KAIST develops 'chameleon' AI chip with programmable response speeds
A new programmable dynamic memtransistor allows hardware to adapt its timing to match real-time data, slashing prediction errors by up to 40-fold.
Researchers at the Korea Advanced Institute of Science and Technology (KAIST) have developed a programmable dynamic memtransistor (PDM) capable of adjusting its own response speed to process time-varying data. Led by Chair Professor Shinhyun Choi, the team created a "chameleon AI semiconductor" that retains multiple timing states, allowing hardware to adapt to the speed of incoming information rather than relying on rigid, pre-set speeds.
The PDM achieves this flexibility through a specialized dual-layer architecture. The device utilizes a charge storage layer dedicated to data processing and a separate electron trapping layer that controls the response speed nonvolatily. In experimental trials, the team demonstrated the ability to tune the current recovery time over a five-fold range and the characteristic frequency over a ten-fold range. This adaptability resulted in a dramatic performance boost, reducing prediction errors for time-varying data by up to 40-fold compared to conventional fixed-response semiconductor devices.
The hardware bottleneck
Most current AI hardware is designed with fixed response speeds determined during fabrication. To handle time-series data—such as the varying speed of a person's handwriting or the movement of an object—systems typically rely on complex software layers to analyze and compensate for these timing differences. This software-heavy approach creates significant computational overhead, leading to high power consumption and latency that limits the efficiency of real-time processing.
Implications for edge computing
By shifting temporal processing from software into the hardware itself, the PDM enables high-accuracy, low-energy AI processing in real time. This shift is critical for edge devices, including autonomous vehicles, robotics, and wearables, where rapid responses to dynamic environments must be balanced with strict power constraints. Because the PDM is compatible with commercial semiconductor materials, the researchers suggest a viable path toward mass production.
Next steps
The findings were published in the journal Nature Communications on July 4, 2026, under the title "Programmable memtransistor array with temporal dynamics modulation for efficient time-series data processing." As the industry moves toward more autonomous systems, the focus will likely shift toward integrating these programmable arrays into larger-scale commercial architectures to verify if the 40-fold error reduction holds across more complex, real-world datasets.