Expert Insights: photonics & AI

Laser focused on the future of light-based technologies, this is following the photons a photonics podcast. I'm your host, Justine Murphy, an editor with laser focus world and multimedia director of Endeavor Business Media's Digital Infrastructure Group. Through interviews and discussions with researchers, company executives, students, educators, and other experts, we share diverse and unique perspectives on the issues, technologies, and trends shaping the photonics community. Follow me on this journey into the realm of one of the most exciting industries in the world. Hey there! Thanks for tuning into following the photons. I'm Justine Murphy. This episode kicks off a monthly series that will examine specific topics via the perspectives of industry experts. In this first installment, we're focusing on the emerging relationship between photonics and AI. AI is maturing quickly and the required computing capacity needs to keep up. But standing in the way right now are the limitations of conventional electronic circuitry, CPUs, GPUs. So what about photonics? Could photonics and AI someday be a match made in heaven? There are a lot of opinions and perspectives when it comes to the AI-photonics relationship potential. So I chatted recently with two experts, each working in this realm, to get their thoughts. First, James Regan, CEO of Oriel, a London-based startup focused on photonic networking. He says the speed and efficiency of optical computing continues to surge, making it a solid candidate for AI collaboration. But optical circuit switches, which is computing technology that facilitates networking by a routing light-based data paths directly to a destination without converting them to electricity, are not quite ready. Right now, we're going through a whole of transitions as the amount of data that's flowing around in AI networks just goes through the roof. And if I look at the way that we built networks or in a many years ago, we had copper pipes going to electrical switches that then switch data around from point to point. And so what we're seeing is as you observed the transition of photonics into these systems. If we think about the pipes between switches, the physics are driving that, right? So as you increase the the board rate, the kind of the impairments that come across the copper tracks, the loss that comes in the copper tracks comes as the square of the board rate. So the faster you go, the shorter distances that you can go with copper. And the earlier you transition to carrying that data it, it photonically, which is a sure say a fundamental fact of the physics. And there are many tactical pragmatic solutions as to how you do that is it's CPU, is it plugable, is it pan four, is it pan eight? You know, these are all shall we say details behind that fundamental transition. So far, that has been the transition to photonic pipes to electrical switches. We are beginning to see the transition into photonic switching. So we're aware of the arrival of OCS. So you know, Google has been deploying OCS in the networks for quite some time. There's a lot of enthusiasm around the idea of optical switching. The challenge we have is that optical circuit switches as they just today are quite slow and they're quite disconnected from the compute networks. If you think about what is in a electrical switch, what's in a, you know, the kind of product that you get from an aristroism like that, it's not just a crossbar switch. There's a whole lot of system features which is enabling that switch to be functional in the network. And so what we have emerging first wave is these OCS's which are simple crossbar switches and they enable us to do, should say, reconfiguration of the network. And now, Oriol itself is taking data from source to destination via that switching. We call it a pure photonic network, rather noptical network, pure photonic network. And we are doing that because we're able to switch photonically very, very, very fast at the network level and implement all the system features that required to use that. So we're managing, we're basically pulling data out of the memory of one server and putting it into the memory of another server across a pure photonic network. So are there disadvantages? Only that it's new and that it's different. People just have to get over the fact that it's different, it's going to be different, but it's going to be way better. The more we can use light, the more we can use photons of the tonics, the faster it will be the low latency, it will be the low power it'll be. I also got some perspectives from Vikram Vithulia, CEO of Severs, semiconductor. He says the photonics AI revolution already exists in AI data centers and an increasing potential is already there too. If you just went into a data center today and he started walking into the data center rooms, you'll find racks and racks of servers. These are compute engines that are doing all the computation and then they also store stuff. Already there is optical connectivity inside the data center today. And that's been there for many, many, years and that is mainly preserved for longer distances within the data center and also in between data centers. So within the data center, it's not for the shortest of distances, but let's say you need to go from one rack to a rack that six racks over 10 racks over 15 racks over. They all rely on optical connectivity. And what is that connectivity? You might hear the name plugable transceivers. Plugable transceivers are simply optical modules that convert electric signals from these servers and transported as light between those servers. So that already exists and has been there for many, many years. The speed of those optical links have continued to increase. They used to be long time back just 10 gigabits per second, 40 gigabits per second. And now most of the world is in the 408 100 gigabits per second optical links. So it's gotten faster and faster as these server racks get higher and higher in compute abilities because they need to move the data as fast as the data is being created to enable these optical links. You need a light source and that light source is a laser. But yesterday's technology is having limits. So what the industry has decided is we don't want the laser to supply the light and also be able to switch itself on and off. We just want the laser to supply the light. We will use silicon photonics to switch the light on and off because silicon circuits can be much faster. So that's what's happening in the existing optical links inside an AI data center. You need them. They need to get faster, but they now need different type of technology to make them faster. The company like severs, we didn't play in the yesterday's types of lasers, but the lasers that are needed from now on, they're called continuous wave lasers, but the fundamental need is they need faster optical networks and yesterday's technology doesn't cut it. In an AI data center, you'll see copper and you'll also see optical for connecting these compute engine. Now as the compute power gets even higher and higher and the compute speeds get higher and higher. Copper is cheap, but copper heats up. Copper is banned with this lesser than an optical bandwidth. They want to see can they even replace copper the remaining copper in an AI data center over time and go all optical. That's the dream scenario of all optical in a data center. As time has gone and the speeds have increased more and more optical network has taken the place of copper and where you see the remaining copper is only in the nearest distances within a data center. And even there optical is saying, I can be 20 times faster. I can provide you 10 times lower energy per bit transferred. That's the push for optical throughout a data center. Now that is pretty revolutionary. So if you think about it, the current market that's already available for a company like Severs is making the existing optical links faster. We are developing technology to solve that. At the same time, we're also partnering with customers in the ecosystem for this ultimate revolution co-packaged optics. Seepio is the way they want to get to zero copper in the future. Seepio timelines depends on who you ask, which is why for us it's important to have a leg in the existing market that needs innovation, but also be ready for the revolutionary upgrade years down the line. Plugable optical trans Severs are also in the mix. For us, plugables is now. So we are actually working with customers in our pipeline to help them even with ramps next year and so forth. My opinion is plugables are here to stay for 10 plus years. So with plugable, there's going to be the biggest chunk of the market in my opinion for the foreseeable future. It really comes down to manufacturing capacity and availability of raw material. That's where the main challenges. I think in five to seven years, people have solved how to build these things, whether it's Seepio, NPO, etc. And if I were to look back five years from now, what I would see as successful is the routes to build these things, the routes to deploy these things, the routes to power these things are solved. Five years from now, I would say, enough intelligent minds are working on this that they will find a way because it needs this modernized infrastructure to support what's needed. So that part of it, I believe, will become fundamentally viable in five to seven years. Things would have been figured out. So what's the final verdict? If you can keep data is light, it's fundamentally more efficient. Get the data into light, keep it into light, send it across the data center of light, switch around little blocks of light. That's how photonics can revolutionize AI. That does it for this edition of following the photons. Be sure to subscribe to our podcast feed to keep up with what's happening around the industry and to listen to new and archived episodes and follow us on social media too. Until next time, keep following the photons. Following the photons is produced by Endeavor Business Media, a division of Endeavor B2B.