Chips that integrate both light and electricity have demonstrated improved computational efficiency along with decreased power usage when contrasted with traditional electronic chips. These photonic computing chips were detailed in two separate studies. Nature This week, we may tackle the increasing computational requirements spurred by the progression of artificial intelligence technologies.
As artificial intelligence and deep learning models become more complex, they strain traditional electronic computing systems and require growing amounts of energy. Computing with light particles instead of electrons—known as photonic computing—offers a promising alternative to address these issues.
Multiplication and accumulation—which are key computational tasks in artificial intelligence—can be executed more swiftly and efficiently through the use of photonic circuits. However, integrating these photonic components into traditional electronic systems has posed significant challenges when trying to showcase their benefits. Nonetheless, recent advancements include two new studies that delve into the capabilities of photonic computing chips combined with silicon-based electronics.
Bo Peng and colleagues demonstrate A photonic accelerator named PACE enables extremely fast computing with minimal delay, which is crucial for real-time operations.
This extensive accelerator includes over 16,000 optical elements arranged within a 64x64 grid layout. It facilitates rapid processing speeds reaching up to 1 gigahertz and shows an impressive decrease in minimum lag time—by approximately 500 times—in comparison to smaller circuit setups or individual light-based parts. The experiments indicate that PACE can tackle complex computational challenges referred to as Ising issues, highlighting the potential practical use of this technology.
In an independent paper Nicholas Harris and his team present a photonic processor capable of executing AI models with notable efficiency and precision. This processor comprises four 128 x 128 matrices and can run the natural language processing model BERT as well as a neural network known as ResNet for image processing tasks, achieving levels of accuracy comparable to traditional electronic processors.
The researchers demonstrate that their photonic device can be used for various purposes, like producing text similar to Shakespeare’s work, precisely categorizing film critiques, and even mastering retro video games such as Pac-Man.
Each team claims that their systems can scale up, although some additional fine-tuning is required.
"Photonics in computing has been under development for many years, however, these recent demonstrations could indicate that we are finally ready to utilize the force of light to create more powerful and energy-efficient computational systems," observes Anthony Rizzo in an accompanying News & Views article.
More information: Shiyue Hua and colleagues presented an integrated large-scale photonic accelerator featuring ultralow latency. Nature (2025). DOI: 10.1038/s41586-025-08786-6
Sufi R. Ahmed and colleagues present universal acceleration for artificial intelligence using photonics. Nature (2025). DOI: 10.1038/s41586-025-08854-x
Anthony Rizzo, Photonic chips offer an enhanced performance boost for artificial intelligence technologies. Nature (2025). DOI: 10.1038/d41586-025-00907-5
Supplied by Nature Publishing Group
This tale was initially released on Tech Xplore . Subscribe to our newsletter For the most recent updates on science and technology news.