Process-in-Memory Using a Magnetic-Tunnel-Junction Synapse and a Neuron Based on a Carbon Nanotube Field-Effect Transistor

Abstract

Bio-inspired computing attempts to emulate the computations performed by biological neural networks, like those in the human brain, to benefit from efficiencies in flexibility, parallel processing, and, most importantly, learning. We simulate a fully nonvolatile spin-based synapse utilizing a magnetic tunnel junction (MTJ) and a neuron based on a carbon nanotube field-effect transistor (CNTFET). The modeled synapse benefits from the features of MTJs like nonvolatility, high endurance, and reconfigurability. The features of CNTFETs are utilized to streamline the design. The functionality and characteristics of the synapse in the inevitable presence of process variations are validated using Monte Carlo simulations. An associative memory, designed and simulated using the spin-based synapse, performs very similar to its ideal software-based counterpart, with about 8% difference in successful recall rate.