Breaking the Speed Limit: Harnessing Light for Enhanced Matrix Multiplications in AI Systems

Perform matrix multiplications as fast as light for enhanced cyber security

Pythagoras declared, \”All things are number.\” Today, 25 years later, we are surrounded by algebra and mathematics, whether or not we can see them. Artificial intelligence (AI), which exploded in a Cambrian explosion, brought numbers closer to all of us.

Over time, operations between scalars were parallelized to operations between vectors and then matrices. The most energy and time-consuming operation in AI systems today is multiplication between matrices. The \”tiled matrix multiplier\” (TMM), which is a technique that speeds up computation, breaks down matrix operations into smaller pieces to be calculated by the same system over successive time slots. Modern electronic AI engines that use transistors are reaching their inherent limits. They can only compute at clock frequencies higher than 2 GHz.

Light’s compelling credentials–ultrahigh speed and significant energy and footprint saving–offer an answer. A team of photonics researchers from the WinPhos Research Group, led by Professor Nikos Pleros of the Aristotle University of Thessaloniki harnessed light’s power to create a compact silicon-based photonic computer engine that can compute TMMs with a clock frequency of 50 GHz, a world record.


Leave a Reply

Your email address will not be published. Required fields are marked *