Stacked AI Chip Processes Temporal Data With Fast and Slow Transistor Layers

Researchers from the Korea Advanced Institute of Science and Technology, in collaboration with UNIST and POSTECH, have developed a stacked artificial intelligence chip capable of recognizing temporal changes in motion. Published in Advanced Materials, the technology employs a monolithic 3D-integrated architecture based on solid thin-film carbon nanotube ion-gated transistors to process information across multiple timescales within a single hardware platform. The device addresses the challenge of distinguishing immediate actions from accumulated movements, such as differentiating a brief arm swing from sustained walking. Led by Professor Jimin Kwon, the team engineered a solid-state ionogel by embedding ions within a polymer matrix. This solid thin-film approach overcomes the fabrication limitations associated with liquid or soft gel materials, enabling precise vertical stacking. By adjusting ionic content and film thickness, researchers tuned transistor response speeds. Stacking layers with distinct kinetics allows the chip to simultaneously detect rapid signals and retain memory of earlier inputs through ionic relaxation dynamics. Electrical characterization demonstrates strong gate control, high current drive at low voltages, and operational stability over 55 months. The 3D reservoir architecture successfully distinguished 16 unique patterns derived from four-bit temporal sequences. In simulations using the moving MNIST dataset of sequential handwritten digits, the system achieved validation accuracies exceeding 90% across varying playback speeds, confirming robust classification of time-dependent data. Fabrication utilizes existing thin-film semiconductor processes, enabling production on 4-inch wafers and flexible substrates. This scalability supports integration into compact, low-power wearable devices for on-device analysis of physiological signals and movement patterns. Professor Kwon noted that the technology offers a viable pathway to mass-produce AI chips capable of efficient temporal information processing for applications such as smartwatches.