Towards Artificial Intelligence Hardware With 3D Integrated Ferroelectric Transistors.

Area of Science:

• Materials Science and Engineering
• Electrical Engineering
• Computer Science (AI Hardware)

Background:

• Modern AI workloads demand significant energy and bandwidth, stressing traditional von Neumann architectures due to data movement bottlenecks.
• Compute-in-memory and neuromorphic systems offer potential solutions, but reliable 3D integration of analog synaptic devices is a key challenge.

Purpose of the Study:

• To develop a monolithic 3D (M3D) integration platform for compact, energy-efficient neuromorphic hardware.
• To demonstrate the feasibility of vertically stacking Indium Gallium Zinc Oxide (IGZO) access transistors and Hafnium Zirconium Oxide (HZO)-based ferroelectric transistors (FeFETs).

Main Methods:

• Fabrication of two-tier and four-tier IGZO/FeFET architectures using a monolithic 3D integration platform.
• Characterization of device structural integrity, elemental profiles, and HZO ferroelectricity across all tiers.
• Evaluation of device switching reproducibility, retention (>10 years), endurance (>10^11 cycles), and multilevel conductance states.

Main Results:

• Achieved excellent structural integrity and preserved ferroelectricity in the M3D-integrated IGZO/FeFETs.
• Demonstrated stable synaptic device characteristics, including reproducible switching and long retention/endurance.
• Attained high accuracy (95.0%-95.5%) for CIFAR-10 inference using mapped FeFETs in a CNN, approaching the software baseline (96.1%).
• Successfully performed analog-domain convolution by encoding kernel weights into FeFET conductance states for edge-aware image processing.

Conclusions:

• The M3D integration platform provides a scalable and reliable method for fabricating advanced neuromorphic hardware.
• The demonstrated IGZO/FeFET devices are suitable for synaptic computing and next-generation compute-in-memory applications.
• This technology paves the way for energy-efficient AI hardware, particularly for neuromorphic vision systems.