Statistical memristor model and its applications in neuromorphic computing

Abstract

More than forty years ago, Professor Chua predicted the existence of the memristor to complete the set of passive devices that previously includes only resistor, capacitor, and inductor. However, till 2008 the first physical realization of memristors was demonstrated by HP Lab. The unique properties of memristor create great opportunities in future system design. For instance, the memristor has demonstrated the similar function as synapse, which makes it promising in neuromorphic circuits design. However, as a nano-scale device, the process variation control in the manufacturing of memristors is very difficult. The impact of the process variations on a neural network system that relies on the continuous (analog) states of the memristor could be significant due to the deviation of the memristor state from the designed value. So a complete process variation analysis on memristor is necessary for the application in neural network. Due to the different physical mechanisms, TiO2-based memristor and spintronic memristor demonstrate very different electrical characteristics even when exposing the two types of devices to the same excitations and under the same process variation conditions. In this work, the impact of different geometry variations on the electrical properties of these two different types of memristors was evaluated by conducting the analytic modeling analysis and Monte-Carlo simulations. A simple algorithm, which is based on the latest characterization method of LER (line edge roughness) and thickness fluctuation problems, was proposed to generate a large volume of geometry variation-aware three-dimensional device structures for Monte-Carlo simulations.We investigate the different responses of the static and memristive parameters of the two devices and analyze its implication to the electrical properties of the memristors. Furthermore, a process-variation aware device model can be built based on ourwork. Both corner model and statistical model can be provided depending on users’ requirements. Our device models make it possible for scientists and engineers to design neuromorphic circuits with memristive devices, and therefore, to convert virtual neural network in super computer to the real hardware memristive system in the future. Rather than the existing crossbar-based neuron network designs, we focus on memristor-based synapse and the corresponding training circuit to mimic the real biological system. The basic synapse design is presented, and the training sharing scheme and explore design implication on multi-synapse neuron system have been explored.