High Area Efficiency (6 TOPS/mm2) Multimodal Neuromorphic Computing System Implemented by 3D Multifunctional RRAM Array

作者：Wenxuan Sun, Yi Li, Woyu Zhang, Xu Zheng, Danian Dong, Jie Yu, Jinru Lai, Shaoyang Fan, Hongzhou Wang, Xiaoxin Xu, Ming Liu · 年份：2023 · DOI：10.1109/iedm45741.2023.10413870 · 被引用次数：3 · 研究领域：Advanced Memory and Neural Computing、Neural Networks and Reservoir Computing、Ferroelectric and Negative Capacitance Devices

For the first time, we demonstrated a multifunction three-dimensional (3D) vertical random-access memory (RRAM) array (MF-ЗDRRAM) where different layers exhibit nonvolatile properties and volatile characteristics respectively, to implement multimodal neuromorphic computing. The RRAM cells in the 1stlayer (WL: TiN) and the 2ndlayer (WL: Ru) have different dynamic characteristics, which are used to construct multi-scale reservoirs (M-RC). The RRAM in 3rdlayer (WL: W) exhibits analog switching behavior, applying for convolutional neural network (CNN) and full connection (FC) layer. A multimodal neuromorphic computing system with the network of M-RC+CNN is implemented by the MF-3DRRAM. The multifunction of the fabricated MF-3DRRAM chip is validated through the multimodal video recognition task, exhibiting high accuracy (98%), high area efficiency (6 TOPS/mm2) and low energy consumption (1.4pJ/operation). This proposed MF-3DRRAM is of great significance for miniaturized, low-power hardware implementations for edge computing.