SIPH Creates Working Chip for Global Use

Researchers at the University of Pennsylvania have developed a new computer chip that operates in the world, and not in electricity. This discovery promises to significantly accelerate the training of artificial intelligence models (AI), increasing the data transfer rate and reducing the electricity necessary for their operation.

Today, humanity creates ex-class supercomputers, capable of performing a quintillion of operations per second. Despite the growth in the power of computing, modern technologies are still based on the principles of the 1960s.

Recently, scientists are actively engaged in the development of computing systems based on quantum mechanics, however, they are still far from wide implementation. At the same time, the growing need for the processing of huge amounts of data for AI increases the requirements for computing capacities and leads to an increase in energy consumption.

A group of researchers, led by Professor Naderar Enghat, from the School of Engineering and Applied Sciences, has developed a chip based on Silicon Photonics (SIPH), which performs mathematical calculations using light. Light is the fastest way to transmit data, and the use of widely accessible silicon allows you to quickly scale the technology.

The researchers’ main goal was to create a chip capable of performing vector-mature multiplication operations widely used in neural networks that underlie the development of modern AI models.

Instead of completely rethinking the manufacturing process, the researchers made specific modifications to the chip’s structure to control the spread of light within it. This ensured that the light moved in a straight line without scattering inside the chip.

The developed chip, which cannot be hacked, is now ready for implementation. The researchers collaborated with a commercial factory for the manufacture of their SIPH chips, adapting their design to the existing market sizes.

Firosis Aflatuni, assistant professor of electrical engineering and systemic engineering at the university, explained that these chips have the potential to replace graphic processors (GPUs) used by companies for training and classifying models. SIPH chips can serve as a complementary addition to existing AI infrastructure.

In addition, SIPH chips enable faster calculations with lower energy consumption and address data privacy concerns. As multiple calculations can be carried out in parallel, data does not need to be stored in RAM during processing, making unauthorized access to data impossible.

Research results were published today in the journal Nature Photonics