RRAM 시냅스와 TS 소자를 사용한 뉴로모픽 하드웨어 시스템

Abstract

Currently, the 4th industrial revolution is underway, and artificial intelligence is at the center of it. Significantly, an artificial neural network that mimics the human brain structure shows performance that surpasses human cognitive ability in multiple fields such as image recognition, natural language processing, face verification, Etc. However, the conventional von-Neumann computer structure has a disadvantage in that bottleneck occurs when performing vector-matrix multiplication, a core function of an artificial neural network, resulting in lowered system efficiency. In this study, we propose a resistive random-access memory (RRAM)-based neuromorphic system to overcome the limitations of the existing CMOS-based computer architecture. First, we analyze the characteristics that RRAM devices used in neuromorphic systems should have and propose a method to complement the characteristics of each synaptic device through a hybrid synapse structure. In addition, we explain an operation method that can increase the performance and energy efficiency of a neuromorphic system by using an OTS-based selector device. The convolution operation in a convolutional neural network by introducing a 3D integrated TiOx-based resistive switching device (RSD) array to act as the kernel is investigated. The 3D integrated TiOx-based RSD exhibits a gradual SET/RESET behavior, which enables multi-level characteristics. A pair of layers in the 3D structured array is used to achieve the positive and negative weights. Based on these results, we demonstrated the 3 x 3 Laplace kernel for edge detection using images from the Columbia Object Image Library (COIL) dataset in a simulator to validate the convolution operation. Evaluate a weight quantization method for neural network and investigate requirements as a synapse device to ensure recognition accuracy of 3-layer perceptron neural network is studied. We use TaOx-based RRAM with RuOy electrode as a synapse device in the neural network and study its operation. According to our simulation results, a synapse device for normal operation of quantized neural network (QNN) without degrading its recognition accuracy requires a minimum of 2-bit conductance states with good retention and linear I-V characteristics. A novel neuromorphic architecture with TiOx-based interfacial RRAM and CBRAM-based filamentary RRAM for highly accurate NN training and long-term inference reliability to overcome the performance degradation in hardware neural networks (NNs) with non-ideal synapse devices is proposed. We used a threshold-triggered training scheme, in which interfacial and filamentary RRAMs were programmed in a complementary fashion. This took advantage of the long retention time of the filamentary RRAM and the high-resolution, symmetric weight update in the interfacial RRAM. Additional device parameters were evaluated, such as linearity, precision, variation, and retention time. An excellent pattern recognition accuracy of ~97% was achieved during training with the MNIST dataset. Thus, reliable inference accuracy after training was maintained using the filamentary RRAM. A method to overcome the short failure of synaptic devices, which is fatal in artificial neural network systems, using OTS-based selector devices is investigated. We characterize a 32 x 32 array of PCMO (PrxCa1-xMnO3)-based resistive random-access memory (RRAM) devices and obtained short failure statistics. Our analysis shows that the inference accuracy of a neural network system can be significantly degraded when only a small number of devices become short-failed in the array. We analyze that the leading cause of the pattern recognition accuracy degradation is the unexpected neural excitation by a high current through the shorted device. To prevent such accuracy degradation, we fabricate an ovonic threshold switching (OTS)-based electrical fuse device and experimentally verify its fuse functionality utilizing the void formation phenomenon in chalcogenide materials under high current conditions. After disconnecting short-failed devices with the fuse operation, the recognition accuracy is recovered from 10% to 97%, which is on par with the performance with no short-failed devices in the array. An experimental demonstration of a highly linear (R2=0.995) and area-efficient 1S1R analog-to-stochastic converter (ASC), a core enabler for fully parallel weight update operation in cross-point synaptic array-based neural networks, is studied. We confirm that the previously reported, sigmoid-like ASCs cannot be used for stochastic updates due to limitations of nonlinear characteristics. We analyze the characteristics of the device-based ASC for stochastic update operation to show that our ASC consumes 1,000 times less power than a CMOS-based ASC and does not require a power-hungry digital-to-analog converter (DAC). A software-level image recognition accuracy (97.96%) is achieved when performing neural network training with our ovonic threshold switching (OTS) device-based ASC. Stochastic resonance exploiting inherent stochastic characteristics of OTS threshold voltage to enhance inference performance of neural networks is proposed. First, threshold switching of OTS device is characterized, and a method to detect signal using OTS device is proposed. Next, we emphasize the importance of stochasticity in the threshold voltage by explaining how stochasticity generates stochastic resonance that enables the system to detect weak signals. Finally, through the evaluation of the inference performance of the artificial neural network, it is confirmed that the inherent stochasticity can effectively restore the degraded MNIST image in poor visibility conditions in the OTS device. As a result, the recognition accuracy was improved from 10.28% to 95.78% when the stochasticity characteristic was reflected. These results show that the stochasticity in the device can improve system performance, contrary to the traditional belief that it negatively affects systems.