Stochastic Switching of Nanomagnets for Post-CMOS Computing

Magnetization reversal in spintronic devices is stochastic and is characterized by time-varying thermal noise. Rather than viewing the device stochasticity as a disadvantage, the inherent probabilistic switching dynamics of a nanomagnet can be used to mimic the computational primitives of several Post-Moore computing paradigms. For instance, neuromorphic computing platforms enabled with stochastic neurons and synapses can be used to perform probabilistic inference in Restricted Boltzmann Machines and Deep Belief Network architectures. More generally, they can be used to implement Boltzmann machines enabled by stochastic units that can be used to find optimal solutions in combinatorial optimization problems. Direct mapping of the stochastic computational units of such probabilistic computing paradigms results in reduced energy and area overhead with respect to CMOS implementations (enabled by deterministic hardware). Further, due to stochastic state updates over time, such probabilistic computing paradigms offer the possibility of state compression of their units in comparison to their deterministic counterparts. Here, we report our recent work on using the stochastic magnetization dynamics of a Magnetic Tunnel Junction to enable various genres of post-Moore computing like Spiking Neural Networks, Boltzmann Machines and Bayesian Inference Networks.