All-Electrical Control of Spin Synapses for Neuromorphic Computing: Bridging Multi-State Memory with Quantization for...

Area of Science:

• Spintronics
• Neuromorphic Computing
• Materials Science

Background:

• Neuromorphic computing requires energy-efficient memory devices mimicking synaptic behavior.
• Current devices face challenges in accuracy and adaptability for brain-inspired systems.

Purpose of the Study:

• To develop and evaluate all-electrically controlled, field-free spin synapse devices for neuromorphic applications.
• To benchmark device performance, focusing on cycle-to-cycle variation and multi-state memory capabilities.

Main Methods:

• Three spintronic device structures were designed: Néel orange-peel effect, interlayer Dzyaloshinskii-Moriya interaction (i-DMI), and tilted anisotropy.
• A benchmarking framework was used to assess cycle-to-cycle (CTC) variation.
• Devices were implemented in convolutional neural networks (CNNs) with post-training quantization.

Main Results:

• The tilted anisotropy device demonstrated an 11-state memory with minimal CTC variation (2%).
• Per-channel quantization in ResNet-18 on the CIFAR-10 dataset achieved high classification accuracy (up to 81.51% with min-max, 81.12% with MSE observers).
• Performance closely approached baseline accuracy, validating the device's potential.

Conclusions:

• Field-free spintronic synapses offer a promising, area-efficient solution for advanced neuromorphic architectures.
• These devices integrate multi-state functionality and robust switching, advancing energy-efficient, high-performance computing inspired by neural processes.