Beyond binary: multistate computing and neural systems with compound magnetic tunnel junctions

Abstract The magnetic random access memory (MRAM) architecture, comprising nonvolatile, fast, and energy-efficient magnetic tunnel junctions (MTJs), presents a promising platform for implementing native in-memory and neuromorphic computing tailored for edge AI applications. Integrating the MTJ-based neural network of neurons and synapses within the MRAM architecture overcomes the conventional von Neumann bottleneck, enabling computation to be performed directly where the data is stored. Here, we present an overview of MTJ technologies proposed for artificial neural networks, specifically focusing on the recent advancements in the compound MTJ concept. The compound MTJ − comprising an array of MTJ cells on a 3-terminal spin-orbit torque device structure − is capable of hosting multiple resistance states and tunable conductance levels, which are well-suited for emulating the flexibility and plasticity of artificial synaptic weights. We highlight developments on compound MTJ designs with demonstrable multistate stability, and enhanced synaptic strength and resolution through the bimodal tuning of tunnel magnetoresistance and discrete states. Next, we discuss a promising strategy to expand the switching voltage window between successive states through planar rotations of constituent MTJs, engendering low write errors and high synaptic tolerance. Finally, we discuss emerging material platforms − field-immune, low-power antiferromagnets and altermagnets − as enablers of high density, multistate compound MTJs scalable and robust in-memory and
neuromorphic computing technologies.