Digital ReRAM-based compute-in-memory design

• Main Author: Xu, Jiawei
• Other Authors: Kim Tae Hyoung
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; Compute-in-memory; ReRAM
• Online Access: https://hdl.handle.net/10356/175060
• Institution: Nanyang Technological University

As Artificial Intelligence (AI) continues to advance, the continuous pursuit of computational power makes in-store computing a hot topic in present research. This work aims to address the limitations of traditional compute-in-memory (CIM) architectures, proposing an innovative all-digital ReRAM-based CIM architecture tailored for edge AI applications. Firstly, the recent paper about ReRAM CIM is reviewed and the basic properties of ReRAM and its working principle are introduced. Following, a novel ReRAM structure 2T2R is introduced. The proposed 2T2R cell eliminates the problems associated with traditional analog designs and has better stability. The CIM architecture can support reconfigurable computing with weight precision ranging from 1 to 8 bits. Simulation results have demonstrated the architecture’s capability to perform 9-bit multiply-accumulate (MAC) operations as well as read/write operations without any loss of accuracy, showcasing its precision and reliability. In terms of computational efficiency, the architecture achieves an exceptional energy efficiency of 15.6 TOPS/W in computational mode.