What question did this study set out to answer?

The aim is to develop a compact neuromorphic hardware solution to overcome data-movement challenges in AI workloads.

April 23, 2026Small0 citationsOpen Access

Towards Artificial Intelligence Hardware With 3D Integrated Ferroelectric Transistors

HSHyunho Seok GKGeonwook KimSungkyunkwan University SSSihoon SonSungkyunkwan University

Key Points

The aim is to develop a compact neuromorphic hardware solution to overcome data-movement challenges in AI workloads.
Developed a monolithic 3D integration platform stacking IGZO and HZO-based ferroelectric transistors.
Fabricated two-tier and four-tier architectures with assessments of structural integrity and ferroelectric properties.
Mapped device characteristics to a convolutional neural network for image inference tasks.
Achieved over 95% accuracy in CNN inference with integrated architectures, approaching the software baseline.
Demonstrated stable multilevel conductance states and reproducible switching over 10 years.
Showed endurance up to 10<sup>11</sup> cycles for the devices.

Abstract

The increasing energy and bandwidth demand of modern AI workloads highlight the need for hardware that mitigates the data-movement bottleneck of von Neumann architectures. Compute-in-memory and neuromorphic systems offer a compelling solution, yet reliable multi-tier 3D integration of analog synaptic devices remains challenging. Here, we report a monolithic 3D (M3D) integration platform that vertically stacks In-Ga-Zn-O (IGZO) access transistors and Hf0.5Zr0.5O (HZO)-based ferroelectric transistors to realize compact, energy-efficient neuromorphic hardware. Two-tier and four-tier IGZO/ferroelectric field effect transistors (FeFET) architectures were fabricated with excellent structural integrity, uniform elemental profiles, and preserved orthorhombic HZO ferroelectricity across all tiers. The devices exhibit reproducible switching, >10-year retention, endurance up to 101 1 cycles, and stable multilevel conductance states suitable for synaptic computing. Mapping device characteristics to a convolutional neural network (CNN) for Canadian Institute for Advanced Research (CIFAR-10) inference yields 95.0% (tier-2) and 95.5% (tier-4) accuracy, approaching the 96.1% software baseline. Analog-domain convolution was further demonstrated by encoding kernel weights into FeFET conductance states for edge-aware image processing. These results establish M3D-integrated FeFETs as a scalable and reliable platform for next-generation compute-in-memory and neuromorphic vision applications.

KI fragen

Bookmark

View Full Paper