Random and Systematic Variation in Nanoscale Hf0.5Zr0.5O2 Ferroelectric FinFETs: Physical Origin and Neuromorphic Circuit Implications

Sourav De,Tarek Ali,Franz Müller,Md Aftab Baig,Thomas Kämpfe,Hoang-Hiep Le,Po-Jung Sung,Bo-Han Qiu,Yao-Jen Lee,Darsen D Lu,Maximilian Lederer,Chun-Jung Su

doi:10.3389/fnano.2021.826232

Abstract

This work presents 2-bits/cell operation in deeply scaled ferroelectric finFETs (Fe-finFET) with a 1 µs write pulse of maximum ±5 V amplitude and WRITE endurance above 109 cycles. Fe-finFET devices with single and multiple fins have been fabricated on an SOI wafer using a gate first process, with gate lengths down to 70 nm and fin width 20 nm. Extrapolated retention above 10 years also ensures stable inference operation for 10 years without any need for re-training. Statistical modeling of device-to-device and cycle-to-cycle variation is performed based on measured data and applied to neural network simulations using the CIMulator software platform. Stochastic device-to-device variation is mainly compensated during online training and has virtually no impact on training accuracy. On the other hand, stochastic cycle-to-cycle threshold voltage variation up to 400 mV can be tolerated for MNIST handwritten digits recognition. A substantial inference accuracy drop with systematic retention degradation was observed in analog neural networks. However, quaternary neural networks (QNNs) and binary neural networks (BNNs) with Fe-finFETs as synaptic devices demonstrated excellent immunity toward the cumulative impact of stochastic and systematic variations.

Highlights

The advent of convolutional neural networks (Lecun et al, 2015) has made machine learning or neural network–based computation an inevitable choice for solving many complex tasks in recent times
The gate stack consisting of hafnium zirconium oxide (HZO) and TiN was formed by the atomic layer deposition (ALD) and physical vapor deposition (PVD) method on top of the interfacial layer
The online training simulation of Fe-finFET–based synaptic arrays shows that the continuous weight adjustment can avoid the accuracy drop due to device-to-device variations during the training process

Summary

Introduction

The advent of convolutional neural networks (Lecun et al, 2015) has made machine learning or neural network–based computation an inevitable choice for solving many complex tasks in recent times. The real-time processing of the enormous amount of data generated from the internet search engines, social networks, and edge devices in health care systems requires massive computing power in conventional von-Neumann computing architecture. The memory-bandwidth bottleneck in vonNeumann computing and the torrent of data generated on the internet every second have reinvigorated the research in brain-inspired computing. Artificial neural networks implemented in such a traditional von-Neumann computing system endure severe bottlenecks during the data transfer between segregated memory and processing units. There are two pathways for implementing the artificial neural network (ANN)

Methods

Results

Conclusion