Leveraging 3D Vertical RRAM to Developing Neuromorphic Architecture for Pattern Classification

The crossbar architecture with resistive random-access memory (RRAM) devices presents many advantages in realizing matrix-based computations and achieves success in neural network implementation. However, the rapid growth of network size demands even denser structures. In this paper, we investigate the neuromorphic hardware design based on the three-dimensional vertical RRAM (3D VRRAM) with an even/odd word line (WL) structure. The increased interconnects of VRRAM aggravate the chronic problems of the crossbar structure like the sneak path currents. We address this issue by attaining a balanced structure with high nonlinear RRAM devices. Furthermore, the impact of complicated signal routing and control due to the vertically stacked structure can be alleviated through architectural level optimization. A three-layer VRRAM structure is demonstrated for neuromorphic design by showing that 8X8-pixel images were successfully classified into three alphabet characters on this structure. The example design also verifies that the 3D VRRAM with even/odd WL structure is beneficial to acquire high area efficiency.