Lightmatter’s Photonic AI Hardware Ready to Shine in New $154M Funding

Image credit: light matter

Photonic computing startup Lightmatter is poised to take the fast-growing AI computing market to the next level with a combination of hardware and software that it claims can help raise the bar for the industry and significantly reduce power requirements to boot. I am taking on the challenge.

Lightmatter’s chip essentially uses optical flow to solve computational processes such as matrix-vector multiplication. This computation is at the heart of much AI work and is currently performed by GPUs and TPUs that are dedicated to this computation but use traditional silicon gates and transistors.

These issues are approaching the limits of density, or speed, for a given wattage or size. Progress continues, but at great cost and pushing the boundaries of classical physics. Supercomputers like GPT-4 that enable training models are huge, consume a lot of power, and generate a lot of waste heat.

“The world’s largest companies are hitting energy power barriers and facing major challenges with AI scalability. Conventional chips are pushing the limits of cooling, creating an ever-larger energy footprint in data centers. Unless we introduce new solutions to our data centers, progress in AI will slow significantly,” said Nick Harris, CEO and founder of Lightmatter.

“Some predict that training a single large language model will require more energy than 100 US homes consume in a year. Moreover, there are also estimates that 10%-20% of the world’s total power will be used for his AI inference by the end of the decade, unless a new computing paradigm is created. ”

Of course, Lightmatter aims to be one of those new paradigms. That approach, which uses an array of microscopic optical waveguides to essentially allow logical operations to be performed simply by light passing through them, is at least potentially faster and more efficient. It’s a kind of analog-digital hybrid. Since waveguides are passive, the main power consumption is generating the light itself and reading and processing the output.

One very interesting aspect of this form of optical computing is the ability to increase the output of a chip simply by using more than one color at a time. Blue performs one operation, red performs another operation. However, in practice, it is like a wavelength of 800 nanometers performs one operation of his, and a wavelength of 820 nanometers performs another. Of course, doing so is not trivial, but these “virtual chips” can significantly increase the amount of computation performed on the array. Twice the color, twice the power.

Harris founded the company on the optical computing research he and his team did at MIT (which we have licensed related patents to) and won an $11 million seed round in 2018. succeeded in At the time, one investor said: But in 2021, Harris admitted that while he knew the technology should work “in principle,” there was a lot of work to do to make it practical. Luckily, he told me that in the context of an investor putting another $80 million into the company.

Now, Lightmatter has raised $154 million in a C round and is gearing up for its real debut. The company used a full stack of Envise (computing hardware), Passage (interconnection, essential for large-scale computing operations), and Idiom, a software platform that allows machine learning developers to adapt quickly. I’m doing some pilots.

A unit of captured Lightmatter Envis. Image credit: light matter

“We built a software stack that seamlessly integrates with PyTorch and TensorFlow. The workflow for machine learning developers from there is the same. We import it, so all our code runs on Envise,” he explained.

The company doesn’t make specific claims about speed or efficiency, and the differences in architecture and computing methods make it difficult to make like-for-like comparisons. But we’re definitely talking not just 10% or 15%, but orders of magnitude. Interconnects are upgraded as well. Because it doesn’t make sense for him to isolate that level of processing onto one board.

Of course, this is not the kind of general-purpose chip that can be used in laptops. It’s very specific to this task. However, the lack of specificity in tasks on this scale appears to be slowing his AI development. However, the development of AI is progressing at such a rapid pace that “holding back” is a misnomer. However, its development is hugely expensive and difficult to work with.

The pilot is in beta, and mass production is planned for 2024, at which point we should have enough feedback and maturity to deploy it in our data centers.

Funding for this round was provided by SIP Global, Fidelity Management & Research Company, Viking Global Investors, GV, HPE Pathfinder and existing investors.