[F-10] CMOS and Memory Applications of Low Dimensional Materials II

[F-10-02] Neural Network Hardware Accelerator using Memristive Crossbar Array based on Wafer-Scale 2D HfSe2

〇Sifan Li¹, Samarth Jain¹, Mei-Er Pam¹, Li Chen¹, Yu-Chieh Chien¹, Xuanyao Fong¹, Dongzhi Chi², Kah-Wee Ang^1,2 (1. National Univ. of Singapore (Singapore), 2. Inst. of Material Res. and Engineering (Singapore))

Presentation style: On-site (in-person)

https://doi.org/10.7567/SSDM.2022.F-10-02

For the first time, hardware acceleration of neural network calculations is demonstrated using large memristive crossbar arrays (CBAs) based on two-dimensional (2D) hafnium diselenide (HfSe2). Growth of wafer-scale polycrystalline HfSe2 thin film and metal-assisted van der Waals (vdW) transfer method are developed. The memristor exhibits a small switching voltage (0.6 V) with low switching energy (0.82 pJ) and achieves synaptic long-term potentiation/depression (LTP/D) with high offline classification accuracy (93.34%) in modeled artificial neural network (ANN). Furthermore, the CBA integrates custom-designed periphery circuits to implement full-hardware convolutional image processing, achieving large-scale image edge detection, soft and inverse functions with a tight current distribution of σ = 0.25 %. This work opens a potential route to the development of large-scale and energy-efficient neuromorphic computing systems.