New Light-Based Chip Supercharges AI Efficiency by Up to 100x (scitechdaily.com)

A team led by Volker J. Sorger at the University of Florida announced a prototype AI chip that performs one of the most compute- and energy-heavy operations in neural networks—convolution—using light instead of electricity. Published in Advanced Photonics (Sept. 8, 2025), the design uses on-chip lasers and microscopic Fresnel lenses to implement photonic Fourier transforms that carry out convolution with roughly 10–100× better energy efficiency than equivalent electronic calculations, while matching accuracy (≈98% on a handwritten-digit test). The result promises substantially lower power draw for vision and pattern-recognition workloads and could ease the growing strain AI places on datacenter and grid power as models scale. Technically, the chip converts digital data into laser light, passes it through two stacked 2D Fresnel-lens arrays fabricated with standard processes, and reconverts the optical result back to electronic signals—effectively performing convolution in the optics domain. Because photonics supports wavelength-division parallelism, the approach can process multiple data streams simultaneously and reduce latency. The team demonstrated a practical, manufacturable path for integrating optical convolution into existing hardware stacks, suggesting a near-term route to hybrid electronic–photonic AI accelerators that boost throughput and energy efficiency for future large-scale models.