Implementation of convolutional neural network and 8-bit reservoir computing in CMOS compatible VRRAM. (15th December 2022)

• Record Type: Journal Article
• Authors: Park, Jongmin; Kim, Tae-Hyeon; Kwon, Osung; Ismail, Muhammad; Mahata, Chandreswar; Kim, Yoon; Kim, Sangbum; Kim, Sungjun
• Is Part Of: Nano energy. Volume 104(2022)Part B
• Publication Date: 2022-12-15
• Subjects: VRRAM -- Resistive switching -- CNN -- Reservoir computing; Nanoscience -- Periodicals; Nanotechnology -- Periodicals; Nanostructured materials -- Periodicals; Power resources -- Technological innovations -- Periodicals; Nanoscience; Nanostructured materials; Nanotechnology; Power resources -- Technological innovations; Periodicals; 621.042
• DOI: 10.1016/j.nanoen.2022.107886
• Languages: English
• ISSNs: 2211-2855

Abstract

Abstract: We developed W/HfO2 /TiN vertical resistive random-access memory (VRRAM) for neuromorphic computing. First, basic electrical properties, such as current–voltage curves, retention, and endurance, were determined. To examine the conduction mechanism, a device with a large switching area was fabricated, and its current level and that of the VRRAM were compared. Moreover, we analyzed the current behavior relative to the ambient temperature. Subsequently, the number of states upon potentiation and depression was linearly converted via conductance modulation due to an applied pulse. The practicality of the device was assessed using a convolutional neural network. Finally, 16-state reservoir computing was combined with multilevel characteristics to implement 8-bit reservoir computing with 256 states. We verified that in terms of time and power consumption, 8-bit reservoir computing is more efficient than 4-bit reservoir computing. Hence, we concluded that the W/HfO2 /TiN VRRAM cell is a promising volatile memory device. Graphical Abstract: ga1 Highlights: 3-dimensional VRRAM structure was fabricated for high-density synapse. High-performance memory with low-power and self-rectifying characteristics is implemented. 99.15 % accuracy for MNIST is achieved in CNN. Short-term memory characteristics are demonstrated. Reservoir computing with 256 states was demonstrated for more energy efficiency.