Dresden-Rossendorf: New platform for magnet-based AI hardware is created

Together, the projects will receive around 3.75 million euros in funding from the EU and the Federal State of Saxony. The aim is to lay the foundations for new, particularly energy-efficient hardware for artificial intelligence - a technology that should function more sustainably in the future than current AI technology.

With the new infrastructure, a competence center for magnetic AI hardware is growing in Dresden-Rossendorf. The funding from the EU, federal and state governments now totals around ten million euros - a strong signal for the future viability of this research. "This makes the HZDR a central location for magnetic AI in Germany," says the Scientific Director of the HZDR, Prof. Sebastian M. Schmidt. "A platform is being created here that brings together basic research and industrial application - exactly what we need for the next generation of AI technologies."

The latest advances in language models, translation software and image generators have a catch: they require enormous amounts of energy. "Training AI models such as ChatGPT alone consumes gigawatt hours," says Dr. Helmut Schultheiß, head of the Spin Interaction and Control working group at the HZDR. "And each new model requires even more computing power and therefore even more electricity." As a result, the energy consumption of AI systems threatens to skyrocket - a problem not only for further development, but also for the global climate. To improve the situation, the HZDR is focusing on a disruptive approach - AI hardware based on magnetic components that is designed to function more sustainably.

Today's computer chips are based on billions of transistors that switch electrical currents and continuously generate heat in the process - which makes them real power guzzlers for AI applications. Magnetic components work differently: they use collective oscillations of electron spins, known as magnons. These do not transport electrical charges, but energy and information via magnetic interactions. As a result, hardly any current flows and only little energy is lost in the form of heat.

• Making sense of the puzzle chaos

Instead of changing individual transistor states, complex wave patterns spread out in the magnetic materials, which can process information directly. "You can think of it like a jigsaw puzzle," explains Schultheiß. "The magnetic AI pre-sorts the pieces before classical logic takes over. This saves a lot of calculation steps - and therefore a lot of energy." The approach builds on the successes of the EU project NIMFEIA. Here, the HZDR has already been able to show that such magnetic elements can be seamlessly integrated into the chip production of partners such as GlobalFoundries and Infineon.

With the tenders it has won, the center can continue to drive development forward. The MagKI investment programme enables the acquisition of four specialized equipment platforms for the production and characterization of magnetic nanostructures. A system is being built in the institute's clean room that can be used to structure magnetic materials in a targeted manner. Researchers can use focused ion beams to create tiny nanodefects - a key technology for precisely adjusting the properties of memory cells. In addition, a femtosecond laser is combined with a superconducting magnet. This allows magnetic oscillations to be examined in the terahertz range, i.e. at previously unattainable speeds. The measuring station is intended to show how magnons behave on extremely short time scales - a prerequisite for magnetic AI that learns and reacts in real time.

• Linking with quantum technology

A quantum magnetometry platform is also being acquired, in which highly sensitive probes record magnetic fields on an atomic scale. This allows magnetic memories to be coupled with quantum technology, for example for high-precision quantum sensors. Finally, a nano-ellipsometer will be able to record the light reflections of individual nanostructures. This will make it possible to measure the optical properties of magnetic and quantum-based systems with spatial resolution for the first time. The data will help to characterize and optimize new materials.

Magnons and quantum phenomena complement each other perfectly. This could make it possible to read out magnetic states optically or quantum-based in the future - a huge technological step forward.

• Competence center for magnetic AI

The second, closely linked project "Magnon4KI" bundles the research activities. It investigates how magnetic and quantum-based processes can be combined to create new types of AI components. Magnon4KI provides the scientific foundation for MagKI. The project develops the concepts, methods and prototypes that will later be tested on the new device platform.

"Both projects are closely interlinked, cross-departmental and part of the HZDR 2030+ strategy, which focuses on sustainable information and quantum technologies," emphasizes Lindner. "Our goal is to secure Saxony a leading international position in the field of magnetic AI in the long term." This is why partners from industry have been involved from the outset: GlobalFoundries is already supplying wafers and chips that are being tested in Dresden-Rossendorf. Infineon wants to pre-characterize its new silicon quantum chips at the HZDR in the future. Bosch is also a partner in joint experiments.