Spintronic Bayesian Hardware Driven by Stochastic Magnetic Domain Wall Dynamics

As AI expands into safety-critical domains, reliability and uncertainty estimation have become essential, leading to the emergence of probabilistic computing. Conventional hardware platforms, such as CMOS, are inefficient for such implementations due to their deterministic nature, where the suppression of intrinsic fluctuations incurs high energy and computational costs to simulate probabilistic dynamics. At the core of this challenge is the absence of novel materials platforms and device architectures capable of natively supporting probabilistic behavior. To address this challenge, we present a Magnetic Probabilistic Computing (MPC) platform-an energy-efficient, scalable hardware accelerator, where its inherent magnetic dynamics directly enable probabilistic computing. The MPC platform is based on Magnetic Tunnel Junction (MTJ) materials, which brings together three key mechanisms-thermally induced Domain Wall (DW) stochasticity, Voltage-Controlled Magnetic Anisotropy (VCMA), and Tunneling Magnetoresistance (TMR)-to achieve fully electrical, tunable probabilistic functionality. As a representative demonstration, we implement a Bayesian Neural Network (BNN) inference structure and validate its functionality through CIFAR-10 classification tasks by feeding experimentally acquired Gaussian distribution signals into a simulated BNN framework. Compared to standard 28 nm CMOS implementations, our approach achieves a seven-orders-of-magnitude improvement in the overall figure of merit of the basic MPC unit, demonstrating substantial gains in area efficiency, energy consumption, and speed. These results highlight the promise of harnessing intrinsic spintronics material stochasticity within the MPC platform, opening new pathways toward reliable and trustworthy physical AI systems.