Light the Way

Photonic chips could speed AI while decreasing its energy usage

AI might ramp up our computing ability to levels we could previously only imagine. But it uses up a great deal of energy in doing so. And as we incorporate AI into more facets of life, we’ll need more power to keep it running. The IEA estimates that energy demand from AI will be 10 times in 2026 what it was in 2023. It doesn’t seem likely that we’re going to ditch AI anytime soon, which means we need a more efficient way of producing it. The solution might lie in the fastest thing in the universe: light.

Photonic Power

Traditional computing runs on electrons. Electronic chips require a lot of electricity to function, and they need even more when powering AI because their processing and memory are separated. They have to work hard to transfer data between the two. Light, on the other hand, not only travels faster than anything else but encounters little resistance as it does so.

Light-based chips that run on photons instead of electrons have a higher bandwidth than electronic chips, meaning they can transmit more information. And with optical frequencies higher than electrical ones, these chips could perform computations faster with less latency, a boon to edge computing.

Photonic chips are “paving the way for breakthroughs in fields that demand high-speed and high-efficiency processing, such as artificial intelligence,” University of Cambridge physicist Natalia Berloff told Quanta Magazine.

Photons also generate less heat than electrons do. Because of this, an electronic chip’s transistors can’t all operate at the same time. But a photonic chip’s potentially could. And considering the amount of water data centers consume to keep servers from overheating, the promise of photonic chips is quite appealing.

“If we could harness” this ability, Stanford electrical engineer Gordon Wetzstein told Quanta, “this would open a lot of new possibilities.”

Speeding Up

Right now, the fastest optical neural networks like HITOP can perform 1 trillion operations per second, 300 times slower than the most sophisticated electronic chips made by Nvidia. But HITOP does that while using 1,000 times less energy. The potential is there.

Lightmatter, which in July announced the hiring of Nvidia executive Simona Jankowski as CFO, developed its photonic chips using electrons for memory and photons for matrix multiplication, crucial for training networks and processing data.

“The biggest companies in the world are hitting an energy power wall and experiencing massive challenges with AI scalability, especially when it comes to chip memory and performance. This has profound effects for everyone, with major implications for carbon emissions and rising financial costs,” Lightmatter CEO Nick Harris told Pulse 2.0. “Lightmatter is providing fundamentally new interconnect and computing technologies that enable sustainability for these power-hungry workloads by leveraging photonics.”

Taichi, a neural network chip developed at China’s Tsinghua University, uses two types of light-based processing to lower energy use 1,000-fold while computing at an accuracy comparable to leading electronic chips. Charged with sorting images into 1,000 categories, Taichi did so at a 92% rate, and it can recreate the styles of musical and visual artists.

A silicon-photonic chip developed by University of Pennsylvania researchers uses lasers to reconfigure what the optical system does, a task that is extraordinarily difficult in previous electronic and photonic chips. With its ability to change quickly and perform many calculations simultaneously, it might not need to store information in a computer’s memory.

“No one can hack into a non-existing memory to access your information,” Firooz Aflatouni, an associate professor in electrical and systems engineering and one of the chip’s developers, told Phys Org.

Achieving AGI

If photonic chips can harness the power of light well enough to improve AI’s utility while drastically decreasing the amount of energy needed, they could enable the development of artificial general intelligence. Beyond the capabilities of current AI, AGI would possess human-level intelligence.

Lightmatter boasts on its website that “we sell the photonic chips that will make building AGI possible.” Taichi’s creators argued in the Journal of Science that “Taichi indicates the great potential of on-chip photonic computing for processing a variety of complex tasks with large network models, which enables real-life applications of optical computing. We anticipate that Taichi will accelerate the development of more powerful optical solutions as critical support for the foundation model and a new era of AGI.”

While there is some question whether AGI is even possible, there are those who argue that not only is it possible, it’s coming soon. Futurist and computer scientist Ray Kurzweil, author of 2005’s “The Singularity is Near” and this year’s “The Singularity is Nearer,” built predictive models that suggest AGI will be possible by the end of the decade. SingularityNET CEO Ben Goertzel agrees.

“To me it seems quite plausible we could get to human-level AGI within, let’s say, the next three to eight years,” Goertzel said at this year’s Beneficial AGI Summit in Panama.

The speed of the rise of photonic chips will have a big say in that.