Memristive passive crossbar circuits hold great promise for neuromorphic computing, offering high integration density combined with massively parallel operation. However, scaling up the integration complexity of such circuits remains challenging due to low device yield, stemming from the intrinsic properties of filamentary switching and limitations in current crossbar fabrication technologies. Here, we report a scalable passive crossbar device technology achieved through a co-design approach for memristors and crossbar structures. The proposed hardware platform is fabricated using CMOS-compatible processes without complex and high-temperature steps, enabling high device yield along with reliable and multibit operation. Importantly, the fabrication process is successfully scaled to a 4-inch wafer, maintaining an average device yield (>similar to 95%) and preserving key switching characteristics. The potential of this platform is showcased by implementing image classification of the fashion MNIST benchmark with an ex-situ trained spiking neural network. We believe that our work represents a significant step toward brain-scale neuromorphic computing systems.
