Researchers at the Korea Advanced Institute of Science and Technology (KAIST) have developed a programmable AI semiconductor that can adjust how it responds to data changing at different speeds.
The research team, led by Chair Professor Shinhyun Choi, developed a programmable dynamic memtransistor (PDM) capable of changing and retaining its response characteristics across multiple states.
Tests showed that the new device reduced prediction errors by up to 40 times when processing data containing both fast and slow changes. The technology could support more efficient real-time AI processing in applications including autonomous vehicles, robots, and wearable devices.
KAIST researchers also fabricated an integrated PDM array and used it to predict complex data. The system achieved accuracy comparable with conventional software-based approaches while using significantly less energy, demonstrating the potential of programmable semiconductor hardware for AI applications.
Making AI semiconductor hardware programmable
Computers and smartphones currently rely on complex software processing to analyse data that changes over time. This can create substantial computational demands and increase power consumption.
Researchers have been exploring ways for semiconductor hardware to process data directly. However, conventional devices generally have fixed response speeds that cannot be changed after fabrication.
The KAIST team addressed this limitation by developing a dual-layer structure inside the transistor. The design combines a charge storage layer that accumulates and processes data with an electron-trapping layer that controls the device’s response speed.
This allows the PDM to adjust its response characteristics according to incoming data and retain those settings.
Response characteristics can be adjusted
The PDM processes incoming data through its charge storage layer, while the electron-trapping layer controls the recovery speed of the semiconductor across multiple levels.
During testing, researchers were able to adjust the device’s current recovery time across an approximately five-fold range. Its characteristic frequency could also be controlled across a range of more than 10-fold.
This ability to alter the response characteristics is particularly important when data contains changes occurring at different speeds.
In experiments involving data with fast and slow changes occurring together, the PDM reduced prediction errors by as much as 40 times compared with conventional semiconductor devices with fixed response characteristics.
The technology was also able to process information accurately when the speed of handwriting or object movement varied. By configuring the device’s response characteristics to match different input timescales, the researchers could improve how the incoming information was processed.
Retaining settings without continuous power
The PDM has another important feature: once its response characteristics have been configured, it can retain them without requiring a continuous external power supply.
The device also does not require complex preprocessing of incoming data. This could help simplify the processing of changing information while reducing the energy required for AI operations.
The researchers demonstrated the approach by fabricating a PDM array and using it to predict complex data. Its performance reached accuracy comparable with conventional software-based systems while consuming considerably less energy.
Potential for AI applications
The researchers say the technology could have applications in devices that need to process changing information efficiently, including autonomous vehicles, robots, and wearable devices.
The PDM is also compatible with materials used in widely adopted commercial semiconductor manufacturing processes. This compatibility could support future efforts to scale the technology towards mass production and commercialisation.
The development represents a step towards an AI semiconductor whose response characteristics can be programmed according to the speed of incoming data.
Rather than relying on a fixed response, the PDM can be configured across multiple states and retain its settings. This combination of programmability, information storage and data processing could help improve AI performance while reducing energy consumption in future devices.
