Optical computer does image recognition without converting light to electricity and back

Researchers built a neural network that processes images entirely as light, using silicon photonics components like interferometers and microring resonators, achieving 94% accuracy on standard image tests while using 100 to 242 times less energy than GPU chips. This matters because every conversion between optical and electrical signals wastes power; if you keep the data as light the whole way through, you save energy at scale.

The fundamental problem with electronic chips is that they hit a power wall — moving electrons through silicon generates heat and wastes energy. Photonic computing avoids this by processing information as light signals, which don't degrade the same way. This paper shows the first complete CNN implemented in optics that stays optical end-to-end (no O/E/O conversions). The real breakthrough is the training method: they built a digital twin for simulation, then fine-tuned the physical device using perturbation algorithms, which means you can actually train these systems despite manufacturing imprecisions and thermal drift. What becomes possible: if this cost curve continues, image processing at the edge could be powered much more efficiently — cameras, medical scanners, autonomous systems. What remains unclear: whether the silicon photonics manufacturing process scales to the same volumes and costs as CMOS, or whether photonic chips stay niche tools for specific tasks.