Fab-Friendly Hafnium Oxide based Conductive Bridging Random Access Memory for Neuromorphic Devices

Abstract

Recently, there are various applications of artificial intelligence (AI) in industries related to voice, image recognition and data analysis. These AI technologies play an important role in the implementation of neural network technology that mimics the human brain. However, the conventional von Neumann structure is connected in series between memory and CPU, resulting in a bottleneck phenomenon. On the other hand, the neuromorphic chips are capable of parallel operations, which can process large amounts of data simultaneously and quickly. Therefore, there are many research is in progress and especially, conductive bridging random access memory (CBRAM) have been widely used in AI technologies related to neural social network.[1] We investigated the characteristics of hafnium oxide (HfOx) based CBRAM which has many advantages to be used as a neuromorphic memory.[28,29] The neuron device has a large horizontal electric field, which causes the filament to disappear slowly, making it easier to control the resistance. In addition, 4F2 process is possible with 2-terminal device, so integration degree can be increased. In this thesis, I researched that HfOx based CBRAM cell with 113nm cell size demonstrated neuronal functionality by negative differential resistance (NDR) region. It has good CMOS compatibility as well as its process is simple.[30-33] HfOx based CBRAM device was fabricated with the simple structure of metal-insulator-metal (MIM), and HfOx was used as an active layer, which is sandwiched between an reactive electrode CuTe and inert electrode TiN. I measured not only I-V curve but the stochastic integrate and fire characteristic. Finally, I presented how the device can operate as a neuromorphic device. In addition, the uniformity and the reliability of the device such as endurance cycling and retention time were investigated. |최근 음성, 이미지 인식 및 데이터 분석과 관련된 산업에서 인공 지능 (AI)의 다양한 응용 활발하다. 이러한 AI 기술은 인간의 두뇌를 모방하는 신경망 기술 구현에 중요한 역할을 한다. 하지만, 기존의 폰 노이만 (von Neumann) 구조는 메모리와 CPU 사이에 직렬로 연결되어 데이터 처리 속도가 느려지는 병목 현상(bottleneck phenomenon)이 발생하게 된다. 반면에, neuromorphic 칩은 병렬 연산이 가능하여 대량의 데이터를 동시에, 뿐만 아니라 신속하게 처리 할 수 ​있다. 따라서 이에 관련된 많은 연구가 진행되고 있으며, 특히 신경 망 브릿징 네트워크와 관련된 인공 지능 기술에 CBRAM (Conductive Bridging Random Access Memory)이 널리 연구되고 있다. [1] 본 논문은 neuromorphic 메모리로 사용되는 많은 장점을 가진 하프늄 산화물 (HfOx) 기반 CBRAM의 특성을 연구했다. [28] 뉴런 소자는 큰 수평 전계를 가지므로 필라멘트가 천천히 사라져 저항을 쉽게 제어 할 수 있다. 또한 2 단자 소자로 4F2 공정이 가능하므로 집적도를 향상시킬 수 있다. 본 논문에서는 113nm 셀 크기의 HfOx 기반 CBRAM 셀이 보여주는 뉴런 소자의 특성에 대해 연구했다. 이 소자는 CMOS 호환성이 뛰어나고 공정이 간단한 장점이 있다. 구조 또한 단순하여, 금속-절연체-금속 (MIM) 구조로 제작했으며, HfOx는 고체 전해질으로 쓰이고, 이는 활성 전극 CuTe과 비활성 전극 TiN 사이에 끼워진다. 우리는 I-V 곡선뿐만 아니라 integrate-and-fire (IF) 특성 또한 측정했다. 마지막으로, 이 소자가 어떻게 neuromorphic 장치로서 작동할 수 있는지를 구현할 것이다. 또한, 내구성 사이클링 및 데이터 유지시간과 같은 소자의 균일성 및 신뢰성이 연구되었다.; Recently, there are various applications of artificial intelligence (AI) in industries related to voice, image recognition and data analysis. These AI technologies play an important role in the implementation of neural network technology that mimics the human brain. However, the conventional von Neumann structure is connected in series between memory and CPU, resulting in a bottleneck phenomenon. On the other hand, the neuromorphic chips are capable of parallel operations, which can process large amounts of data simultaneously and quickly. Therefore, there are many research is in progress and especially, conductive bridging random access memory (CBRAM) have been widely used in AI technologies related to neural social network.[1] We investigated the characteristics of hafnium oxide (HfOx) based CBRAM which has many advantages to be used as a neuromorphic memory.[28,29] The neuron device has a large horizontal electric field, which causes the filament to disappear slowly, making it easier to control the resistance. In addition, 4F2 process is possible with 2-terminal device, so integration degree can be increased. In this thesis, I researched that HfOx based CBRAM cell with 113nm cell size demonstrated neuronal functionality by negative differential resistance (NDR) region. It has good CMOS compatibility as well as its process is simple.[30-33] HfOx based CBRAM device was fabricated with the simple structure of metal-insulator-metal (MIM), and HfOx was used as an active layer, which is sandwiched between an reactive electrode CuTe and inert electrode TiN. I measured not only I-V curve but the stochastic integrate and fire characteristic. Finally, I presented how the device can operate as a neuromorphic device. In addition, the uniformity and the reliability of the device such as endurance cycling and retention time were investigated.