Nano-Magnetic Logic based Architecture for Edge Inference using Tsetlin Machine

High learning accuracy with less time, low energy and small memory footprint made Tsetlin Machine (TM) an emerging edge inference alternative to traditional machine learning algorithms such as neural networks (NN). Advances in nanomagnetic logic are considered a potential alternative to CMOS-based digital circuits due to their non-volatility, low power and high integration density. Adapting the advantages of these two technologies, we propose a novel nano-magnetic logic-based architecture for edge inference using the Tsetlin machine to build energy-efficient applications. This design can be expanded to construct energy-efficient Convolution TM applications. The proposed TM inference architecture is designed by exploiting the shape and positional hybrid anisotropy and the antifer-romagnetically coupled fixed input majority gate (MG). The micromagnetic tool is used for the simulation framework. The proposed design to implement 2 clause TM inference requires 15 majority gates comprising a total of 60 nanomagnets and 14 interconnect magnets, requiring $1.24X10^{-19}J$ energy per operation which is $10^{5}$ to $10^{6}$ times less compared to CMOS logic, and the power consumed by the total number of nanomagnets for implementing the proposed architecture is 0.00082 $\mu \mathrm{W}$.