Daegu Gyeongbuk Institute of Science and Technology (DGIST) Department of Electrical Engineering, Electronics and Computer Engineering Professor Choi Sang-hyeon's team said on the 28th that it succeeded in integrating "memristors," cited as a core technology for next-generation artificial intelligence (AI) semiconductors, at the wafer scale. It is being assessed as laying the technical groundwork to move beyond the structural limits of conventional semiconductors toward "brain-inspired semiconductors" that can process vast amounts of information efficiently like the human brain.
The human brain consists of about 100 billion neurons and 100 trillion synapses. It stores and processes an enormous amount of information simultaneously within a small space. Semiconductors that operate in this "brain-like" manner, that is, brain-like AI chips, are regarded as the ultimate goal of next-generation artificial intelligence technology.
However, current AI Semiconductors struggle to match the efficiency of the brain due to complex circuits and high power consumption. The memristor has emerged as an alternative to overcome these limitations.
A memristor is a semiconductor device that remembers how much current has flowed. Because it can perform memory and computing simultaneously, it is much faster and more efficient than conventional semiconductors. Its structure is simple, allowing more circuits to be integrated into the same area, and particularly, when arranged in a "crossbar" pattern where horizontal and vertical lines intersect like a grid, it can store tens of times more information than existing memory.
The problem was that mass production was not easy. Memristor technology to date has remained at a small-scale experimental level. Due to complex processes, low Production yield (completion rate), and issues of current leakage and voltage loss, it was difficult to scale up to large-area wafer units.
To solve these technical limitations, the research team collaborated with Professor Dmitri Strukov's group at UC Santa Barbara. The two teams introduced a new approach that co-designs the entire process, from the materials used to make semiconductors, to the devices (small components), circuit design, and operating algorithms.
As a result, they succeeded in uniformly implementing memristor circuits across the entire surface of a 10-centimeter (4-inch) diameter wafer without going through complex manufacturing steps. The completion rate (Production yield) exceeded 95%, far higher than existing technologies.
In addition, the team succeeded in fabricating a three-dimensional (3D) stacked structure by layering memristors. By stacking semiconductors vertically, they opened the way to process more computations at once. This is an important achievement showing that memristor-based semiconductors can be expanded into large-scale artificial intelligence systems in the future.
The team applied this technology to actual artificial intelligence computation and conducted experiments. Applying a spiking neural network (SNN) approach that mimics the operating principles of neurons in the human brain, they confirmed high computational efficiency and stable performance while significantly reducing power consumption compared with existing methods.
Choi said, "This study advances previously limited memristor integration technology by a step," adding, "It could lead to the development of next-generation AI Semiconductor platforms at the level of the human brain."
The results were published on the 1st in the international journal Nature Communications.
References
Nat Commun (2025), DOI: https://doi.org/10.1038/s41467-025-63831-2