Optical method runs AI tensor operations at the speed of light

Researchers at Aalto University have developed a novel optical computing method called “single-shot tensor computing” that performs complex tensor operations—fundamental to AI tasks—at the speed of light. Unlike conventional GPUs, which face limitations in speed, power efficiency, and scalability due to increasing data demands, this photonic system encodes digital information into the amplitude and phase of light waves. These waves interact naturally during propagation, executing mathematical operations such as convolutions and attention layers without requiring electronic switching or external control circuitry, thereby reducing power consumption and simplifying integration. The team further enhanced the system by employing multiple wavelengths of light, each acting as an independent computational channel, enabling parallel processing of higher-order tensor operations. This multi-wavelength approach allows the system to perform complex AI computations in a single optical pass, akin to merging multiple processing steps into one. The method is versatile across various optical platforms and is planned for integration onto photonic chips to create low-power, light-based AI processors. The researchers anticipate commercial deployment within