Highly‐Uniform Passive Crossbar Arrays of Resistive Switching Random Access Memory (RRAM) for In‐Memory Computing Applications

Resistive switching random-access memory (RRAM) is gaining attention for its ability to support various data-intensive computing tasks via in-memory computing (IMC). The use of passive RRAM arrays is hindered by sneak-path issues, programmability challenges, and concerns regarding reliability and robustness. In this study, we address the initialization problem and propose a strategy to efficiently program passive arrays, thereby overcoming these limitations. We validate our approach using 32 × 32 crossbar arrays of Pt/HfO2/Ti RRAM, achieving 99.5% functionality and encoding various conductance weights with an error below 3%. To demonstrate the potential of RRAMs, we experimentally perform 24 × 24 discrete cosine transform (DCT) image reconstruction, achieving an accuracy higher than 95%. Also, we implement a single-layer classifier for feature recognition on the MNIST digit dataset, achieving 84.6% accuracy compared to the theoretical 86.1%. We finally develop an in-memory solver for combinatorial optimization problems, such as the quadratic assignment problem, leveraging the full parallelism of the crossbar configuration and implementing a quantum-inspired parallel annealing method, and demonstrate non-linear regression up to the fourth order.