KAIST develops AI semiconductor that adapts to data speed

Researchers have developed an AI semiconductor device that adjusts its response characteristics to match the rate at which incoming data changes.

KAIST announced Friday that a research team led by Chair Professor Choi Shin-hyun of the School of Electrical Engineering and Graduate School of Semiconductor Engineering has developed an AI semiconductor device called the programmable dynamic memtransistor, or PDM, along with an integrated array built from the device.

A memtransistor is a next-generation semiconductor device that combines the functions of a memory chip — which stores information — and a transistor, which performs calculations. Because it can retain previous information while processing new data, it can adjust its response speed to match incoming signals.

Data in the real world does not change at a uniform rate. Some information, such as heartbeats or speech, fluctuates rapidly, while other data — factory equipment status or weather readings, for example — changes slowly. Conventional AI semiconductors, however, have fixed response speeds and retention times, limiting their ability to handle data streams that vary at different rates simultaneously.

The research team addressed this by integrating two distinct layers inside a single transistor: a charge storage layer that processes electrical information, and an electron trap layer that regulates how quickly the semiconductor returns to its baseline state.

When data enters the PDM, the charge storage layer handles the processing while the electron trap layer controls the rate at which the device resets. Much like a driver adjusting speed between a highway and a narrow alley, the semiconductor autonomously selects the most appropriate response speed for the pace at which the data is changing.

The core advance is the ability to tune the device's temporal response characteristics based on data speed, without requiring separate or complex data preprocessing, allowing it to efficiently handle signals that vary across a wide range of rates.

In experiments, the team demonstrated that the time for the semiconductor to return to its baseline state after responding to an input signal could be tuned across a roughly fivefold range. Frequency response — the device's ability to process rapidly repeating signals — could be adjusted by more than tenfold.

In time-series prediction tests involving data that mixed fast- and slow-changing signals, the PDM reduced prediction errors by up to 40 times compared with conventional fixed-response semiconductors.

The team also fabricated an array integrating multiple PDM devices. Testing showed the array operated with far less energy than software-based AI systems while maintaining comparable accuracy.

The technology is non-volatile, meaning the programmed response characteristics are retained without a continuous power supply. It is also compatible with CMOS, the widely used commercial semiconductor fabrication process, which the researchers said significantly raises the prospects for mass production and commercialization.

"This research realizes an AI semiconductor that autonomously responds in the most appropriate way based on the speed at which data changes," Choi said. "We expect it to become a core technology for improving performance and reducing power consumption across a wide range of AI devices, including self-driving cars, robots and wearables."

Kim Dae-won, a doctoral candidate at KAIST's Graduate School of Semiconductor Engineering, served as lead author. Co-authors include Jo Yun-ho, Seo Seok-ho, Kim Yu-jin, Park Si-on, Jang Tae-hwan and Park Chae-bin. Researchers Oh Young-taek and Fellow Lee Jae-deok of Samsung Electronics' Semiconductor Research Center also contributed as co-authors, with Choi serving as corresponding author.

The findings were published in Nature Communications in July. The research was supported by the Ministry of Science and ICT, the Institute for Information and Communications Technology Planning and Evaluation, the Ministry of Trade, Industry and Energy, and Samsung Electronics, among others.