In-Memory Computing Architecture for Tsetlin Machine Inference using MTJ-based Crossbar

In this paper, we investigate the suitability of magnetic tunnel junctions (MTJ) as a promising memory technology that is known for its non-volatility, swift operation, minimal power consumption, robust endurance, high density, and compatibility with vanilla CMOS. We introduce a novel in-memory computing architecture for inference of a Tsetlin machine using an MTJ crossbar for machine learning applications. In this architecture, Boolean feature inputs are represented as digital voltages, which enable parallel current paths that are determined by the MTJ resistive memory states. The resultant proportional column current is subsequently converted back into the Boolean domain for additional digital processing. Our proposed architecture reliably achieves desired inference functionality, consuming up to 9.18% less energy for MNIST and 9.08% less for F-MNIST compared to CMOS equivalents, while maintaining non-volatility, high density, and CMOS compatibility.