Hardware Non-Idealities and Reliability in RRAM Crossbar Neuromorphic Character Recognition Systems

Resistive Random Access Memory (RRAM) crossbar arrays have emerged as a promising platform for in-memory computing due to their high density, non-volatility, and ability to perform massively parallel analog matrix operations. In this work, we investigate the reliability of an RRAM-based template recognition framework for EMNIST-derived uppercase letters. Representative binary character templates are stored in a 784 x 26 crossbar array, and recognition is performed through XNOR-based similarity evaluation and current accumulation. We study the impact of multiple RRAM non-idealities, including random faults, data-dependent faults, spatially correlated clustered faults, and conductance variation, on recognition accuracy. We further evaluate two established fault-tolerance mechanisms, Triple Modular Redundancy (TMR) and Algorithm-Based Fault Tolerance (ABFT), and compare their effectiveness across a range of RRAM fault scenarios. Experimental results show that TMR provides the highest resilience under severe fault conditions at the cost of significant hardware overhead, while ABFT offers a lightweight alternative with lower implementation cost and moderate fault tolerance. The results highlight the trade-offs between reliability improvement and implementation overhead, providing practical guidelines for selecting appropriate protection mechanisms in RRAM-based recognition systems.