아날로그 딥러닝 가속기 구현을 위한 산소 이온 기반의 3단자 시냅스 소자

Abstract

On account of the exponential development of artificial intelligence, the amount of computation required has increased dramatically. It is no longer possible to keep up with the requested computation with a processing and memory unit that is concomitant to the existing von-Neumann architecture. Therefore, the proposed analogue neuromorphic computing device promotes device characteristics completely different from the achievement conditions so far. Although various materials and ions have been studied, we would like to introduce a CMOS-compatible and easy-to-control MO-ECRAM. MO-ECRAM consists of three-terminals, so it is relaxed to regulate by separating read and write operations, and has strengths in terms of linearity and symmetry. The device we proposed can control the maximum conductance by utilizing the part where the gate stack is not covered. Consequently, it was confirmed that the resistance changes in the corresponding region, but this change is insignificant and negligible. MNIST simulation was conducted to prove that this design has a positive effect on neural network training. As a result, it was authorized to be fault-tolerant due to existence of the ungated region, which will act as an important factor when considering the massive array operation. The device configures selector-free when MO-ECRAM is implemented as a neuromorphic chip beyond a synaptic transistor, it is advantageous in terms of power consumption and scaling.