Lightmatter Inc., a startup with technology that allows chips to exchange data using light, has raised $154 million in new funding.

The Series C round was announced this morning. It included contributions from the venture capital arms of Alphabet Inc. and Hewlett Packard Enterprise Co., as well as a number of other institutional investors. The round reportedly triples the undisclosed valuation that Lightmatter received after its previous raise in 2021.

Many modern processors are made of chiplets, compact compute modules each optimized for a specific set of tasks. Those modules exchange data with one another in the form of electrical signals. Lightmatter has developed an optical interconnect, Passage, that allows chiplets to share data in the form of light rather than as electricity.

The company claims that its technology provides up to 100 times more bandwidth than traditional alternatives. As a result, it enables a processor’s chiplets to exchange data with one faster than would otherwise be possible. Speeding up the movement of information in a chip increases application performance.

Lightmatter’s Passage interconnect is a flat, rectangular piece of silicon made from an entire wafer. Up to tens of chiplets can be placed atop the wafer to create a processor. The data that the chiplets send to one another travels through Passage in the form of light.

Passage transits light beams via compact, embedded channels called waveguides. Lightmatter says 40 of its waveguides take up as much space as a single optic fiber. Passage also includes various other components, including transistors that help coordinate the flow of data traffic.

There are already optical technologies on the market that allow chips to exchange data using light. However, they require attaching bulky cables to each processor. Lightmatter says that Passage takes up significantly less space than traditional interconnects and uses as one-fifth the power.

“Rapid progress in artificial intelligence is forcing computing infrastructure to improve at an unprecedented rate,” said Lightmatter co-founder and Chief Executive Officer Nick Harris. “The energy costs of this growth are significant, even on a planetary scale. Generative AI and supercomputing will be transformed by photonic technologies in the coming years.”

According to the company, another benefit of its technology is that it reduces network complexity.

A cluster of chips such as graphics processing units can be linked together in many different ways. Some network configurations are significantly faster than others. To maximize performance, companies often customize their chip clusters’ networks to each workload.

With traditional interconnects, the cluster setup process requires engineers to manually connect hundreds of fiber optic cables. Lightmatter says that Passage automates the task. The interconnect provides the ability to change network configurations in under a millisecond.

Lightmatter offers Passage alongside a server called the Envise 4S. It’s optimized to run advanced artificial intelligence models. According to the company, the system can complete three times more computing instructions per second than Nvidia Corp.’s DGX-A100 machine learning appliance.

The Envise 4S is powered by 16 custom chips that each comprise 256 cores. Those cores are based on the RISC architecture, a popular open-source blueprint for creating central processing units. Lightspeed’s chips also include optical components that reportedly carry out calculations by splitting and mixing beams of light.

Alongside the custom silicon, Lightmatter says that each Envise 4S includes two off-the-shelf Advanced Micro Devices Inc. processors. Data is stored in three terabytes of onboard flash memory.

The company will use the capital from its new funding round to support the commercialization of Passage and Envise. Additionally, Lightspeed will work to expand the adoption of Idiom, a software toolkit it offers in conjunction with its hardware products. The toolkit makes it easier for developers write applications that can run on the its Envise servers. 

Image: Lightmatter