(Invited) Heterogeneous Integration and Co-design of Memristor and MOSFET for AI Application. (8th September 2020)

• Record Type: Journal Article
• Authors: Xi, Yue; Gao, Bin
• Is Part Of: ECS transactions. Volume 98:Number 5(2020)
• Page Start: 3
• Page End: 8
• Publication Date: 2020-09-08
• Subjects: Electrochemistry -- Periodicals; Electrochemistry; Periodicals; Electronic journals; Electronic journal; 541.37
• DOI: 10.1149/09805.0003ecst
• Languages: English
• ISSNs: 1938-5862

Abstract

Abstract : Recent advances in AI bring great challenges on the computing hardware. Memristor based neuromorphic computing can improve the computing density and energy efficiency significantly. However, how to integrate memristive materials and design mixed circuit with advanced CMOS platform is a key challenge. In this presentation, I will talk about our recent progress on heterogeneous integration of memristor and MOSFET. Million-level one-transistor-one-resistor (1T1R) cells can be integrated on a chip. Then I will discuss the co-design technology for memristor/MOSFET mixed circuits. A general purpose neuromorphic computing chip was designed and fabricated to process deep neural networks. I will also introduce our new hybrid training method for eliminating the accuracy loss caused by device variations.