Memristor-Based Crossbar Applications in Machine Learning and Analog Circuit Design

Abstract

The shift to the AI/ML driven data processing resulted in a significant research demand in this field. Due to the inefficiency of the conventional computers, analog and mixed-signal data processing structures became feasible solutions and memristor-based crossbars are among the popular ones. They are able to solve Vector-Matrix-Multiplications (VMMs) with a higher energy efficiency and competitive accuracy compared to digital computers. This dissertation proposes the implementation of the memristor-based crossbars in various ML applications including linear algebraic operations, Markov Chains, Eigencentrality measures, Gaussian Processes, complex networks, power method and custom analog circuit applications to perform critical tunable filtering tasks.

The core block of this proposed work, memristor crossbars, are evaluated and analyzed in terms of non-idealities and noise in the system and the accuracy limits of the system is discussed. The overall performance comparison with the existing conventional methods are also discussed and given for each application. The main non-idealities discussed in this work are but not limited to: interconnect resistance and capacitance, source driver resistance, finite gain and bandwidth peripheral circuitry, random op-amp offset, memristor programming error, thermal noise, and programming noise.

The comparison between the memristor crossbars and other conventional digital and analog systems revealed that memristor crossbars provide significant energy savings while maintaining a competitive accuracy. When they are being used as an analog processing core in a digital system, the round trip resulting from the analog to digital conversion limits their efficiency. The advances in material science and circuit design is projected to improve memristor-based processing accuracy to a digital precision scale.