February 23, 2024


Technologists David Lazovsky and Preet Virk, with backgrounds in semiconductor engineering and photonics, realized together a few years ago that: artificial intelligence Machine learning workloads can quickly run into “data movement” problems. They predict that moving data in and out of computing hardware will become increasingly challenging as AI models grow larger than can be held on any one memory chip chip.

Their solution, designed by Phil Winterbottom, a former researcher at renowned Bell Labs, is an optical interconnect technology for compute-to-compute, compute-to-memory, and on-chip data transfer. Lazovsky and Virk founded a startup with Winterbottom, Celestial AI, to commercialize the technology. Now, the startup is attracting big backers.

Celestial AI today announced that it has raised $100 million in Series B funding led by IAG Capital Partners, Koch Disruptive Technologies, and the Temasek Xora Innovation Fund. The funding, which brings Celestial AI’s total funding to more than $165 million, will be used to support production of Celestial’s photonic platform by expanding the company’s engineering, sales and technical marketing departments, according to CEO Lazovsky.

Celestial currently has about 100 employees, a number that Lazovsky expects to grow to 130 by the end of the year.

“Compute and memory are tightly coupled these days. The only way to add more high-bandwidth memory is to add more compute, whether additional compute is required or not,” Lazovsky told TechCrunch via email. “Celestial’s technology enables memory disaggregation.”

In a data center, memory is usually one of the most important expensive Resources – This is partly because resources are not always used efficiently. Because memory is so closely tied to compute, it is challenging, and sometimes impossible, for operators to “disaggregate” and pool memory across hardware within a data center because of bandwidth constraints and extremely high latencies.

according to a Internal Microsoft research shows that as much as 25 percent of memory in Azure is “idled,” or left over, after server cores are leased to virtual machines. Reducing idle memory can cut data center costs by 4% to 5%, the company said estimated — Potential substantial cost savings in multi-billion dollar operations.

Celestial was originally a portfolio company of The Engine, a venture capital firm that spun out of MIT in 2016. Celestial has developed what appears to be a solution in its photonics-based architecture that scales across multi-chip systems. Celestial’s technology, which uses light to transmit data, can move information within a chip and from chip to chip, making memory and computing available for artificial intelligence and other workloads.

skyline

Image Source: skyline

Celestial also claims that its technology can reduce the power required to move data, thereby indirectly improving the chip’s performance. Typically, a chip uses a portion of the power it consumes to move data between circuits, which takes away power that the chip could use directly for computing tasks. Celestial’s photonics reduce the power needed to move data, allowing the chip (at least in theory) to increase its computing power.

Celestial’s photonic technology is compatible with most industry interconnect standards (e.g., CXL, PCIeThe company claims 25 times higher bandwidth and 10 times lower latency and power consumption than fiber-optic alternatives.

“With the growth artificial intelligence , especially large language model (LLM) and recommendation engine workloads, are shifting towards accelerated computing,” Lazovsky said. That’s what we’re solving with Celestial’s photonic structures. “

Celestial offers its interconnect products through a licensing program and said it is working with a number of “tier 1” customers, including hyperscalers as well as processor and memory companies.

Interconnect products appear to be Celestial’s number one priority. The company sells its own AI accelerator chip, called Orion, built on Celestial’s photonics architecture. But as investors told TechCrunch in a recent article on TC+, AI photonic chips have yet to overcome the engineering challenges to make them practical at scale. Unless Celestial breaks through in the areas of data-to-analog conversion and signal regeneration—the biggest stumbling blocks for photonic chips today—Orion is unlikely to get farther than its rivals.

Beyond chips, Celestial has many competitors in the photonic integrated circuit market Can It will be worth $26.42 billion by 2027.

Ayar Labs produces chip solutions based on optical networking principles and has raised more than $200 million in venture capital since its founding in 2015. Another competitor, Renovus, recently raised $73.9 million.

However, future consolidation in the broader optical interconnect space is likely. About three years ago, Marvell acquired Inphi, Optical Network Expert, $10 billion. After a period of silence, Microsoft last year acquired Lumenisity, a startup developing high-speed fiber optic cables for data centers and carrier networks.

Both Inphi’s and Lumenisity’s technologies target different use cases. But Big Tech’s enthusiasm for optics and photonics is notable.

“Our photonic technology is truly different, unique and has remarkable properties,” said Lazovsky. “Given the growth of generative AI workloads brought about by LL.M. The need for connectivity is growing rapidly to support the shift from general-purpose computing data center infrastructure to accelerated computing.”

Samsung Catalyst, Smart Global Holdings, Porsche Automobile Holding SE, The Engine Fund, imec.xpand, M Ventures and Tyche Partners also participated in Celestial’s Series B round.