video_OCP2026_GenuineOptics_NPO論點_20261015

Good afternoon. So, uh number 10 on the list of NPO presentations. I wonder what I can possibly say that would be new at this point. Uh this going to be a bit of a challenge. Um if you look through my presentation, I probably took two slides from each of the previous presentations and made my presentation, so apologies in advance if it's something of a repeat, but uh um hopefully you know, I'll at least express a point of view here. All right, so we all know the challenges about power, right? And that's clearly an issue, and that's kind of initially the thought process behind a lot of the NPO, CPO thinking was, "Okay, we got a power problem." But the reality is, you know, something like 5 to 10% of the data center power is in the optics, and it's not a huge number, not compared to say the GPUs or something like that, but it's important. Signal integrity and bandwidth, obviously that's another one uh that's going to be a problem, especially as we get to higher and higher bandwidth off the chips, and how do we route that bandwidth into something useful. I want to talk a little bit about physical space, of course, and cable density, and all the issues that come with, you know, massive amounts of copper going to the top of the rack. Latency is is an issue, but I'll kind of point out that we've pretty much already solved that issue. Um and it's just a question of of these other things that come up to problem. But I think when it finally comes down to it is, where can we scale fast enough to be important? Um and the issue is, if you look talk to Andy Bechtolsheim, he says this next generation needs a billion 1.6 transceiver equivalent units in something like the 27, 28 time frame. I think those are a little bit bigger numbers than that uh Vladimir had earlier, but it's still in the same ballpark, and I think, you know, we need to be considering not just can we do it technically, but can we do it at that kind of scale and volume? All right, so looking at what the solutions that people have been talking about, of course linear optics has been around LRO LPO for a little while now. Uh you know, compared to fully re-timed optics, we get substantially better latency as well as the power consumption. So, power consumption of an LPO pluggable module is something on the order of 6 to 8 picojoules per bit today. Um genuine optics is actually kind of a leader in this area, so we we will have 800 gig and 1.60 LPO modules available very as as the market moves forward. Um with XPO, we're also a founding member of the XPO coalition. You can see our demo in the booth right now of an LRO module. Uh same kind of performance numbers in terms of power and in terms of uh latency, but a big jump in terms of front panel density. I think one of the earlier speakers mentioned this. This is going to be a key problem as we try to again compress the switch um front panel fiber density as much as we can. However, when you switch to NPO or CPO, now you can start to use some really compact uh switches. MMC-type connectors or sorry, optical connectors. MMC-type connectors can get us down, you know, to in some cases as much as 8 or 9X compared to 1X from the OSFP kind of of numbers that we see today. And I think, you know, that that's actually a crucial point to be considering. I think Peter mentioned this. And it's uh you know, as we as we scale further along, the ability to put more and more inside a package and ability to um manage the 1U rack as efficiently as possible is going to be key. What's ready for prime time, right? What can we do today with the ecosystem in place? LPO, obviously we can do that. It's pretty minor changes to the existing uh ecosystem for pluggable modules. XPO a little bit more of a change, but still very high reliability, availability, maintainability, a little bit of technical novelty. The ecosystem's ready today to take care of that. It's the same kind of technology we put in modules today. NPO a little bit of technical novelty, a little bit of change. Now we're talking about maybe 2 and 1/2 D or 3D packaging. We're talking about some question about the ecosystem readiness, but by and large it's still a pluggable module and it still builds the same way we build our pluggable modules today. Not a lot of difference. CPO beautiful technology. I've been working on it since October 21. So I know. Um but the amount of work required to get an ecosystem standardized around that is still huge, right? You can buy proprietary systems today, but can you buy something where I can mix and match with vendors and suppliers in the process? Not today and not for a long time, I think. I actually work on the APC as well. We're trying to standardize this, but we're a long way away from it, too. I do want to put a little shout out to OpenCPX. They're kind of addressing a lot of these problems by making it pluggable on the substrate. So maybe IPO, I think somebody else had that kind of idea as a as the approach as a name for that approach. So