A group of scientists from the University of Pennsylvania’s School of Engineering and Applied Sciences has developed a new chip that uses light waves instead of electricity to perform complex mathematical operations. The silicon photonic chip can be fabricated on modern hardware and used as a co-processor for the GPU, for example in machine learning-related tasks.

The scientists built and tested the chip on vector matrix multiplication operations for 2×2 and 3×3 matrices. The possibility of working with a 10×10 matrix was also demonstrated. These examples showed that the proposed methods have the potential to create large-scale analog computing platforms based on light waves, the scientists reported in a journal article. Nature Photonics.

The basis of the work is the proof of concept for fabricating waveguides and amorphous lenses directly on a silicon wafer using standard etching and wafer processing techniques. Traditional metastructure fabrication methods suffer from limitations such as narrow bandwidth and high sensitivity to fabrication errors. In particular, it limits the scalability of such architectures.

Instead of using a silicon wafer of the same height, the developers explain: “You thin the silicone, for example, to 150 nanometers”, but only in certain regions. These changes in height provide a way to control the propagation of light across the chip without adding any other material; because changes in elevation can be spaced out to ensure the light is dispersed in a specific way, allowing the chip to perform mathematical calculations. Speed ​​of light.

Simply put, waveguides are etched into silicon and a lensing system is created that will allow a light signal to pass through a waveguide maze with a strictly defined algorithm, resulting in a specific result depending on the input signals. Such a coprocessor can be adapted to a regular GPU to get rid of the energy-consuming operations of vector matrix multiplication and thus speed up the computation of artificial intelligence and machine learning tasks.