Micro-LEDs allow faster artificial intelligence

04 Dec 2024

TU Braunschweig miniaturized light sources will enable tomorrow's optical processing systems.

Micro-LED light sources have been a subject of research for developers of miniaturized micro displays for some time, exploiting in particular their suitability for creating precise optical patterns at high modulation rates.

A project at the Technical University of Braunschweig (TU Braunschweig) has now shown how the same kind of light sources could have uses beyond display technology, by demonstrating applications in neuromorphic computing and artificial intelligence.

Published in Journal of Physics: Photonics, the findings point to a role for micro-LEDs in making future computers more powerful and energy-efficient, avoiding the problem of energy consumption from massively parallel information processing that currently hangs over AI development.

"Our optical neuromorphic computing mimics the functioning of biological neural networks, such as those in the human brain, by using electronic circuits or photonic components," commented Andreas Waag from TU Braunschweig Institute of Semiconductor Technology.

The research, like the study of micro-LEDs for displays, exploits the emission properties of gallium nitride (GaN) semiconductor materials, increasingly attractive for uses in power electronics because they offer higher power density and better efficiency than traditional non-optical silicon semiconductors.

At TU Braunschweig's Nitride Technology Centre (NTC), the combination of GaN components with conventional silicon microelectronics is seen as a route to completely new applications, and forms one theme of work in the region's QuantumFrontiers cluster of excellence.

Avoiding immense energy demands

In trials the project assembled InGaN-based micro-LEDs into benchtop monolithic arrays from single continuous epitaxially grown semiconductor layers. To address each pixel individually, a microLED array was bonded to CMOS-based driver circuitry, connecting each microLED pixel to its respective current driver.

These prototype segmented LED chips were directly structured on their sapphire growth substrates and then attached to a CMOS chip containing the driver circuits. Modelling the computing power of a system based on these emitters in terms of tera-operations per watt, a metric for AI systems, indicated significant potential improvements over conventional approaches, said the team.

The NTC research group has already developed a macroscopic optical micro-LED demonstrator equivalent to a 1,000 neuron network. The demonstrator passed a standard AI pattern recognition test, identifying numbers from zero to nine written in a jumbled fashion, some of which are difficult for a human to decipher.

Demonstrating the advantages of this approach using a prototype assembled from currently available off-the-shelf components is a good indication of the advantages of micro-LEDs, according to the TU Braunschweig project, although significant future research is now needed into optimized electrical circuitry and suitable miniaturized lenses and filters.

If these challenges are met, then a micro-LED approach "avoids the weaknesses of conventional digital computer technology, which lead to immense energy demands in massively parallel information processing for AI applications," said Christian Werner from project partner Ostfalia University of Applied Sciences. "It is expected that in 10 years' time around a third of the world’s electrical energy will be used for supercomputers and their cooling."