AI Calculations at the Speed of Light

New Silicon-Photonic Chip Promises Breakthrough in AI Computing

Researchers at the University of Pennsylvania have developed a cutting-edge chip that leverages light waves, instead of electricity, to perform complex mathematical computations—a breakthrough that could revolutionize the speed and efficiency of AI computing.

The team, led by Nader Engheta and Firooz Aflatouni from Penn’s School of Engineering and Applied Science, designed a silicon-photonic (SiPh) chip capable of performing vector-matrix multiplication, a critical function in neural networks, the backbone of AI systems. By using light waves, which are the fastest form of communication, the chip holds the potential to vastly accelerate the computational power of AI systems while significantly lowering energy consumption.

Silicon photonics, a field merging optics with silicon chip manufacturing, has long been touted as a potential next step in overcoming the limitations of traditional electronic chips, which have not significantly changed since their inception in the 1960s. Engheta’s work, which focuses on manipulating light at the nanoscale, has now been applied to the SiPh platform, utilizing silicon, one of the most abundant and cost-effective materials used in mass-produced computer chips.

The key innovation lies in controlling the propagation of light through a silicon wafer by altering its height at the nanoscale. The precise distribution of these height variations scatters light in controlled patterns, enabling the chip to perform calculations at unprecedented speeds.

In a paper published in Nature Photonics, Engheta and Aflatouni demonstrated that their new chip could carry out computations at the speed of light. This development represents a potential leap forward for AI computing, allowing for faster processing of neural network tasks with far less energy usage, paving the way for more powerful and efficient AI applications.

See “Inverse-designed low-index-contrast structures on a silicon photonics platform for vector–matrix multiplication,” Nature Photonics, February 16, 2024. https://rdcu.be/dyNgT