삼중층 금속 산화물 벌크 RRAM 소자를 이용한 뉴로모픽 엣지 컴퓨팅 연구 (Bulk Switching Trilayer Metal Oxide RRAM for Neuromorphic Computing at the Edge)

서지정보

· 발행기관 : 한국세라믹학회
· 수록지 정보 : 세라미스트 / 28권 / 1호 / 106 ~ 116페이지
· 저자명 : 박재성

초록

The Complementary metal oxide semiconductor-Resistive random access memory (RRAM) integration presents significant potential for energy-efficient and high-speed neuromorphic computing. However, conventional filamentary RRAM technologies that rely on filamentary switching face challenges such as high variability, noise, reduced computational accuracy, and increased energy consumption. To address these limitations, we developed a filament-free, bulk-switching RRAM technology. By designing a trilayer metal-oxide stack, we optimized switching behavior across different oxide thicknesses and oxygen vacancy distributions, enabling stable bulk switching without filament formation. Our approach achieves reliable switching in the MΩ regime with high current nonlinearity and supports up to 100 distinct levels without requiring a compliance current. Additionally, we built a neuromorphic compute-in-memory platform and demonstrated its potential for edge computing by deploying a spiking neural network for an autonomous navigation and racing task. Our work addresses challenges posed by existing RRAM technologies and paves the way for neuromorphic computing at the edge under strict size, weight, and power constraints.

영어초록

The Complementary metal oxide semiconductor-Resistive random access memory (RRAM) integration presents significant potential for energy-efficient and high-speed neuromorphic computing. However, conventional filamentary RRAM technologies that rely on filamentary switching face challenges such as high variability, noise, reduced computational accuracy, and increased energy consumption. To address these limitations, we developed a filament-free, bulk-switching RRAM technology. By designing a trilayer metal-oxide stack, we optimized switching behavior across different oxide thicknesses and oxygen vacancy distributions, enabling stable bulk switching without filament formation. Our approach achieves reliable switching in the MΩ regime with high current nonlinearity and supports up to 100 distinct levels without requiring a compliance current. Additionally, we built a neuromorphic compute-in-memory platform and demonstrated its potential for edge computing by deploying a spiking neural network for an autonomous navigation and racing task. Our work addresses challenges posed by existing RRAM technologies and paves the way for neuromorphic computing at the edge under strict size, weight, and power constraints.

참고자료

· 없음