Enabling Low-Power Charge-Domain Nonvolatile Computing-in-Memory (CIM) With Ferroelectric Memcapacitor

The explosive growth in data-centric computing driven by artificial intelligence (AI) and big data has surpassed the capabilities of the traditional von Neumann architecture. The architectural separation of computing and memory units results in substantial energy consumption, latency, and additional...

Full description

Saved in:
Bibliographic Details
Published in:IEEE transactions on electron devices 2024-04, Vol.71 (4), p.1-7
Main Authors: Wang, Xuepei, Cui, BoYao, Jing, Lingling, Wang, Xiaolin, Wu, Maokun, Wen, Yicheng, Wu, Yishan, Liu, Jinhao, Zhang, Feilong, Lin, Zidong, Sun, Yanan, Ren, Pengpeng, Ye, Sheng, Wang, Runsheng, Ji, Zhigang, Huang, Ru
Format: Article
Language:English
Subjects:
Artificial intelligence
Big Data
Binary neural network (BNN)
Computer memory
computing-in-memory (CIM)
Edge computing
Energy consumption
Ferroelectric materials
Ferroelectricity
memcapacitor
multiplication and accumulation (MAC)
Network latency
Neural networks
P-i-n junctions
Power consumption
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:The explosive growth in data-centric computing driven by artificial intelligence (AI) and big data has surpassed the capabilities of the traditional von Neumann architecture. The architectural separation of computing and memory units results in substantial energy consumption, latency, and additional hardware expenses. Therefore, a novel computing paradigm with low power consumption and high parallelism is urgently needed. One promising solution to address the memory wall challenge is capacitor-based computing-in-memory (CIM). In this article, we propose a hafnium-based ferroelectric memcapacitor (FE-memcap), which achieves a 103 ON/OFF ratio by combining ferroelectric polarization switching and p-i-n junction charge shielding. Furthermore, gate work-function engineering enables high ON/OFF ratios with small driven voltages, typically around 0.05 V. When integrated into the XNOR-binary neural network (BNN), the FE-memcap demonstrates the equivalent computational accuracy as traditional memristor-based counterparts. Importantly, the FE-memcap exhibits negligible static power consumption and significantly reduces dynamic power consumption by over 104 times. This attribute renders it highly suitable for event-driven edge computing applications and augments its potential for future AI applications.
ISSN:0018-9383
1557-9646
DOI:10.1109/TED.2024.3367965