Luminous Computing, a one-year-old startup, is aiming to build a photonics chip that will handle workloads needed for AI at the speed of light. It’s a moonshot and yet, the young company already has a number of high-profile investors willing to bet on the prospect.
The company has raised $9 million in a seed round led by Bill Gates, Neo’s Ali Partovi and Luke Nosek and Steve Oskoui of Gigafund.
The round also attracted other new investors, including Travis Kalanick’s fund 10100, BoxGroup, Uber CEO Dara Khosrowshahi, and Emil Michael as well as pre-seed investors Class 5 Global, Joshua Browder, Ozmen Ventures, Schox Investments and Third Kind Venture Capital.
Luminous was founded by Michael Gao, the company’s chief strategist, CEO Marcus Gomez and CTO Mitchell Nahmias, whose research at Princeton is the basis of the chip. Gomez started a software-as-a-service business in the fashion industry and more recently worked as a data scientist at Tinder. Gao also founded software startup AlphaSheets.
Luminous’ approach in basic terms is based on using light to move a dense amount data quickly and efficiently. The idea is that by using photonics for all of the major bottlenecks that traditional processors struggle with will be removed.
“While many photonics research efforts focus on general-purpose data movement, Luminous appropriately targets the AI compute market, which is where the demand is,” Partovi of NEO said.
Luminous is not the only startup out there trying to build a supercomputer on a chip, nor is it the first to be focused on photonics. For instance, Lightmatter has raised $33 million, including investment from Google’s venture arm (with participation from Spark Capital and Matrix Partners) to make photonic chips.
The driving factor is a boom in companies seeking to develop chips specifically designed to handle AI and machine learning applications. In 2018, there were at least 45 startups working on AI chips, New York Times reported at the time. Some technology companies, including Apple, Amazon, Facebook and LG are developing their own AI and ML chipsets for specific purposes. The pursuit is fueling interest among venture capitalists and leading to acquisitions.
The architecture of the chip that Luminous Computing is building is based on Nahmias’ research. As part of his thesis at Princeton University, Nahmias built photonic integrated circuits for computing and became a founding researcher in the field of neuromorphic photonics.
“Training an AI system still takes days, when it should take just minutes,” Gomez told TechCrunch in a recent interview.
The foundation of today’s machine learning systems is based on relatively simple operations — but a lot of them. Training these models still takes a lot of time and involves vasts amount of training data. Even when using today’s generation of specialized AI chips, it still often takes days to train a model. Then, that model has to be tested, refined and trained again. So a task that would help accelerate the development of autonomous vehicles, for example, can benefit from chips that can process these operations faster than ever before.
It’s still early days for Luminous. However, Gomez says they already have working silicon. While Gomez wouldn’t disclose when this new chip would be launched, he emphasized that this isn’t some distant fantasy. The company is aiming to ship development kits within the next few years.
Still, Gomez acknowledges the scale of what they’re trying to achieve: to ship a single chip that will replace the computing power of 3,000 boards containing Google’s Tensor processing units (TPU).
The 7-person company plans to use the new round of capital to grow its team, specifically with people who have experience in the semiconductor industry.