Engineers at the University of Florida have built a photonic chip that performs convolutions, the most compute-heavy operation in modern AI, using light instead of electricity and delivering roughly 100 times greater power efficiency than conventional electronic processors. The device arrives as data centers strain under the energy demands of training and running large neural networks, and it joins a fast-moving wave of optical computing research from labs in the United States, China, and Europe that collectively suggest photonic hardware could reshape how AI workloads are processed.
How Light Replaces Electrons for Convolutions
Convolution is the mathematical backbone of image recognition, video analysis, and many generative AI systems. On a standard GPU, each convolution requires billions of multiply-and-accumulate steps, all carried out by shuttling electrons through transistors that generate heat at every gate. Photonic chips sidestep that thermal penalty by encoding data directly into light waves, allowing calculations to occur as photons propagate through optical structures rather than through resistive silicon pathways.
The University of Florida team’s approach relies on a photonic joint transform correlator (pJTC) that performs on-chip Fourier transformation to accelerate convolution and cross-correlation. By shifting the math into the frequency domain optically, the system reduces computational complexity from O(N^4) to O(N^2), a difference that grows sharply as input sizes increase. The chip uses two-dimensional versions of the same lenses found in lighthouses, shrunk to a fraction of the width of a human hair, to steer and process light on the chip surface.
Unlike electronic accelerators, which must serialize many operations to avoid overheating and manage memory bandwidth, the pJTC architecture exploits the inherent parallelism of light. Multiple wavelengths and spatial modes can carry independent data streams through the same waveguides, effectively stacking many convolutions into a single optical pass. The result is a form of analog computing where the physics of interference and diffraction perform the work that would otherwise be done by vast arrays of digital multipliers.
Efficiency Gains and Accuracy Benchmarks
Raw speed means little if accuracy suffers. Separate photonic processor work at MIT demonstrated that key computations can finish in less than half a nanosecond, with the photonic system achieving more than 96 percent accuracy during training tests and more than 92 percent accuracy during inference on a classification task. Those numbers approach the performance of digital baselines while consuming far less power, because photons do not encounter the resistive losses that electrons face in copper interconnects and transistor channels.
At the extreme low end of energy consumption, researchers have shown that an optical neural network can operate using less than one photon per multiplication, a figure that reframes the energy floor for AI inference. The UF chip’s 100-fold efficiency improvement, reported through SPIE materials, sits within this broader trajectory: each new photonic design pushes the energy cost of a single multiply closer to the physical minimum set by quantum noise rather than by transistor switching.
Maintaining accuracy in these regimes requires careful handling of noise, calibration drift, and device nonlinearity. Many photonic accelerators use hybrid schemes in which the heavy linear algebra is performed optically, while nonlinear activation functions and error correction remain in the electronic domain. This division of labor allows the system to reap most of the energy benefits of optics without sacrificing the numerical robustness of mature digital circuits.
Solving the Fan-In Bottleneck
One persistent barrier to scaling photonic convolution hardware has been fan-in, the process of combining many optical signals into a single output without losing information. A team led by professors Zhang Xinliang and Dong Jianji at Huazhong University of Science and Technology addressed this directly with a compact multimode convolver that uses lossless mode-division fan-in. Published in Nature Communications, their device occupies just 0.42 mm2, tolerates fabrication variances of plus or minus 15 nm, and operates across 35 nm of optical bandwidth.
Those fabrication tolerances matter because they determine whether a design can move from a single lab demonstration to volume manufacturing. A chip that fails when lithography drifts by a few nanometers is a research curiosity; one that holds performance across a 15 nm window starts to look compatible with commercial foundry processes. The 0.42 mm2 footprint is also significant: smaller devices mean more convolvers per wafer and lower unit costs, a prerequisite for any technology aiming to displace entrenched electronic accelerators.
By encoding different input channels into distinct spatial modes rather than separate waveguides, the multimode fan-in architecture also eases routing congestion. That, in turn, simplifies the layout of large-scale photonic neural networks, where hundreds or thousands of convolution kernels must be interconnected without incurring excessive loss or crosstalk.
Reconfigurable Designs and System-Level Power
Speed and footprint are necessary but not sufficient. Real-world deployment also demands flexibility, because different AI models require different convolution kernel sizes and network depths. A reconfigurable silicon photonic chip developed within the SmartLight/iPronics platform tackles this by allowing the optical mesh to be reprogrammed for different CNN architectures. End-to-end power and accuracy benchmarking against digital CNN baselines showed that system design choices, particularly reducing the number of photoreceivers and analog-to-digital converter blocks, cut total power draw without proportional accuracy loss.
Earlier foundational work on an integrated photonic tensor core demonstrated parallel convolution processing using phase-change memory arrays paired with optical frequency combs. That architecture established baselines for energy and throughput that newer designs now build on, and it proved that photonic in-memory computing, where weights are stored optically rather than fetched from electronic DRAM, can eliminate one of the largest power sinks in conventional AI inference.
System-level studies increasingly highlight that the surrounding electronics, including drivers, control logic, and data converters, can dominate power consumption if left unchecked. The most promising photonic AI engines therefore co-design optics and electronics, trimming peripheral circuits and exploiting sparsity or reduced precision where model tolerance allows. The University of Florida pJTC chip fits this trend by concentrating optical resources where convolutions are densest and relying on streamlined electronic interfaces elsewhere.
From Lab Bench to Orbit and Beyond
The University of Florida’s ambitions extend well past the lab. On October 27, 2025, Florida engineers tested photonic AI chips in space, marking what Volker J. Sorger, Ph.D., called “a first-of-its-kind validation of photonic computing in space.” Orbital environments are ideal stress tests for photonic hardware: radiation, thermal cycling, and power constraints are all more severe than in terrestrial data centers. If a chip survives and performs in orbit, its path to edge deployment on Earth, in autonomous vehicles, drones, and remote sensors, becomes far more credible.
Internationally, the competition is intensifying. Researchers in China, led by Chen and colleagues, have produced an all-optical chip called LightGen that can run advanced generative AI models, not just the classification tasks that most photonic demos target. European teams are exploring neuromorphic photonics for ultrafast signal processing, while U.S. groups continue to refine hybrid electro-optic accelerators that can slot into existing server architectures with minimal disruption.
Underpinning much of this progress is the open circulation of preprints and designs through repositories such as arXiv member institutions, which allow hardware concepts to spread quickly across borders and disciplines. Community support, including individual donations to arXiv, helps sustain this shared infrastructure at a time when photonic AI research is expanding rapidly and demands fast, open dissemination.
For now, electronic GPUs remain the workhorses of large-scale AI, but the trajectory of photonic convolution chips points toward a more heterogeneous future. As fan-in bottlenecks are addressed, reconfigurable meshes mature, and space-qualified prototypes prove their resilience, light-based accelerators are likely to take on specialized roles wherever energy efficiency and latency are paramount. The University of Florida’s pJTC chip adds another strong data point to that trend, suggesting that the next major leap in AI performance may come not from shrinking transistors, but from letting photons do more of the heavy lifting.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
