Since AI models have become increasingly computationally complex, engineers are looking at different types of hardware and materials that can speed up the creation process. One type of component with promise is photonic chips which use light to send signals instead of the electric current that traditional processors use. In the theory of things, photonic chips can improve performance because light generates less heat than electricity, can move faster and is more resistant to temperature fluctuations, as well as an electromagnetic field.
Celestial AI lands $56M to create lights-based AI accelerator chips
However, some disadvantages need to be resolved in order for the technology to be accepted by the general public. They’re larger physically than their electronic counterparts and hard to manufacture in mass quantities, for instance, because of the age of the photonic chip manufacturing plants. Additionally, photonic designs heavily depend upon electronic circuits for control which could cause bottlenecks.
The challenges haven’t stopped companies currently exploring photonics, including Lightmatter and LightOn, and tech giants like Intel or Japan’s NTT. The latest company to join the growing market will be celestial AI, founded in 2019 to create a software and hardware platform for AI chips that use light to transfer data through chips. Celestial today announced it has raised $56 million of funding, based on an estimated post-money value of $150 million and bringing the total capital raised to more than 64 million dollars.
Light-based AI processors
Celestial (previously Inorganic Intelligence) is a portfolio company that is part of The Engine, the venture capital firm that was spun out of MIT in 2016. Phil Winterbottom, David Lazovsky, Michelle Tomasko, and Preet Virk started the company to develop products based on the light that “highly effectively” create maps of data and perform computation with no need to perform intricate optimizations.
Winterbottom has previously worked in Bell Labs, Nokia’s distinguished R&D division, before working at fiber access startup Entrisphere (which was purchased from Ericsson) along with cable-networking company Gain speed (which was later purchased through Nokia). Tomasko, in charge of Celestial’s software development, has worked in various engineering positions at Nvidia and Google.
Celestial declares to have created an architecture based on photonics that can scale across multichip systems, allowing data transfer through light within chips and chip-to-chip. Celestial’s hardware product incorporates photonics into the AI acceleration system, making computing and memory available for AI applications.
“Celestial was created to make use of technological advances in silicon photonics powered by data communications as well as the most advanced chipset design to separate system capabilities from those of electronic-only systems,” CEO David Lazovsky said to VentureBeat by email. “Celestial’s goal is to fundamentally change computing efficiency by leveraging our technology platform that utilizes light to enable data movement inside and between chips.”
In conventional hardware, transistors regulate the movement of electrons through the semiconductor and perform functions (and operating software) by cutting information into a set of zeros and ones. Contrary to this, Celestial’s hardware, like other photonics hardware, calculates by mixing and splitting beams of light in nanometer-wide channels. The calculations made by photonics chips are analog, not digital, and therefore more inaccurate. However, they can carry out these calculations, including the calculations required to train AI models, very quickly and in parallel.
“We have created an adaptable general-purpose AI accelerator with wide applicability to a variety of machine learning applications that include computer vision as well as recommendation and personalization and natural language processing and many more,” Lazovsky said. “[Thenew funding is being used to expand our Orion AI accelerators as well as our design services partnership with Broadcom in developing the 5nm ASIC produced at TSMC. Orion AI accelerators will be available in two forms: a standard PCIe device (75W TDP) and a multi-chip computer configuration that is optically connected.”
Also Read: CommerceIQ Raises $115M in Series D Funding
The specifics of Celestial’s tech remain elusive, and we’ve contacted them to inquire about the technology, but the startup is competing against established competitors in photonics. Lightmatter offers server blades that contain photonic chips compatible with traditional data centers. LightOn announced in the last quarter of 2013 the availability of one of their photonic “co-processors” will integrate into the French Jean Zay supercomputer. In December, Lightelligence declared that it had solved a difficult math problem a hundred times more quickly than the typical GPU using its photonic hardware.
On the other hand, while the competition is fierce, Verified Market Research predicts that the market of AI chips could amount to $109.83 billion in 2028, which makes even a tiny portion of the market extremely lucrative. The photonics industry trade association estimates it. The Photonics Public Private Partnership predicts with optimism that just 20 percent of the potential and advantages of light technology are yet to be fully realized.
“The AI industry is experiencing an unprecedented increase in growth, the product of ever-growing data sets from ever-growing sources, coupled with the growing complexity of AI jobs,” Lazovsky added. “Celestial’s technology platform provides an innovative set of software tools into the equation that will allow our customers to tackle this issue. We’re harnessing the capabilities of electronics for high-performance, high-precision computing and photonics to provide high-speed, low-power, and broadband data motion. Its benefits of machine learning extend beyond power and performance to low latency, user-friendly software, and lower costs of operation.”
KDT led Celestial’s most recent financing round with Temasek’s Xora Fund, The Engine, Tyche Partners, M-Ventures, IMEC XPand, and Fitz Gate. The company, which is still in the pre-revenue stage and has focused its efforts on “a few strategic collaborations with customers,” currently says that the capital raised is intended to grow its engineering team across the globe, product development, and strategic supplier relationships, such as the semiconductor maker Broadcom.