Dynamic Analysis of the Effect of the Device-to-Device Variability of Real-World Memristors on the Implementation of Uncoupled Memristive Cellular Nonlinear Networks

Abstract

Cellular Nonlinear Networks (CNNs) are a well established computing approach in the domain of analog computing, known for massive parallelism and data processing locality that enable efficient hardware implementations. Combining CNN with non-volatile memristive devices holds the promise to overcome technological hurdles, like scalability issues, and high energy consumption, while also introducing richer dynamics into the field of CNN. Memristive devices based on the valence change mechanism (VCM) show great properties, like bipolar switching, tuneable resistance and non-volatility that are essential for the design of memristive CNN (M-CNN). In this study we design and investigate an uncoupled M-CNN cell implementing the EDGE detection task. This is the first paper investigating the resilience of M-CNN against device-to-device variability. To this end the first experimentally acquired Dynamic Route Map (DRM) of the M-CNN cell is employed. The comparison with simulations results allows for investigating the effect of mechanisms in the VCM device on the performance of the cell. The result of the computation is stored in the VCM device despite the unavoidable variability in the electrical behaviors of different device samples. Furthermore, the theoretically predicted richer dynamics of M-CNNs over traditional CNNs is demonstrated. This work provides crucial insights into design considerations of M-CNNs, especially as here first steps towards the comprehensive analysis on the effect of imperfections and variability of the memristor on M-CNN cell are taken.