Research on highly reliable synaptic devices for neuromorphic computing

Abstract

Artificial intelligence technology based on big data has been developed rapidly, and accordingly, an enormous amount of data is required. Despite these advances, hardware for implementing this technology is composed of a conventional von Neumann structure in which a memory and processing unit are physically separated, causing many inefficiencies in the process of transporting data. Neuromorphic computing can overcome the limitations of the existing von Neumann structure by enabling efficient data processing through simultaneous storing and computing as the human brain. Therefore, research on resistance change-based two- or three-terminal synapse devices that can mimic human synapses in terms of hardware is being actively conducted. However, the reliability of the existing two-terminal resistive switching device is deteriorated due to the disorder of the conductive filament that causes resistance change, which is a major obstacle to commercialization. In addition, conventional three-terminal synaptic devices have limitations in online training due to nonlinear conductance updates, low endurance, and high-power consumption. In this study, a highly reliable resistance change device is fabricated through structural engineering by inserting a nanoporous structure and a buffer layer using an amorphous material. In addition, a gate injection-based synaptic transistor with a high linear update through thermionic emission mechanism, which is compatible with conventional silicon-based CMOS platforms, is developed for online-line training.