Advances and Challenges in ECRAM-Based Resistive Crosspoint Arrays for AI Computation

Artificial Intelligence (AI) technology, which heavily relies on deep neural networks, is having a significant impact on both business and society, transforming our daily lives. To support the ever-increasing computation demands for AI applications, there has been a growing interest in novel analog computing architectures enabling massively-parallel neural network computation by adopting cross-point arrays of emerging non-volatile memory devices. Both device-level and system-level research efforts are actively pursued to achieve the promised power-efficiency and substantial performance improvements. Among various cross-point device candidates, electrochemical random-access memory (ECRAM) has recently attracted researchers owing to its 3-terminal nature and superior programmability [1-3]. In this talk, I will first introduce our recent device specification study when different training algorithms, optimized for practical resistive memory devices, are used to train neural networks in resistive cross-point arrays [4]. This quantitative study is expected to provide a guide for material and device research. I will continue the discussion on the recent advances and challenges in implementing large ECRAM-based cross-point arrays for analog AI computations with the consideration of the non-ideal device characteristics such as limited retention time and selector-less array-level operations.<br/><br/>[1] S. Kim et al., 2019 IEEE International Electron Devices Meeting, 2019. [2] H. Kwak et al., Semiconductor Science and Technology, 2021. [3] J. Tang et al., 2018 IEEE International Electron Devices Meeting, 2018. [4] C. Lee et al., Frontiers in Neuroscience, 2022.