IMX Talks - A magnonic matrix device for parallel convolutional processing

The growing computational demands of artificial intelligence are exposing fundamental limits of electronic and photonic accelerators, spurring a search for alternative physical computing hardware capable of efficient matrix-vector multiplication (MVM). Magnonic systems, which encode information in propagating spin waves, offer low-power signal transport and wavelengths well suited to dense on-chip integration, yet their potential for parallel computing remains largely unexplored. Here we present magnonic matrix devices, built from low-loss yttrium iron garnet (YIG) waveguide networks, that exploit spin-wave transport and interference to perform real-time, parallel MVM and convolutional operations. These devices achieve high area efficiencies, comparable to state-of-the-art electronic physical computing platforms, alongside competitive power efficiencies. Integrated into convolutional neural networks, they perform benchmark tasks including handwritten digit recognition (97.1% accuracy), MRI image classification (83.2%), and Parkinsonian gait detection (90.6%). Weights can be tuned globally through microwave frequency and bias magnetic field, enabling network-wide reconfiguration for task switching, while local optical and electrical control is demonstrated as a proof of concept, pointing toward spatially selective weight updates for future on-chip training. Together, these results establish magnonic convolutional processing as a scalable platform for efficient AI acceleration and open a route toward compact, wave-based processors that intrinsically support parallel computation.

Bio: Sebastiaan van Dijken is a professor in the Department of Applied Physics at Aalto University in Finland, where he leads the Nanomagnetism and Spintronics group. His research focuses on spin-wave dynamics and magnon transport, with the goal of exploiting magnons for energy-efficient signal processing and data transfer in next-generation computing systems. By designing and fabricating engineered magnetic thin films and nanostructures, his work advances both fundamental knowledge of spin transport and magnetization dynamics and enables practical implementations in magnonic circuits and spintronic devices.
His group also investigates complementary directions, including electric-field control of magnetism and magnetoplasmonics. More recently, his research has expanded into neuromorphic and in-sensor computing, exploring how magnetic and magnonic systems, as well as photomemristor networks, can perform tasks such as pattern recognition and adaptive learning, offering new architectures inspired by biological neural systems.