Neuromorphic Computing and Spiking Neural Network using Spin Devices

Abstract

Neuromorphic computing is the research trending research area which aims for next generation hardware solution. The present CMOS technology is facing issues due to saturating Moore’s law, Dennard’s scaling and von Neumann bottleneck. Therefore, it is necessary to figure out the unconventional devices, circuits, architectures and algorithms that can render the path for future nanotechnology. This work entitled “Neuromorphic computing and Spiking neural networks using spin devices” presents the basic approach of spin based neuromorphic computing, its advantages and challenges. The hardware implementation of computation intensive tasks using CMOS are highly inefficient in terms of area, power and latency. Hence, neural networks are preferred to be implemented using crossbar array architecture to save area and power consumption. Spin device-based hardware implementation is highly efficient as compared to other non-volatile memories (NVMs) and CMOS. Moreover, the bio-plausible spiking neural networks are much more energy efficient than conventional artificial neural networks. Spintronic devices provide high endurance, non-volatility, CMOS compatibility, energy and area efficiency. Hence, spin-based devices are explored for implementation of neural networks. The computations involved in neural networks are multiply and accumulate operations which are power hungry. Hence, spin based neuromorphic implementation of such dense deep neural networks (DNNs) is a suitable approach. This work explains the approach and tools used for simulating spiking and non-spiking neural networks using NVMs with emphasis on spin devices. The spin-based devices, neuron models, synaptic devices, neural network training algorithms, and crossbar array-based implementation is described in this work. With the help of the DESTINY tool, the performance of SOT-MRAM has been improved compared to other NVM devices.