Spin-Orbit-Torque-Driven Two-Terminal Giant Magnetoresistance Memristive Devices for In-Memory Computing

Area of Science:

• Materials Science
• Spintronics
• Neuromorphic Computing

Background:

• Increasing complexity of AI and data-intensive applications necessitates advanced computing architectures.
• Current spintronic processing-in-memory (PIM) solutions, while promising, face scalability issues due to their large device footprints.
• Memristive devices are crucial for high-performance and scalable PIM.

Purpose of the Study:

• To demonstrate a scalable two-terminal spin-orbit torque (SOT)-driven giant magnetoresistance (GMR) memristive device.
• To integrate data storage and logic functions into a single unit for efficient in-memory computing.
• To explore the potential of this device for artificial intelligence (AI) and neuromorphic applications.

Main Methods:

• Fabrication of a two-terminal GMR memristive device using a Pt/Co/Cu/CoTb stack.
• Modulation of SOT current amplitude to achieve multiple nonvolatile resistance states.
• Tuning of CoTb alloy composition to create additional resistance states with opposite polarities.
• Demonstration of synaptic plasticity (long-term potentiation and depression) via current pulse modulation.
• Simulation of a deep neural network (DNN) utilizing the GMR device.

Main Results:

• Achieved ten nonvolatile resistance states by modulating SOT current.
• Created additional states with opposite polarities by tuning alloy composition.
• Successfully demonstrated synaptic plasticity functions.
• Simulated DNN achieved 92% accuracy in handwritten digit recognition and image visualization.
• Exhibited basic Boolean logic functions, confirming processing capabilities.

Conclusions:

• The developed two-terminal GMR memristive device offers a scalable solution for advanced memristive memory and in-memory computing.
• This device integrates storage and logic, paving the way for more efficient AI hardware.
• The demonstrated capabilities in synaptic plasticity and Boolean logic highlight its potential for neuromorphic computing applications.