New invention promises ultra-fast and much greener AI

Scientists at Université Laval have created an optical chip that can transmit data at 1,000 gigabits per second. That’s fast enough to transfer the equivalent of 100 million books in seven minutes, so the implications for AI development are huge.

AI is currently doing the environment no favors when it comes to electricity use. Its growing appetite is driving historic spikes in power demand and will be responsible for nearly half of data center power consumption by the end of 2025.

In short, the future might be quite nightmarish. It doesn’t have to be, though, and scientists are doing their best to align AI development with energy sustainability.

A team from Laval’s Centre for Optics, Photonics and Lasers has come up with an optical chip that can transfer massive amounts of data at ultra-high speed.

As thin as a strand of hair, this technology offers unrivalled energy efficiency, as the chip only consumes the energy needed to warm a single milliliter of water by one degree.

Naturally, researchers think they’ve created a solution to a critical bottleneck in AI systems. Power-hungry data centers require massive amounts of information to flow between processors.

Current AI systems like ChatGPT demand enormous computational resources, but the connections between processors often become the limiting factor.

Today’s AI data centres rely on tens or even hundreds of thousands of processors, communicating like neurons in the brain. Each processor spans a few millimetres, but when you add it all up, the infrastructure quickly becomes enormous, and so does the energy to power it.

However, the innovation, published in Nature Photonics, cuts through the barriers by harnessing the power of light to transmit information.

Unlike traditional systems that rely solely on light intensity, this chip also uses the phase of light, in other words, its shift. Its performance is unprecedented.

“We’re jumping from 56 gigabits per second to 1,000 gigabits per second,” said PhD student Alireza Geravand, the first author of the study, who sees massive potential for training AI models.

“At 1,000 gigabits per second, you could transfer an entire training dataset – the equivalent of over 100 million books – in under seven minutes. That’s about the time it takes to brew a cup of coffee.”

This technology could hit the market in the coming years. Companies like Nvidia are already starting to use microring modulators – key components in optical communication systems, particularly for high-speed data transmission on integrated photonic chips – albeit limited to light intensity.