Sputtered Hafnium Oxide Based Memristive Crossbar Array for Energy Efficient MAC Operations

Abstract

Memristor crossbar-based analog computing architectures have emerged as promising candidates for energyefficient in-memory computing due to their high integration density and parallel computing capability. However, leakage and sneak path current in the memristive crossbar remain a critical challenge that affects the accuracy of multiply and accumulate operations. In this work, a $4 \times 4$ memristive crossbar array with a single dummy column assisted leakage compensation technique is proposed to improve the accuracy of multiply and accumulate operations. The memory element is a physically fabricated $\text{Ag}/\text{HfO}_{x}/\text{Pt}$ memristor, grown by reactive RF magnetron sputtering, whose electrical behavior is captured through experimentally measured I-V characteristics and modeled using the VTEAM framework. The proposed architecture was validated by applying random 8-bit binary input vectors across the four word lines, with quantized synaptic weights encoded as memristive conductance states. The resulting MAC outputs across all four bit lines match the target computed values with zero errors. The crossbar array exhibits a propagation delay of $368.626 f s$, and a total energy consumption of $1.275 p J$ per read operation, with individual active cells dissipating $77.169 f J$ and dummy cells consuming $5.15 f J$ per read cycle. These results demonstrate that sputtered $\text{HfO}_{x}$ based memristive crossbars, combined with dummy column leakage compensation enable accurate, lowpower analog MAC computation, underscoring their suitability for next generation neuromorphic and in-memeory computing systems.