Light-emitting diodes (LEDs), renowned for their energy efficiency, are opening up more application possibilities beyond lighting and displays. On November 25, the Technical University of Braunschweig (TU Braunschweig) announced that a research team at its Nitride Technology Center (NTC) is using Micro LEDs to build AI neural networks, aiming to develop more powerful and energy-efficient computer systems for future artificial intelligence.
LED-based neuromorphic computer demonstration device. (Image credit: TU Braunschweig)
Miniaturization, scalability, and energy efficiency of hardware are key to developing more powerful AI applications. A research team at the Technical University of Brunswick (NTC) has adopted a novel approach, using Micro LED technology to construct computer systems. By miniaturizing and expanding applications through Micro LEDs, they are creating neural network computers.
The findings, published in the Journal of Physics Photonics by the Technical University of Braunschweig, Ostafaria University of Applied Sciences, and AMD Osram, explain how this computer can elevate AI applications to a higher level.
The research team stated that optical neuromorphic computing mimics the workings of biological neural networks (such as the human brain) and is implemented using electronic circuits or photonic components. This method avoids the enormous energy demands of traditional computer technologies in AI applications. It is predicted that with the development of artificial intelligence technology, approximately one-third of the world's electricity will be used for supercomputer operation and cooling within the next 10 years.
The research team combined GaN components with traditional silicon microelectronics technology to create a highly integrated array with hundreds of thousands of Micro LEDs. GaN-based Micro LED technology has enormous potential to reduce the massive energy consumption of AI systems, potentially reducing energy consumption by up to 10,000 times.
These Micro LEDs can perform tasks that were originally done by silicon crystals. By combining parallel memory processing with efficient photon production and detection, they can create hardware that can physically map different levels of neural networks and enable parallel information flow.
Currently, this research is still in its early stages. The NTC research team has developed a macroscopic optical Micro LED demonstration device containing 1,000 neurons. This device has passed standard AI pattern recognition tests, demonstrating its ability to recognize digital content written in a scrambled manner. (Translated by Irving)
