Stochastic Neuromorphic Computing Architecture Based on Voltage-Controlled Probabilistic Switching Magnetic Tunnel...

Area of Science:

• Spintronics
• Materials Science
• Computer Engineering

Background:

• Integrated circuits face limitations in power, area, and stability.
• Spintronic devices offer a path to overcome traditional computing paradigms.
• Novel integration of spintronic devices with computing architectures is essential.

Purpose of the Study:

• Investigate the switching mechanism of Magnetic Tunnel Junctions (MTJs) using Voltage-Controlled Magnetic Anisotropy (VCMA) and Spin Hall Effect (SHE).
• Analyze the dynamic characteristics of VCMA-assisted switching SHE-MTJ devices.
• Develop a foundation for stochastic and neuromorphic computing architectures.

Main Methods:

• Established a macrospin approximation model based on the Landau-Lifshitz-Gilbert (LLG) equation.
• Investigated the synergistic effects of VCMA and SHE on MTJ switching.
• Incorporated a thermal fluctuation field to enable stochastic switching behavior.

Main Results:

• VCMA voltage pulses significantly reduce required spin Hall current density and pulse width, minimizing ohmic losses and Joule heating.
• Demonstrated voltage-controlled SHE-MTJ devices exhibiting stochastic switching with a sigmoidal voltage-probability response.
• Proposed and designed an in-memory computing architecture for binarized Convolutional Neural Networks (CNNs).

Conclusions:

• The study provides a feasible spintronic implementation for low-power, high-energy-efficiency edge-side intelligent chips.
• The developed architecture achieved 72.49% Top-1 accuracy on CIFAR-10 using SqueezeNet with only 1.25 million parameters.
• This work lays the groundwork for advanced spintronic-based computing, particularly for AI and neuromorphic applications.