video_OCP2026_Lumentum_CPO與NPO光源_20261015

Okay. >> Forward and backward >> All right, thank you. So, I was just going to start. Um if you think being the 10th speaker is hard, try being the 11th speaker. >> [laughter] >> On CPO and NPO. Uh so, all kidding aside, uh great to be here today. Just, you know, huge thank you to the organizing committee for letting us, you know, come in and talk about some of the work we're doing on optical sources uh for CPO and NPO. Uh my name is Matt C Sack, and I'm the uh CTO of Lumentum. Okay. So, you know, I think we've all seen a picture of many speakers have shown something like this um during the session today. And basically, it's just a diagram of your typical kind of AI data center. But, I think [clears throat] in order to have a discussion about CPO and NPO and sources and things is you you really need to understand where these things are going. You know, which which networks are people deploying CPO, which networks are people deploying uh NPO into, and sort of like why is this happening? So, let's just kind of quickly go through this, and then uh we're going to talk about where some of these different packaging form factors uh actually fit inside of the network. So, I think you know, we we know kind of the front end, right? The front end network is where you're downloading all the information into. Uh many speakers have talked about, you know, the back end or the scale-out network. That's the one that's connecting accelerator racks together. Uh we're all familiar with the scale-up network. That's connecting accelerators. And then, of course, there's uh you know, emerging applications where people are thinking about connecting accelerators even inside the tray. And that's the scale-in uh application that people have uh started talking about. But, as I was saying before, use this advanced thing. As I was saying before, how do you think about CPO and NPO across some of these different networks? So, what I've done here is I've just taken kind of some of the standard packaging form factors were, you know, relatively familiar with. Uh and some of the different networks and tried to kind of align them, so you can kind of see where they all, you know, where they all play. So, starting with pluggables, I think we all know kind of pluggables. Pluggables, you know, they have a a DSP inside them to retime the signal coming from a switch. You have your optics that sit inside of there. You know, pluggables have a long history inside of the front-end network. Obviously, that was one of the first places they were deployed at scale. Um and front and pluggables actually are not just used in the front-end, but they're also used in the back-end. Uh but of course, you know, as power goes up and as, you know, the deployment size associated with some of these networks gets bigger and bigger, uh people look at ways, okay, how can I save power? How can I save cost? How can I increase my faceplate density? Um and so they turn to solutions like uh CPO uh and NPO. And you know, if you look at really what's going on in CPO and NPO through a pluggable lens, essentially what people did is they just opened up a pluggable, they took the optics out, they stuck the optics next to the accelerator or next to the switch, took the laser, put the laser in an external uh box at the front panel, uh and they removed the DSP. And of course, removing the DSP has a lot of advantages. Helps with cost. That's one of the most expensive elements inside of any pluggable transceiver. Also helps with power. Uh it's one of the things that burns the most power uh inside of uh any pluggable. And so, you know, if you think about it, one of the most natural places to put this uh is kind of in the scale-out network. And one of the reasons is because uh from a deployment standpoint, from a type of pluggable transceiver at the back-end is the most um it's the most homogeneous. It's the one that uses the most common type of transceiver, so you can have the same PMD at your switch, the same PMD in your rack, and you can kind of connect everything together. Now, I know this is about CPO and NPO, but I do want to introduce this other term, and a couple of other speakers have also talked about this. I think it's really important to distinguish where things are going, right? And this other term some people talked about IPO. I'm going to talk about it in terms of OPO. So, on-package optics. And the difference between on-package optics and something like co-package optics is really the use of serdes. So, even in CPO, right? Even in some of these NPO pluggable form factors, basically, people are still using high-speed serdes to connect together between an accelerator or between a switch and the optics. And this has a lot of advantages, right? I mean, optics can be pretty fast. You don't have to do any retiming. Some people have shown some very nice results in terms of being able to get good uh good bit error rates. So, it's a good solution, but ultimately, if you really want to get to things that have very, very, very low power, you want to go wide and slow electrically. And in order to do that, that's where the optics comes kind of onto the package. And that's what we're calling uh OPO. And then, from a network standpoint, kind of the most convenient place to use something like that would either be uh eventually in scale up, although you could use CPO and NPO also in scale up. Or you could use something like this for uh scaling, where you're starting to use some uh high-density wider and slower kind of uh source technology. Okay. So, let's move on, and let's talk about CPO. So, uh Lumentum and Nvidia announced a partnership back in 2025 to enable their co-package optics Spectrum-X switch. That's kind of here on the slide. You can kind of see the co-package engines that sit around the switch there. You can see them kind of on the top uh and on the side. And a big part of that CPO switch was the laser source. Um when you think about CPO and you think about laser sources, actually the key is economics. It's how do you get as many photons as you possibly can into that external pluggable to take the contribution of that external pluggable laser from a power and from a cost standpoint as low as you can possibly get. And so when Lumentum