy degradation. To address the above issues, we first evaluate the extra buffer overheads resulting from the computation reuse scheme for storing repetitive MVM results in the buffer. Based on our evaluation, we find an optimal RIS reuse ratio that balances between buffer costs and the efficiency gain from computation reuse, leading to more energy reduction. In addition, we introduce the RIS-based Hybrid-CIM (H-RIS), which mixes up the analog CIM and digital near-memory-computing (NMC) at the pattern level to maintain accuracy. Based on the above techniques, when we set the RIS ratio to 25%, H-RIS increases 18% accuracy compared with the pure analog CIM and also reduces 97% energy compared with the pure digital NMC.