Image by Josch13 from Pixabay

Discovery from Polytechnique Montréal could reduce the energy consumption of AI

A team from Polytechnique Montreal has identified a new organic material that could significantly increase the performance of photonic chips, without requiring major modifications to existing infrastructure.

This could one day reduce the electricity consumption of data centers and generative artificial intelligence (AI) systems — a discovery that makes perfect sense just days after a UN agency warned that the water and electricity consumption of data centers is expected to double by 2030, mainly due to the explosion in demand for AI.

“The problem really revolves around current computing capabilities,” explained the head of this work, Professor Stéphane Kéna-Cohen.

“The speed of electronic chips as we know them (…) is limited. The strategy since 2005 has been to add more and more transistors and also to enlarge computer chips (but) the larger the chips become, the more difficult communication between the different parts of the chips becomes.”

And the solution to this problem, he continues, “is to use light to communicate between the different computer chips, (…) what we call photonic chips.”

Every second, the data behind billions of emails, TikTok videos, and AI queries travels the world as light pulses via fiber optic networks.

Along the way, these signals pass through tiny components that act as channels for light: photonic chips. These devices do more than simply carry the signals; they also direct and combine them, ensuring efficient information flow through complex networks. 

But photonic chips have their limits since, for example, each question asked to an agent like ChatGPT requires multiple round trips between components, which increases the number of times the signals have to be converted and transformed.

This is where the discovery by Professor Kéna-Cohen and his colleagues comes in.

“We are transforming a truly passive system for light into an active system that allows us to modulate light, amplify light, and add functions that are currently absent on these silicon-based chips,” he explained.

The team identified a new material that can be directly integrated onto silicon, they explained, “enabling it to perform advanced optical functions.” Instead of constantly converting electrical signals into photonic signals and vice versa, this material allows light to be processed directly. 

This breakthrough relies on an organic molecule that allows light beams to interact as they pass through the material. This opens the door to functions such as amplification and modulation directly on the chip.

Moreover, it is claimed, the material could be easily integrated into current chip manufacturing processes.

“It’s a kind of thin organic layer that can be deposited on current photonic chips without having to rethink the architecture, without having to change the material,” said Professor Kéna-Cohen. “It’s something that can be added at the end of the manufacturing process to give it the functionality we would like.”

In short, he explained, “what we can do with these thin films is amplify the light, so we can use a high-power light signal to amplify a second signal on the chip, the second signal being the one we want to transmit over long distances.”

We can therefore imagine a new generation of optical components used to encode information, amplify signals and generate custom light shapes.

The report from the UN Institute for Water, Environment and Health estimates that, within a few years, data centers could consume up to 9 billion cubic meters of fresh water per year, equivalent to the annual needs of some 1.3 billion people living in sub-Saharan Africa.

Moreover, according to a press release, data center electricity consumption is expected to triple by 2030 compared to 2023 levels, reaching approximately 945 terawatt-hours. This is nearly three times the combined annual electricity consumption of Pakistan, Bangladesh, and Nigeria.

“These chips generate a lot of heat, requiring active cooling to control this heat dissipation,” said Professor Kéna-Cohen. “When information is processed using light, no heat is dissipated. The only heat dissipated is to generate the light signals at the beginning of the process. Therefore, the more we can transpose the computation from the electronic domain to the optical domain, the greater the potential for reducing energy consumption.”

The technology developed at Polytechnique Montréal is not yet ready for commercial deployment. Professor Kéna-Cohen estimates that we are currently exploiting barely 1% of its potential, a performance he hopes to increase tenfold within a year or two.

“In the short term, we’re mainly focused on simply enabling communication between the different, rather conventional chips,” he concluded. “And in the long term, what we’d like to do is completely replace the chip itself, or at least the most energy-intensive operation within the chip, with an operation performed using light.”

The details of this discovery were published by the journal Science Advances.

–This report by La Presse Canadienne was translated by CityNews