Investigation on the Impact of Memristive Device Variability on Uncoupled Memristive Cellular Nonlinear Networks

Conventional von Neumann architectures suffer from an inherent bottleneck arising from the physical separation of memory and processing units, motivating alternative bio‐inspired computing paradigms. Cellular nonlinear networks (CellNNs) offer one such alternative, but implementations that rely on von Neumann‐style processors inherit the memory bottleneck. The integration of memristive devices into CellNN implementations can mitigate this bottleneck, enabling in‐memory computation with volatile and non‐volatile storage, and enhanced cell dynamics. In this work, an uncoupled Memristive CellNN (M‐CellNN) breadboard cell is physically implemented, and the device‐to‐device (D2D) variability of five nominally identical Pt///Ti/Pt (HOTO) VCM devices is experimentally characterized. These measurements calibrate variability‐aware SPICE‐level simulations of a M‐CellNN array implementing EDGE detection. A pixel‐level EDGE detection accuracy of 100% is demonstrated across five input images, with all true edge pixels correctly identified and no spurious detections, under D2D variability conditions calibrated from experimental data. Since cells operate independently, errors remain spatially isolated and do not cascade across the array. Furthermore, 100% accuracy is maintained under global passive component deviation of 20% and realistic local component mismatch of 2%, with the worst‐case robustness boundary located at 4% mismatch for the most demanding image, confirming suitability for monolithic hardware integration.