A next-generation AI memory device developed through a collaboration between Professor Min-Kyu Song of the Korea University School of Electrical Engineering and researchers at MIT has been published in Science Advances on December 5 (U.S. local time). (Related link: https://www.science.org/doi/10.1126/sciadv.adz2553)
 

Unlike conventional memories that store data using electrical charge (voltage) or resistance states, the newly demonstrated device encodes information using the dielectric constant (capacitance) of an ultrathin film. Because it enables read operations without significant current flow, the device is expected to reduce energy consumption by several thousand times compared to existing memory technologies.
 

The research team successfully fabricated a ferroelectric capacitor (FeCAP) using single-crystal barium titanate (BaTiO₃) in a manner fully compatible with standard silicon semiconductor processes—the first demonstration of its kind. This achievement establishes a technological foundation for the practical realization of dielectric-based memory. The BaTiO₃ single-crystal FeCAP operates at an electric field 100 times lower than that required by existing dielectric-based memory devices, yet still delivers a record-high memory window, which distinguishes the different memory states.
 

Additionally, by separating the BaTiO₃ single-crystal film into an ultrathin freestanding membrane and directly stacking it onto a silicon chip, the team demonstrated the feasibility of monolithic 3D integration. This approach goes beyond traditional wafer-level or package-level stacking methods such as HBM, opening a path toward multi-layered, high-performance memory arrays on a single wafer—a key enabling technology for future ultra-dense, high-performance AI processors.
 

Professor Song stated, “Our work addresses the longstanding challenge of achieving a sufficiently large memory window in ultra-low-power dielectric memories by combining single-crystal materials with innovative processing techniques. The fact that this platform is compatible with silicon semiconductor processes makes the achievement even more significant. We aim to further develop this into an ultra-low-power, high-density memory platform essential for next-generation AI computing.”
 

This research was supported by the Intelligence Advanced Research Projects Activity (IARPA) in the United States and the National Research Foundation of Korea (NRF), with participation from MIT, Georgia Tech, Cornell University, Texas A&M University, Korea University, Seoul National University, and POSTECH.