Raised $154 million! This company will deploy photonic AI hardware at scale next year

On June 1, Lightmatter, a photonic technology company, announced that it has raised $154 million in Series C financing from SIP Global, Fidelity Management & Research, Viking Global Investors, Google Ventures, Wise & Pathfinder and existing investors. With this round, Lightmatter has raised more than $270 million in funding to date.

Lightmatter is a Boston-based startup that uses light for artificial intelligence computing.

"We use light to connect computer chips together, and we also use light for deep learning computing. The reason we're getting these customer and data center scale deployments with interconnect is because the generative AI boom is driving high-end chips like crazy."

Lightmatter is currently making a big push into the fast-growing AI computing market, which it claims will help the industry level up and also save a lot of power.

Lightmatter's chips essentially use optical flow to solve computational processes such as matrix-vector products. This kind of math is at the heart of much AI work, and is currently performed by GPUs and TPUs that specialize in such work, but they all use traditional silicon gates and transistors.

The problem with this is that we are approaching the limits of density and therefore the speed of computation at a given power or size. This makes training models like GPT-4 huge for supercomputers, requiring huge amounts of power and generating a lot of waste heat.

"Around the world, a large number of companies are hitting energy-power bottlenecks and experiencing huge challenges in the scalability of AI. Traditional chips are pushing the boundaries of what is possible to cool, and data centers are generating increasing amounts of energy consumption." Nick Harris, CEO and founder of Lightmatter, said, "Unless we deploy a new solution in the data center, the progress of AI will slow down significantly."

"Some predict that it takes more energy to train a large language model than 100 American households consume in a year. In addition, others estimate that unless new computing models are created, 10-20% of the world's total electricity will be used for AI reasoning by the end of the next decade."

Lightmatter, of course, is planning to become one of these new paradigms. Its approach is (potentially) faster and more efficient, and they use arrays of microscopic optical waveguides that allow light to essentially pass through them to perform logical operations: a kind of analog-digital hybrid. Since the waveguides are passive, the main power consumption is creating the light itself, and reading and processing the output.

One of the really interesting aspects of this form of optical computing is that we can increase the power of the chip by using more than one color at a time. Blue does one operation and red does the other; although in practice it's more like 800 nm wavelengths for one and 820 nm for the other. Of course, this is no easy task, but these "virtual chips" can greatly increase the amount of computation on the array.

Harris started the company based on the optical computing work he and his team did at MIT, which licensed the patents to them, and managed to raise an $11 million seed round in 2018.

One investor said at the time, "This is not a science project." But Harris admitted in 2021 that while they knew "in principle" the technology should work, there was still a lot to do to make it work.

Now Lightmatter has raised $154 million in Series C funding and is gearing up for its true debut. It has several pilot projects, including Envise (computational hardware), Passage (interconnects, critical for large computational operations) and Idiom (a software platform that Harris says should allow machine learning developers to adapt quickly.

Lightmatter Envise unit

"We've built a software stack that integrates seamlessly with PyTorch and TensorFlow." He explains, "The workflow for machine learning developers is the same: we import the neural networks built in these industry-standard applications into our library, so all the code runs on Envise."

The pilot project is currently in beta testing, with mass production planned for 2024. By then they should have enough feedback and maturity to deploy their own data centers.

Reference links:

[1] https://techcrunch.com/2023/05/31/lightmatters-photonic-ai-hardware-is-ready-to-shine-with-154m-in-new-funding/

[2] https://www.reuters.com/technology/ai-boom-triples-valuation-lightmatter-us-startup-using-light-computing-2023-05-31/