VCs Bet on AI Chips That Use Light Instead of Electronics
When Nvidia CEO Jensen Huang in March said his company would release hardware that lets an AI data center chip send information directly over optical cables instead of needing to convert to optical from electrical signals, it was a shot in the arm for optics developers that have long wanted a bigger role in data centers.
Using light to send information over fiber optic cables is well-established, but attempts to use light for computation in data center servers have struggled. Some investors now think the time is right to use optical technology for active computation in the chips themselves, instead of just using it to ferry data between chips.
The idea is that light can encode information in its wavelength, amplitude and phase. By bouncing light off a clever arrangement of mirrors and other devices, for instance, and allowing the beams of light to interact, the light can perform addition and multiplication—essentially replacing traditional electronic chips.
Compared to moving electrons in copper wires or transmitting radio waves over the air, sending light through fiber optics “is both very contained, very fast, and a piece of fiber is like the size of a strand of hair,” said Bill Long, chief strategy officer at Zayo, which provides fiber optic connections, including to AI data centers.
Plus, light gives off less heat than electronics, reducing the risk of overheating, compared to a standard graphics processing unit for AI. “If you want to just make a GPU 100 times faster, and you do nothing about the fundamental energy efficiency, what you’re saying is the GPU is going to burn 100 times the power, and everything melts,” said Patrick Bowen, CEO of Neurophos, which designs photonics chips. Those chips direct light using preposterously small components, drawing on research on invisibility cloaks at Duke University.
Since March, investment firms have poured hundreds of millions of dollars into photonics-focused startups such as Lumotive, Arago and Q.Ant that aim to use light instead of electrons for computation. Geometry, a new venture capital fund that invests in photonics and other “deep tech” sectors, has nearly finished raising $30 million and invested in Irradiant Technologies, which manufactures photonic components, said Ben Levy, a general partner of the fund.
Some kinds of math are difficult to do with light and still require electronics, Levy said. That, along with the difficulty of competing against Nvidia, is why some photonics firms focus on hardware that can move information between traditional chips in a data center, similar to Nvidia’s upcoming product, which is known as a switch. Huang said the optics switch would save lots of energy in an AI data center, compared to traditional electronic switches.
But Geometry’s other general partner, Tom Walton-Pocock, says that the time is right for the more ambitious idea of light-based chips. One reason is recent advances in the materials that go into photonics components. The other reason comes down to the kind of math that modern AI models use. “If you look at a neural net today, it’s huge, dense blocks of matrix multiplication,” he said.
Those are exactly the kind of operations where photonics—er—shines.