With the fast growth of artificial intelligence (AI), large systems now spread computing work across many units located in separate data centers. These units must exchange huge amounts of data quickly and with little delay or power use. Older methods suffer from delays and use more energy than desired.
Scientists led by The Chinese University of Hong Kong have created an integrated all-optical signal processor built on a silicon photonic chip that moves and handles information using light instead of electricity. The processor fixes distortions in the light signals while they remain in light form and avoids converting them to electrical signals first.
It reaches a total data rate of 1.6 terabits per second. The processing delay stays below 60 picoseconds. Energy use stays very low at only tens of femtojoules per bit, an extremely small amount of energy for each piece of data.
Direct light processing lowers delays and energy needs
The processor draws ideas from brain-like computing and machine learning. It adjusts light paths inside the chip to better read complex signal patterns and correct problems. It acts as a flexible corrector that can be changed to handle different issues, such as the spreading of light colors in fibers known as chromatic dispersion, limits in equipment speed, and extra distortions from heavy use. This method restores signals more effectively than older electronic approaches and can allow more than six times as many light channels to work in one fiber without problems.
Tests showed it handled signals from eight light channels at the same time, each at high speed with a common data encoding method. The processor kept stable low energy use and very short delays across the channels. These results point to strong potential for linking distant data centers that support large artificial intelligence training tasks.
The work follows earlier progress in using light for long distance communication. It shows that light can now carry data and also process it directly, which may support more efficient systems for future computing needs.
This research is published in Science.