A fully hardware-based memristive neural network is capable of delivering high computing throughput and power efficiency.

Large fluctuations of key parameters in cation-based resistive random access memory (RRAM), which originate from stochastic growth of metallic conductive filaments, always impose a significant barrier to the practical application of memory devices. Here, we propose an ordinary bilayer oxide structure of Ag/TaO_x/TaO_y/Pt (x < y) to address this issue and achieve the performance enhancement of memory cells. This memory device is inclined to form nano-cone-shaped filaments under external bias, and the tips of filaments provide an electric field concentration, achieving an effective control of filament growth. Compared with the single-layer device Ag/TaO_x/Pt, the bilayer device manifests a larger ON/OFF ratio, much lower operation voltages and RESET currents, a higher response speed, and better uniformity. The insertion of the oxygen-rich layer also brings about the tunability of switching voltages and the elimination of the negative-SET phenomenon. Our experiments might pave the way for high-density emerging memory commercial applications.