Engineers at the University of Pennsylvania have developed a groundbreaking chip that utilizes light waves instead of electricity to perform intricate mathematical operations crucial for training artificial intelligence (AI). This silicon-photonic (SiPh) chip, designed by Benjamin Franklin Medal Laureate Nader Engheta and H. Nedwill Ramsey Professor, has the potential to significantly enhance computer processing speed while drastically reducing energy consumption.

Published in Nature Photonics, the collaborative effort between Engheta’s group and Firooz Aflatouni, Associate Professor in Electrical and Systems Engineering, highlights the innovative approach to chip design. By integrating Engheta’s nanoscale material manipulation expertise with Aflatouni’s pioneering research in nanoscale silicon devices, the team aimed to create a platform for vector-matrix multiplication, a fundamental mathematical operation in neural network development.

The chip’s uniqueness lies in the controlled variations in silicon height, reducing it to around 150 nanometers in specific regions. These variations, without the need for additional materials, enable the chip to manipulate light propagation, allowing for lightning-fast mathematical calculations at the speed of light. This breakthrough opens avenues for computing beyond the limitations of conventional chips, which have largely adhered to principles dating back to the 1960s.

Aflatouni notes that due to the design’s compatibility with commercial foundries, it is already poised for commercial applications. The chip could potentially be adapted for use in graphics processing units (GPUs), catering to the soaring demand for AI systems.

The advantages extend beyond speed and energy efficiency; the chip offers enhanced privacy features. Simultaneous computations eliminate the need to store sensitive information in a computer’s working memory, making a future computer powered by this technology virtually impervious to hacking.

As Aflatouni emphasizes, “No one can hack into a non-existing memory to access your information.” The collaborative effort also included co-authors Vahid Nikkhah, Ali Pirmoradi, Farshid Ashtiani, and Brian Edwards of Penn Engineering.

By Impact Lab