Photonic Chip Breakthrough Enhances AI Convolution Efficiency with Light-Based Processing

A team of engineers has developed a photonic joint transform correlator (pJTC), a light-based computer chip designed to accelerate convolution operations critical to artificial intelligence (AI) tasks like image recognition and pattern-finding. The chip leverages near-energy-free on-chip Fourier transformation, reducing computational complexity for convolution and cross-correlation from O(N^4) to O(N^2), where N^2 represents input data size.

The pJTC employs two sets of miniature Fresnel lenses, each a fraction of a human hair’s width, fabricated using standard manufacturing processes. Data, such as images, is converted into laser light, processed through these lenses, and converted back into digital signals. This approach achieves up to 100 times greater power efficiency compared to traditional electronic chips, with a demonstrated 98.0% accuracy on a Modified National Institute of Standards and Technology inference task for handwritten digit classification.

The chip’s wavelength-multiplexed architecture allows multiple data streams to be processed simultaneously using different colored lasers, yielding a throughput of 305 TOPS/W and 40.2 TOPS/mm^2 based on current foundry processes. This design reduces computing time and enhances energy efficiency, addressing the high power demands of AI systems.

Led by Volker J. Sorger, Ph.D., from the Uni...