https://www.nextplatform.com/2024/10/21/the-money-keeps-rolling-in-for-optical-interconnects/

With the bottlenecks between compute engines, their memories, and their networking adapters growing larger with each passing generation of AI machinery, there has never been a more pressing need to shift away from copper and towards optics in datacenter systems.

In fact, AI is the killer app that those developing silicon photonics and interconnects based on this family of technologies have been waiting for. For most hyperscalers, cloud builders, and HPC centers, with modest memories and modest networking demands between cluster nodes, the normal two-year cadence of doubling network bandwidth is fine, but there is an impedance mismatch with the PCI-Express peripheral bus taking three years to double. Not only that, it takes a year or two to get a new PCI-Express bump into production.

This has been annoying for those who make system interconnects, but the real problem is that we need an order of magnitude increase in bandwidth in and out of compute engines and memories and into network interfaces if we are to keep AI foundation models fed well and running efficiently. Otherwise, we are just throwing bad money after worse.

To put it bluntly, at the very least, we need the bandwidth of PCI-Express 8.0 today, which should weigh in at 1 TB/sec of bandwidth on an x16 duplex card when the first devices appear on the market sometime in 2029 or 2030. (The PCI-Express 8.0 spec if expected to be ratified in 2028 or so, very likely using 256 Gb/sec signaling and maybe some advanced form of pulsed amplitude modulation above and beyond the current PAM-4.)

Hence, all of the interest and investment in silicon photonics and optical interconnects that can directly tie compute engines and memories directly to each other and over fairly long ranges – the scale of a datacenter or a region.

That excitement about silicon photonics and optical interconnects is captured perfectly by the $400 million Series D funding round by optical interconnect upstart Lightmatter, which was founded in 2017. With that Series D round, Lightmatter has raised a total of $822 million and now has a valuation of $4.4 billion, according to the company. That is a factor of 4X increase in valuation with a factor of 2X increase in funding.

The latest round of funding for Lightmatter was led by T Rowe Price, with contributions from Google Ventures and Fidelity Investments; Hewlett Packard Enterprise and Lockheed Martin have participated in prior rounds as have a number of other venture firms all eager to capitalize on the AI boom.

We first talked to Lightmatter when the company dropped out of stealth mode in 2021, detailing its Passage optical interposer and its Envise matrix math accelerators. But it is the Passage interposer, which we talked about in more detail earlier this year, that may be the more valuable technology that is created by Lightmatter and that will perhaps propel it to an initial public offering. Perhaps after a Series E round is accomplished sometime in the not-too-distant future.

Lightmatter is the current darling of the next wave of silicon photonics, and is raising the most funds. Ayar Labs has raised $219.7 million in three rounds, Celestial AI has raised $338.9 million in three rounds, Eliyan over $100 million in two rounds, and newbie Xscape Photonics $57 million in one round as it just dropped out of stealth two weeks ago. This is by no means an exhaustive list of silicon photonics startups, and of course chip design houses Marvell, Broadcom, and Intel have big silicon photonics projects as well.

The odds are that it will take a number of different approaches to break the tyranny of copper interconnects, and it stands to reason that many of these startups will either sell their technology to the hyperscalers, cloud builders, and compute engine and networking chip makers or be acquired by one of them seeking to get and keep a technological edge over its rivals. Nvidia can be as acquisitive as it wants to be, and so can Taiwan Semiconductor Manufacturing Co, which might want to get its hands on the Passage technology.

Lightmatter co-founder and chief executive officer Nicholas Harris got his PhD at MIT and did a doctoral thesis on programmable nanophotonics for quantum computing and AI compute engines. Chief scientist and co-founder Darius Bunandar got his PhD at MIT as well, in this case in physics and studying quantum computation and communication with nanophotonic circuits. Co-founder Thomas Graham was an investment banker at Morgan Stanley and a product manager for various projects at Google. And now they are joined by a new chief financial officer, Simona Jankowski, who was a managing director at Goldman Sachs who jumped to Nvidia to be its CFO between 2017 and 2024. If you plan to be counting a lot of beans and getting ready to go public, Jankowski is one of the few people who have been on a rocket ride like Nvidia.

t will be interesting to see if Nvidia uses Passage to do away with copper interconnects and NVSwitches in its future optically interconnected systems. This is the one deal that you can bet all of the silicon photonics startups are chasing.

But for all we know, Nvidia is doing its own optical interconnects, or licensing technology to make its own. We will have to wait until the “Rubin Ultra” generation of GPU accelerators in 2027 to find out, if the rumors are right. And that gives Nvidia’s growing list of competitors – including the hyperscalers and cloud builders with their homegrown AI chippery – a chance to adopt optical technologies ahead of it and try to steal away some AI market share.