A research team led by Professor Seyoung Kim from POSTECH and Professor Hyung-Min Lee from Korea University has showcased the potential of analog hardware using Electrochemical Random Access Memory (ECRAM) devices for maximizing the computational performance of artificial intelligence. Their study, published in Science Advances, addresses the limitations of existing digital hardware in handling the scalability of AI technology, particularly in applications like generative AI. Analog hardware, which adjusts semiconductor resistance based on external voltage or current, shows promise for specialized AI computation due to its ability to process data in parallel and continuous processing. However, meeting the diverse requirements for computational learning and inference remains a challenge.
To overcome the limitations of analog hardware memory devices, the research team focused on ECRAM devices, which control electrical conductivity through ion movement and concentration, and feature a three-terminal structure with separate reading and writing paths. The team successfully fabricated ECRAM devices in a 64×64 array using three-terminal-based semiconductors, demonstrating excellent electrical and switching characteristics, high yield, and uniformity. The team applied the Tiki-Taka algorithm, an analog-based learning algorithm, to the hardware, maximizing the accuracy of AI neural network training computations. They also highlighted the impact of the “weight retention” property of hardware training on learning and confirmed that their technique does not overload artificial neural networks.
The research is significant as it represents the largest array of ECRAM devices for storing and processing analog signals reported in the literature to date. The successful implementation of these devices on a large scale with varied characteristics for each device showcases the potential of analog hardware for AI computation. Professor Seyoung Kim emphasized the importance of novel memory device technologies and analog-specific AI algorithms in realizing the potential for AI computational performance and energy efficiency that surpass current digital methods. The research received support from the Ministry of Trade, Industry, and Energy, the Public-Private Partnership for Semiconductor Talent Training Program, and the Korea Semiconductor Industry Association.
In conclusion, the research team’s demonstration of analog hardware using ECRAM devices for maximizing AI computational performance offers promising prospects for the commercialization of this technology. By overcoming the limitations of existing digital hardware and showcasing the potential for improved energy efficiency and computational performance, the team’s work contributes significantly to the field of AI hardware development. The successful fabrication of ECRAM devices in a 64×64 array and the implementation of cutting-edge analog-based learning algorithms highlight the capabilities of analog hardware in handling complex AI computations. The support from various institutions further underscores the importance of this research in advancing AI technology.
