神经形态工程学
电阻随机存取存储器
计算机科学
尖峰神经网络
人工神经网络
计算
计算机体系结构
计算机工程
电子工程
计算机硬件
人工智能
作者
Chunmeng Dou,Xiaoxin Xu,Xumeng Zhang,Linfang Wang,Wang Ye,Junjie An,Jianguo Yang,Qing Luo,Tuo Shi,Jing Liu,Dashan Shang,Feng Zhang,Qi Liu,Ming Liu
标识
DOI:10.1109/iedm19574.2021.9720546
摘要
In this work, we discussed the critical challenges, key enabling techniques, and future trends on developing RRAM-based brain-inspired computation, including computing-in-memory (CIM) and neuromorphic computing (NC), from device, circuit to system. To suppress the device non-idealities in the synaptic array, we proposed using optimized bit-cell design and computing approach to reduce the errors and power of the analogue multiply-and-accumulate (MAC). To lower the neuron power, we proposed the sparsity-aware analog-to-digital converter (ADC) for artificial neural networks (ANNs) and highlighted the energy- and area-efficient bio-plausible neurons based on NbO x devices for spiking neural networks (SNNs). On this basis, we introduce several RRAM CIM designs, followed by a discussion on the remaining challenges and future trends.
科研通智能强力驱动
Strongly Powered by AbleSci AI