Study on Oxide-based 2-terminal Synapse Devices for Neuromorphic Application

Abstract

Nowadays, bio-inspired neuromorphic computing has been researched for massive processing application due to the von-Neumann bottleneck. In particular, deep neural network (DNN) architecture have a remarkable performance in pattern recognition thanks to their powerful capability of data processing. DNN has been developed on the basis of software, but it requires a lot of time and energy due to the limitation of the memory bandwidth. To solve this problem, hardware-based DNN has been proposed using cross-point array architecture owing to the ease of vector multiplication. For this reason, two-terminal synapse device should be required for realizing cross-point array-based neural network. Resistive switching memory (RRAM) have been considered for use as electronic synapse devices due to its scalability, two-terminal device, low-power operation, and CMOS compatibility. To implement DNN hardware using RRAM, many groups simulated synapse characteristics in neural network. In accordance with neural network simulation, synapse devices are required to have symmetry characteristics of potentiation/depression and high dynamic range to obtain high recognition accuracy. Unfortunately, most of RRAM devices exhibit non-linear conductance change and low dynamic range. In this thesis, we have studied the RRAM with 2-termianl synapse devices as a synapse device for achieving high pattern recognition accuracy of neural networks. First, interface RRAM was investigated by using TiOx-based resistive switching devices. We propose TiOx-based resistive switching device for neuromorphic synapse applications. This device is capable of 64-levels conductance states because of their optimized interface between the metal electrode and the TiOx film. To compensate the change in switching power with increasing pulse number, we propose the use of fixed voltage and current pulses in potentiation and depression conditions, respectively. By adopting a hybrid pulse scheme, the symmetry of conductance change under both potentiation and depression conditions is shown to be significantly improved. Both the improved conductance levels and the symmetry of conductance change are directly related with enhanced pattern recognition accuracy, which is confirmed by a neural network simulation. Second, filamentary RRAM was investigated for off-chip learning due to quantization of resistance state. To utilize a filament RRAM device, multi-level device was limited by during reset process. However, multi-level characteristics was limited by breakdown failure during reset process. Thus, we investigated the reset breakdown phenomenon of HfOx-based resistive memory for reliable switching operation in a fully CMOS compatible stack. Through the understanding on the effect of electrode materials and device area, our findings show that observed failure is attributed to additional oxygen vacancies close to the electrode interface, where switching is occurred. Therefore, RuOx serving as an oxygen diffusion barrier was introduced to suppress the generation of unwanted oxygen vacancies by preventing out-diffusion of oxygen through the electrode. As a result, significantly enhanced breakdown strength in HfOx/RuOx stack is achieved and resulting in improved cycle endurance with larger on/off ratio.