came out with its ultra-high power laser technology, its 400 mW laser technology, obviously that's very attractive because that can produce a lot of photons in a really small uh in a really small area. But one of the downsides of co-packaged optics, as some of the other speakers mentioned, is it's only available from a few suppliers. And this is where I think uh NPO really has an advantage. Um you can see the number of people that have contributed to NPO. I think this is a really good uh activity, really helps bring a lot of the benefits that CPO offers into the a kind of more broad supplier base. Uh Lumentum is also a part of uh the NPO, the Open CPX uh MSA. I think this is a really positive, really important thing because it allows multiple suppliers again to play in this uh co-packaged optics story. Okay. So what kind of laser sources do you need for NPO and CPO? Well, kind of the story is the same between CPO and NPO. It actually doesn't matter which one you talk about because the drivers are the same. Basically what you're trying to do is you're trying to use a laser, as few lasers as possible, to share among as many channels as possible, whether it's in an external supply, whether it's in an NPO module, whether it's in an external, you know, box which is feeding into an NPO module. The economics and the drivers are all the same. So that means technology like Lumentum's ultra-high power 400 mW lasers uh very well suited for technologies like NPO. It's also very well suited for technologies like CPO. So, you can kind of see some of the performance of the device here. Again, it's a 400 mW device. It has greater than 20% wall plug efficiency. It has really good linewidth, which is really important for certain types of modulators. It has very good relative relative intensity noise. Uh it's becoming a proven kind of industry benchmark for ultra-high power lasers for these co-packaged and now uh near package optics applications. But, we didn't stop here. Um of course, we've continued to innovate on this platform for CPO and NPO. Um the first thing we did, you can imagine there's kind of two vectors you could go from a laser standpoint. You can either increase the power driving the economics and driving the number of splits you can do from a laser standpoint, or you can go broader and you can do more wavelengths per fiber like things like the OCIA MSA is going after. So, the first thing we did is we demonstrated a more than 800 mW. We call it a super high power laser. We demonstrated this uh at OFC this past year. Uh it has very similar characteristics to our 400 mW laser. It has a very low linewidth. It has very good RIN. Very good side mode suppression ratio. It's mode hop free, which was mentioned in some of the other previous talks and it's also compatible with the ELSFP uh form factor, which is great for kind of CPO applications. In addition to the 800 mW, of course, like I was saying, the other direction you can go is you can go wider. You can add more wavelengths. So, also at OFC this year, we demonstrated a 16 wavelength uh 200 GHz spacing uh laser uh technology where we can take our same ultra-high power laser and we can make 16 of these with very very high accuracy on a very tight grid. So, here we're showing around 300 50 mW of output power. Um the grid adherence is actually very very good. It's plus minus 25 GHz compared to what we were trying to target. And again, you know, continuing to look at sort of standard pluggable form factors either for ELSFP or other applications. Um this can fit inside of the same uh same form factor. Okay. Uh finally on the laser side, you know, of course, once you have multiple wavelengths, once you have ultra-high power, you can now start to take these technologies and you can shape them and target them at specific applications. And so, this was a demonstration actually we just showed at ECOC this past year. Uh this shows our um same ultra-high power laser uh source technology but specifically targeted at the OCIA MSA. Uh and actually all those curves, which are in the background kind of overlapping with each other, these are all different lasers. I think there's something like 20 lasers all overlapped uh in this plot. So, it's very very accurate laser technology. It's very well suited to applications like OCI. Okay. Uh I'm going to shift gears here just for a second um because we're talking about NPO and CPO. And one of the things I'm going to bring up and I think one of the very interesting things about NPO is it's almost like technology agnostic. You know, you have a box, you can stick any kind of technology in there you'd like. And one of the projects that Lumentum has been working on for quite some time is this ultra-high density interconnect technology based on 1060 nanometer Vixels. And actually at the show here at OCP, we have a a live demonstration of our 1060 nanometer laser technology with uh Astera Labs uh running PCIE uh Gen 5, eight channels of PCIE Gen 5, and uh PCIE Gen 6 uh live at the show. So, I would encourage you to go and check it out. It's a really exciting uh continued demonstration of that technology. Uh we've done a number of demonstrations using this. We did some at OFC and ECOC as well. So, we're very excited to continue to showcase uh some of the progress here. Okay. Thank you very much. >> [applause] >> I mean, if you if you think about it, right? Like typically, uh you know, in a pluggable optic, people will do maybe two channels with like 75 mW. Like that's pretty That's pretty common. So, with 800, I mean, you can just do the math, you know? It's a lot. It's It's really It's a really good technology for, you know, like I said, really kind of driving down the complexity associated with some of those NPO packages. >> What's the insertion loss between 800 mW >> Insertion loss, it's I mean, you can split on the chip, right? That That's something you can do. So, you know, as long as the the splitters have pretty low loss, right? You can continue to, you know, increase the channel count. >> Yeah, Matt, how big is that chip for 800 mW? >> Ooh, good question. We don't disclose that. >> Okay. >> [laughter] >> Everybody always wants to know that. >> Okay, so that's going to hold on to the next speaker. Thank you, Matt. Thank you. >> Yeah.