A lens instead of a multiply-accumulate array
The team, led by Volker Sorger at the University of Florida with collaborators at UCLA, George Washington University and the Florida Semiconductor Institute, etched two sets of miniature Fresnel lenses onto a silicon-on-insulator chip. Machine learning data is converted into laser light, passed through the lenses, and converted back to a digital signal. A Fresnel lens performs a Fourier transform on light passing through it, and a convolution becomes a multiplication in the Fourier domain, so the optics carry out the arithmetic that a conventional accelerator does with a multiply-accumulate array. [1][3]
Atlas interpretation: That substitution is the whole idea: instead of clocking a grid of digital multipliers, the chip lets a physical property of the lens do the multiplying while the light merely passes through it. [3]
98 percent on handwritten digits, at MNIST scale
The prototype classified handwritten digits, the standard MNIST-style benchmark, at about 98 percent accuracy, matching conventional electronic chips running the same task. The design also supports multiple colored lasers carrying separate data streams through the same lenses at once, which researcher Hangbo Yang described as a distinguishing advantage of doing the math optically rather than electronically. [3][5]
Atlas interpretation: Matching accuracy on MNIST is a low bar by 2025 standards; the digits task exists to prove the optical path preserves enough precision to be useful, not to demonstrate a capability. The actual claim is about the cost of the computation, not what it computes. [3]
10 to 100 times less energy, and what that number covers
The team reported the optical convolution step running at 10 to 100 times the power efficiency of the same operation on conventional chips, describing it as near-zero energy because the lens performs the transform passively as light travels through it rather than through switching transistors. Sorger said the aim was to keep AI systems scaling as compute demand keeps growing. The work was funded by the Office of Naval Research. [3][4]
Atlas interpretation: The efficiency figure describes the convolution operation in isolation, not a full accelerator or a deployed system. Laser generation, the analog-to-digital conversion at each end, and everything outside that one operation still cost energy, and none of the coverage reviewed quantifies the total system power budget or compares it against a specific commercial chip. That leaves the hundredfold figure as a claim about one arithmetic step performed in a lab prototype, worth tracking rather than treating as a finished efficiency result. [3][4]
Sources
- Near-energy-free photonic Fourier transformation for convolution operation acceleration
Advanced Photonics · Sep 8, 2025
- New light-based chip boosts power efficiency of AI tasks 100 fold
SPIE · Sep 8, 2025
- New light-based chip boosts power efficiency of AI tasks 100 fold
University of Florida News · Sep 8, 2025
- New light-based chip boosts power efficiency of AI tasks 100 fold
SPIE · Sep 8, 2025
- Light-Based Chip Boosts AI Power Efficiency Hundredfold
Photonics Media · Sep 8, 2025