A reprogrammable chip that manipulates light to perform AI’s nonlinear calculations could slash energy costs and supercharge machine learning, ushering in an era of fully optical, lightning-fast computer
Postdoctoral fellow Tianwei Wu (left) and Professor Liang Feng (right) in the lab, demonstrating some of the apparatus used to develop the new, light-powered chip. (Credit: Sylvia Zhang)
Penn Engineers have developed the first programmable chip to train nonlinear neural networks using light. This breakthrough could dramatically speed up AI training, reduce energy use, and pave the way for fully light-powered computers.
Unlike today's AI chips that rely on electricity, this new chip is photonic — it uses beams of light to carry out complex mathematical functions. As described in Nature Photonics, the chip reshapes how light behaves to execute nonlinear operations at the heart of deep learning.
The Missing Piece in Photonic AI
AI systems rely on deep neural networks composed of nodes that perform nonlinear operations, which are critical for learning and decision-making. Without nonlinearity, multiple layers collapse into a single ineffective linear model.
While earlier photonic chips handled linear operations, none could perform nonlinear functions using only light—until now.
“Without nonlinear functions, photonic chips can’t train deep networks or perform truly intelligent tasks,” says Tianwei Wu, a postdoctoral fellow and lead author of the study.
Reshaping Light with Light
The breakthrough centers on a semiconductor material that can be optically reprogrammed. A "signal" beam carrying input data passes through the material, while a separate "pump" beam controls the material's properties.
By adjusting the pump beam's pattern and intensity, researchers control how the material reacts to the signal, effectively programming different nonlinear functions on the chip.
“We’re not changing the chip’s structure,” senior author Liang Feng explains. “We’re using light to create patterns inside the material that reshape how light moves through it.”
Training at the Speed of Light
To test the platform, the team ran benchmark tasks. The chip achieved over 97% accuracy on a nonlinear classification problem and over 96% on the Iris flower dataset, matching or exceeding traditional digital models.
Notably, just four nonlinear optical nodes on the chip performed as well as 20 linear nodes with activation functions in a conventional network, highlighting the system’s efficiency.
Unlike previous fixed-function photonic systems, this chip is reprogrammable: the pump beam draws new instructions in real time, making it a true field-programmable photonic computer.
Future Directions
While the current implementation focuses on polynomial functions, the researchers aim to implement more complex operations like exponential or inverse functions in the future, potentially enabling photonic training of large language models.
By replacing traditional electronics with low-power optics, the chip could also reduce energy usage in AI data centers, transforming machine learning's environmental and economic footprint.
“This could be the beginning of photonic computing as a serious alternative to electronics,” says Feng. “Penn is the birthplace of ENIAC — the world’s first digital computer. This chip may be the first real step toward a photonic ENIAC.”
Source:
Journal reference: