Celestial AI Inc., a startup that develops optical technology for linking together processors and memory modules, has secured $175 million in new funding.

The announcement of the Series C round today comes shortly after the company’s disclosure that it has multiple hyperscale and semiconductor industry customers. Two major semiconductor suppliers, Samsung Electronics Co. Ltd. and Advanced Micro Devices Inc., participated in Celestial AI’s new round. They were joined by more than a half-dozen other backers, including U.S. Innovative Technology Fund, which led the investment.

Processors move data to and from RAM in the form of electrical signals. Celestial AI has developed an interconnect, Photonic Fabric, that makes it possible to transport information as light rather than electricity. The company says its technology can improve memory bandwidth and capacity by a factor of 25 while significantly lowering power consumption.

Celestial AI argues Photonic Fabric could enable a new, more efficient approach to designing artificial intelligence infrastructure.

Many AI chips ship with an integrated memory module that they use to store neural networks’ data. This module is typically based on a RAM architecture known as HBM3. Memory based on that architecture can move data to and from the processor to which it’s attached significantly faster than standard RAM, which speeds up processing.

The tradeoff is that a single AI chip can only be equipped with a limited amount of HBM3 memory. If a company requires more capacity, it must purchase additional chips. Celestial AI’s technology reportedly makes it possible to increase the amount of HBM3 memory in an AI server cluster without buying as many additional processors, which significantly decreases hardware costs.

The most advanced graphics cards on the market feature less than one terabyte of onboard HBM3 memory. According to Celestial AI, its Photonic Fabric can link AI chips with up to tens of terabytes of memory. 

“Today’s advanced AI models require exponentially more I/O bandwidth and memory capacity but are currently constrained by low bandwidth and high latency,” saidI co-founder and Chief Executive Dave Lazovsky.

Current AI processors’ limited memory capacity stems from space constraints. Because of various technical considerations, chipmakers can only place HBM3 modules immediately next to a processor at a distance of a few millimeters. That means there’s relatively little space available for RAM, which limits the amount of memory capacity a chip can provide.

Celestial AI’s Photonic Fabric solves the challenge by allowing HBM3 modules to be placed at a greater distance from processors. Data travels to and from those modules in the form of light. According to Celestial AI, its hardware can work with not only HBM3 but also HBM4, an upcoming iteration of the memory technology that is expected to deliver a significant performance increase.

The company will use the proceeds from its new funding round to support go-to-market initiatives. Celestial AI says that it’s collaborating with multiple customers, including hyperscale data center operators and chipmakers, to commercialize Photonic Fabric. Those customers are working to implement the technology in chiplets, compact semiconductor modules that can be integrated into processors such as graphics cards.

Image: Unsplash