Overview

• Researchers report the prototype’s results in Advanced Photonics on Sept. 8, developed with collaborators at the Florida Semiconductor Institute, UCLA, and George Washington University.
• The chip performs convolution by converting inputs to laser light, passing them through on‑chip Fresnel lenses for the transform, then converting outputs back to digital signals.
• Lab evaluations show roughly 98% accuracy on handwritten digit classification, comparable to conventional electronic implementations.
• Energy use for the convolution step is reported to be 10 to 100 times lower than current chips, with reduced runtime for the operation.
• The design demonstrates parallel processing via wavelength multiplexing and is fabricated using standard semiconductor techniques, with researchers noting existing optical elements in some NVIDIA systems could ease future integration.