A new artificial intelligence (AI) semiconductor device that can adjust its response characteristics according to the rate of change in input data has been developed. In time-series data prediction that mixes fast and slow information, it reduced errors to one-fortieth of those of existing devices.
A research team led by Chair Professor Choi Sin-hyun of the School of Electrical Engineering and the Graduate School of Semiconductor Engineering at KAIST said on the 7th that, together with Samsung Electronics Semiconductor Research Institute, it developed a programmable dynamic memtransistor (PDM) and an integrated array of it. The results were published in the July issue of the international journal Nature Communications.
An AI Semiconductor is a semiconductor designed to efficiently handle AI computation. The memtransistor used by the team combines, in a single device, the functions of memory that stores information and a transistor that performs computation. Connecting multiple of these can be used as an AI semiconductor circuit that remembers the previous input while computing the next data.
The PDM developed this time can adjust how long it retains information and how fast it responds, in line with how quickly the input data changes. Existing devices had fixed speeds for responding to input signals and returning to their original state, but the PDM can change these characteristics depending on the processing target.
The team placed, inside a single transistor, a charge storage layer that stores electrical information and a layer that traps electrons. The charge storage layer processes input information, and the electron-trapping layer controls the time it takes the device to return to its original state.
In experiments, the team succeeded in adjusting, over about a fivefold range, the time it takes the device to return to its original state after data input. It also confirmed that the frequency range for processing repetitive signals can be varied by more than tenfold.
In time-series data prediction that includes information with different rates of change, prediction errors decreased to as low as one-fortieth compared with conventional semiconductors that have fixed response characteristics. In array experiments that consolidation multiple PDMs, it showed accuracy similar to software-based AI systems while demonstrating the potential to reduce energy use.
A PDM retains its set response characteristics even after power is turned off. It can process inputs at various speeds without separate, complex data pre-processing, and it is also compatible with conventional complementary metal-oxide semiconductor (CMOS) fabrication processes.
Professor Choi Sin-hyun said, "It is meaningful that the response of a semiconductor can be adjusted according to the rate of change in data," adding, "It could be applied to fields that require real-time data processing, such as Autonomous Driving cars and robots, and wearable devices."
References
Nature Communications (2026), DOI: https://doi.org/10.1038/s41467-026-75211-5