it's kind of a split between NPO and CPO. I think that's great. Um but ultimately is it scalable today within the current module capacity limits, but expansion will present an issue. >> [clears throat] >> So NPO optical engines uh where are we in terms of the technology? It's completely agnostic, right? I can build an NPO with any internal technology I want to. Uh Cypho is the ones that Jenway typically works on. Uh we also have some Vixel technology which we will probably apply to it as time goes forward. Um the MSAs exist sort of. So, Open CEBPX exists, of course. The OIF old OIF packaging standard exists, but we're still in the process of developing real standards for this. I think that's crucial, but it'll help the ecosystem develop once we have the standards in place. Um and as we several have mentioned, external lasers, integrated lasers, uh one issue with external lasers is it occupies the front panel, and I was really happy to see somebody mention XPOELS cuz I think that's just absolutely something I've been thinking about as a great idea. So. All right, connectors, right? So, [clears throat] we've talked about a lot about connectors here. Um there is a a sort of a geographical divide. The LGA type approach is seeming to win today in Asia, especially China, whereas uh we look at CPX and Mez connectors as doing better in Europe and America. Um I think [snorts] there's reasons, pros and cons for both, and I think, you know, it depends a lot on the individual use cases as to which makes the most sense. What I will say is that we actually have all three. So, uh it's kind of a funny thing, but once you have the platform, the individual connectors can be changed pretty easily. So, here's our solution today. Uh you can see this in our booth. We have 6.4T MPO running. Uh we're using a mirror Mez on this version. Uh we have uh basically 30 two channels at 200 gig running at less than 10 to the minus 10 error rate uh with 18 dB of uh bump-to-bump loss. So, uh this is this is a viable solution today. It'll be GA sometime early next year. Um we're looking forward to, you know, invite you all to come by the booth and check it out. So. Um but it does involve some 2 and 1/2 D packaging, so it involves some packaging innovation that which we had to start to develop internally as we went forward. And speaking of packaging, the nice thing I think about NPO in a way, we talk about it as a bridge technology, but it's also a place to test out these packaging approaches in an incremental fashion. Instead of having to do it all at once in a magical CPU package with, you know, optical interconnects on the ASIC, uh we get to try it one step at a time, work out the bugs, and then move into the next step. So, I actually think this is a really good opportunity for Genuine, for my company, to learn and expand its capabilities. This is not to say we're not working on CPO, we are, but it's not to it's not in the same situation that NPO is today. Uh we talked about standards, and I think again, this is a crucial point. Um OIF, of course, has started their standards work on uh on 12.8 um as well as 6.4 7.2T. Um IPAC is working on standards as well as this is this meeting here, probably meets OpenCPX we just heard about. Uh in Asia, they're working on what they call UPO. Uh same kind of idea though, so ultra performance optics. Um this is cool crucial, I think. Once the standard's in place, then ecosystems can develop, people can have standardized connectors, standardized FAUs, standardized uh even packaging systems and things like that, and you know, the the the magic happens when everybody can do it, and then we're all competing for the same same sockets. So, here's our road map. Um as I said, today we're a leader in the LRLPO space. Uh we have 1.6T sampling of both today. So, if you're interested, um much commonality between that and NPO versions, right? So, we have an NPO version today that you can see in the at 6.4T is based on the Mez connector, but we'll also have CPX and LGA versions very shortly. Uh and we also have 12.8 TXPO working in the booth again if you're interested. Um and in that case, it's an LRO implementation, but we'll be doing an LPO implementation very shortly as well. So. Um Anyway, summary, I think NPO provides a lot of the benefits. Very little uh is missing from that except maybe something like 1 to 4 picojoules per bit uh kind of difference in power. Um some It utilizes much of the existing supply chain. Right? The people who make PC boards, the people who package this stuff, the people who attach FAUs, it's all the same kind of stuff. So, there's not a lot of uh invention required in the supply chain, which is good. Uh we can support a variety of formats. We can support OCI if we want. We can support uh um multiple DWDM or or CWDM standards as well. Uh and the standards that it do exist to date for the for the packaging are becoming more robust and enabling a supply base in a better competition. So, that's all I have. Thank you. >> [applause]