Monolithic-3D Inference Engine with IGZO Based Ferroelectric Thin Film
Transistor Synapses
Abstract
Instigated by the plethora of data generated by edge devices and IoT
devices, machine learning has become the de facto choice of everyone for
solving many tasks. Applications such as intelligent healthcare
monitoring systems, smart watches, or automatic cars require real-time
processing of the data or image, which is done by machine learning
algorithms with higher efficiency than humans. There are two possible
methods for artificial intelligence?1. non-von-Neumann hardware-based
implementation of neural networks 2. traditional computer science based
approach for neural networks or traditional von-Neumann
architecture?based implementation of neural networks The standard
von-Neumann performance of neural networks, where the memory and
computation parts are segregated, suffers from severe latency with
increasing number of edge devices. However, the multitude of edge
devices used in our daily life imposes strict restrictions on latency,
device area, and power consumption for hardware. Therefore, we need to
take the route beyond the CMOS-based mixed-signal implementation of
neural networks, where the memory bandwidth is not limited by the
quintessential von-Neumann archi?tecture. The primary motivation of
present-day research on the non-von-Neumann computing architecture is to
build dedicated hardware modules for implementing low-power,
fast-computing units without affecting the recent trend of scaling. This
chapter focuses on amorphous indium-gallium-zinc-oxide (α-IGZO) based
and their system-level applications. Back-end-of-line (BEoL) compatible
indium IGZO based multibit one-time programmable (OTP) ferroelectric
thin film transistors (FeTFT) with lifelong retention capability were
fabricated. The maximum temperature of the entire fabrication process
was limited to 350oC. The gate-stack engineering by varying the
thickness ratio of ferroelectric hafnium zirconium oxide (HZO) and IGZO
layer fomented excellent data-retention capability and one time
programming property. Further, we have evaluated the performance of
IGZO-based FeTFT as synaptic devices for an inference engine. The
system-level simulation revealed inference accuracy loss of only 1.5%
after ten years without re-training for Modified National Institute of
Standards and Technology (MNIST) hand-written digits in a multi-layer
perceptron (MLP) neural network with a baseline of 97%. The proposed
inference engine also showed superior energy efficiency and cell area of
95.33 TOPS/W (binary) and 8F2 , respectively.