Hybrid Memristor-CMOS Computer for Artificial Intelligence: from Devices to Systems

Abstract

Neuromorphic computing systems, which aim to mimic the function and structure of the human brain, is a promising approach to overcome the limitations of conventional computing systems such as the von-Neumann bottleneck. Recently, memristors and memristor crossbars have been extensively studied for neuromorphic system implementations due to the ability of memristor devices to emulate biological synapses, thus providing benefits such as co-located memory/logic operations and massive parallelism. A memristor is a two-terminal device whose resistance is modulated by the history of external stimulation. The principle of the resistance modulation, or resistance switching, for a typical oxide-based memristor, is based on oxygen vacancy migration in the oxide layer through ion drift and diffusion. When applied in computing systems, the memristor is often formed in a crossbar structure and used to perform vector-matrix multiplication operations. Since the values in the matrix can be stored as the device conductance values of the crossbar array, when an input vector is applied as voltage pulses with different pulse amplitudes or different pulse widths to the rows of the crossbar, the currents or charges collected at the columns of the crossbar correspond to the resulting VMM outputs, following Ohm’s law and Kirchhoff’s current law. This approach makes it possible to use physics to execute direct computing of this data-intensive task, both in-memory and in parallel in a single step. First of all, I will present a comprehensive physical model of the TaOx-based memristor device where the internal parameters including electric field, temperature, and VO concentration are self-consistently solved to accurately describe the device operation. Starting from the initial Forming process, the model quantitatively captures the dynamic RS behavior, and can reliably reproduce Set/Reset cycling in a self-consistent manner. Beyond clarifying the nature of the Forming and Set/Reset processes, a bulk-like doping effect was revealed by the model during Set and supported by experimental results. This phenomenon can lead to linear analog conductance modulation with a large dynamic range, which is very beneficial for low-power neuromorphic computing applications. Second, an integrated memristor/CMOS system consisting of a 54×108 passive memristor crossbar array directly fabricated on a CMOS chip is presented. The system includes all necessary analog/digital circuitry (including analog-digital converters and digital-analog converters), digital buses, and a programmable processor to control the digital and analog components to form a complete hardware system for neuromorphic computing applications. With the fully-integrated and reprogrammable chip, we experimentally demonstrated three popular models – a perceptron network, a sparse coding network, and a bilayer principal component analysis system with an unsupervised feature extraction layer and a supervised classification layer – all on the same chip. Beyond VMM operations, the internal dynamics of memristors allow the system to natively process temporal features in the input data. Specifically, a WOx-based memristor with short-term memory effect caused by spontaneous oxygen vacancy diffusion was utilized to implement a reservoir computing system to process temporal information. The spatial information of a digit image can be converted into streaming inputs fed into the memristor reservoir, leading to 100% accuracy for simple 4×5 digit recognition and 88.1% accuracy for the MNIST data set. The system was also employed for solving other nonlinear tasks such as emulating a second-order nonlinear system.