Creating smarter AI with smarter semiconductors

A research team led by Professor Seyoung Kim from the Departments of Materials Science and Engineering and Semiconductor Engineering at POSTECH, in collaboration with alumnus Kyungmi Noh and PhD student Hyunjeong Kwak from the same department, along with Professor Hyung-Min Lee from Korea University's School of Electrical Engineering, has recently made significant strides in the field of artificial intelligence (AI) hardware. They have demonstrated that analog hardware, utilising Electrochemical Random Access Memory (ECRAM) devices, can greatly enhance AI computational performance, paving the way for its potential commercialisation. Their findings were published in the international journal "Science Advances."

As AI technology, including applications such as generative AI, advances rapidly, existing digital hardware like CPUs, GPUs, and ASICs have reached their scalability limits. This has led to growing interest in developing analog hardware tailored for AI computation. Analog hardware processes AI computations in parallel by adjusting the resistance of semiconductors based on external voltage or current. It also utilises a cross-point array structure with vertically intersecting memory devices. Although this approach offers certain advantages over digital hardware, particularly for specific computational tasks and continuous data processing, it still faces challenges in meeting the diverse requirements for computational learning and inference.

To overcome these limitations, the research team focused on ECRAM, a type of memory that controls electrical conductivity through the movement and concentration of ions. Unlike traditional semiconductor memory, ECRAM features a three-terminal structure that separates the paths for reading and writing data, allowing for operation at relatively low power levels.

In their study, the team successfully fabricated ECRAM devices using a three-terminal semiconductor structure arranged in a 64×64 array. Their experiments showed that these devices exhibited excellent electrical and switching characteristics, along with high yield and uniformity. The team also implemented the Tiki-Taka algorithm, an advanced analog-based learning algorithm, on this high-yield hardware, which led to the maximisation of AI neural network training accuracy. Importantly, the researchers demonstrated how the "weight retention" property of the hardware influences learning and confirmed that their method does not overload artificial neural networks, underscoring its commercial potential.

This research is particularly noteworthy because, until now, the largest reported array of ECRAM devices capable of storing and processing analog signals was 10×10. The team’s success in scaling up to a 64×64 array, while maintaining varied characteristics across devices, marks a significant advancement in the field.

Professor Seyoung Kim of POSTECH commented: “By developing large-scale arrays based on novel memory device technologies and creating AI algorithms specifically for analog hardware, we have uncovered the potential for AI computational performance and energy efficiency that significantly exceeds that of current digital methods.”

The research was supported by the Ministry of Trade, Industry and Energy, the Public-Private Partnership for Semiconductor Talent Training Program, which is backed by the Korea Planning & Evaluation Institute of Industrial Technology (KEIT) and the Korea Semiconductor Industry Association, as well as EDA Tool of the IDEC.