Scientists from the École Polytechnique Fédérale de Lausanne (EPFL) have developed a new framework for optical computing that significantly reduces the energy consumption of artificial intelligence systems. This discovery has the potential to revolutionize the industry and make artificial intelligence more energy-efficient.

Modern AI systems, such as deep neural networks, require substantial amounts of energy for training and deployment. Some estimates suggest that if the current growth rate of energy consumption continues, by 2027, the annual energy consumption of servers for AI models could exceed that of a small country.

EPFL researchers have proposed a new approach to optical computing, which uses photons to process data. This allows computations to be performed much faster and more efficiently than traditional electronic systems. However, until now, optical systems have been unable to perform the nonlinear transformations necessary for data classification in neural networks.

The EPFL scientists have developed a simple solution that enables nonlinear transformations to be performed optically. They spatially encode image pixels on the surface of a low-power laser beam, allowing for the nonlinear multiplication of pixels. This solution requires eight orders of magnitude less energy compared to traditional electronic systems.

"Our method is scalable and 1,000 times more energy-efficient than the most advanced deep digital networks," says Demetri Psaltis, head of EPFL's optics laboratory.

The research, supported by a Sinergia grant from the Swiss National Science Foundation, was published in the journal *Nature Photonics*. The scientists are already working on developing a compiler to translate digital data into code that can be used by optical systems.

This breakthrough has the potential to transform the industry and make artificial intelligence more energy-efficient. However, further engineering research is needed to achieve scalability.