video_OCP2026_Meta_NPO生態系_20261015

Hey, good afternoon everyone. Um, I'm Nan here so from technical sourcing team at a meta. And um, first off, I want to express my sincere appreciation to the iPad team for inviting me to share my uh, thoughts around the near package optics. And also it's a tremendous pressure um, like to present after Jiping. And he's so technical and he did so much better job than me in tech on technical front. So I decided to not talk about technical at all. So you will you may feel bored a little bit but I'm going to share my perspectives around the industry trend. And how to build the ecosystem and why we believe NPO is a good solution for the next generation AI architecture interconnect. All right, first off um, let's set the stage by looking at the um, the key challenges operating at a larger scale. So at a meta we're serving 3.6 billion daily active users on a daily basis. Um, and also this this means that a 200 billion messages and the transported on platforms every day. And on WhatsApp alone that's a 600 million messages per day. That's a huge amount of traffics and every single change in the network architecture and interconnect topologies and that will have a significant impact um, in our like hardware design and also ecosystem engagements. And we're super proud of this large scale but at the same time this represent um, like very significant challenge for us to build a like robust and scalable architecture to support this level of communication needs. So among the uh, family of our apps the AI has been transforming every single product that we're building. So and then just recently in September we announced the news a secure private personal agent that um, like if you haven't to that installed on your phone and please try it. It might be smarter than what you thought. And then just yesterday, uh we announced the Miuz Wonder and the Miuz Wonderful. That's the most of frontier model we are building to support our Miuz product and also other product families. So, um before talking about the optical interconnects, let's zoom out and look at the overall um like trend of the AI innovation. So, over the past a few years, I I think everybody acknowledged that um the the need for faster innovation is more and more um imperative. And um this is the report I just published yesterday for how um the each uh large model performs um in the benchmarking metrics. And as you can see that um Meta Miuz has been doing well, but my intent here is not really show how that compares to other models. I think every model is built for specific use cases, and it it it's can't just be all size fit um one size fits all. And but my um messaging here is actually to demonstrate how fast we need to innovate and to support the future AI needs. So, in the past um the one model generation can take maybe one or two years to accomplish and then published and um now matures into be usable by the users. But right now, that time frame has shrunk to almost like months or even days. And uh as the time frame I just explained earlier, so and we released the um Miuz Park in uh September time frame, but just the one month later, we announced another iteration, and it's currently it's performing well. Okay, based on that demand, let's zoom in to a little bit um on the infrastructure needs. We started with the AI journey with building a 128 GPU cluster, and at that time, there was no back end network at all. So, everybody was comfortable with the front end network and transport those packages in uh like a uh in a reasonable speed. But sooner after that and as you guys can recall that we had a AI booming in late '22 and early '23. And in just the 2 years we scaled 128 DGPU cluster to 24 thousand GPU cluster. That's a building size cluster. And the entire building are running in a like a coherent way. And behave as a single computer interface with users. And then just a one year later we started building multi-building GPU clusters and connected was able to connect 128 29,000 GPUs together. And the fast forward to last year we were starting the journey to build multiple region cluster up to 1 gigawatts. And that that journey still goes on. So given the number of GPUs and the rule of thumb estimation is that we need six transceivers per a GPU. And some somebody may have higher or lower numbers depending on the specific network topology. But based on this large number of GPUs and the latest forecast projected for next year we're going to see over 100 million transceivers needed for the entire data center capacity. So with that a huge demand that's driving actually a couple of innovations across the entire network stack. For the front end network we need to to have a lower latency and also the faster response to react user inquiries. But for the back end how can we connect each GPUs faster and in a more reliable way? And this is a leading to our thought around what would be the right interconnect solution for the next generation AI cluster. And as you guys can can tell that we have multiple options. Like within the transceiver family we have FRO full time a retimed, the half retimed and the linear drive, and also the newly created um um the NPO near package optics architecture, and also the CPO and the OCI. So, among all of this, I think the the key um value for using the newer and cutting-edge technology is kind of like a three-pronged approach. It's the lower bandwidth, lower power, and the lower latency. And specifically for our news um application, as you can tell that uh we have a strong requirements for the lower latency, uh especially in the uh inference network, so that the user when use the personal agent, they don't feel the delay, and it's more like talking to a real agent. And because of that, we need to scale our network to 10 times um by 2030, and uh to address two specific challenges. How can we make the uh lower latency not a gate for the entire network stack? And also um uh also to save money by reducing the latency. And every single second latency in the network can cause a billion dollars of loss in terms of the uh 20 the GPU capacity. So, now zooming to a little bit um into uh the business case for NPO. So, at the meta, uh we have a very strong um use case for scale-out network for using NPO. This is slightly different than our partner hyperscalers. And uh the thinking process behind that is that um for uh scale-out network, we have a better options, like um the alternatives um the CPO and also the pluggable optics. So, in case NPO doesn't work well for time to market, we have some backup position to mitigate the risk. Uh but for scale-out network, it's under our um exploration, but we are not at the time to make the call that we want to transition the copper backplane-based the to NPO yet. And I want to highlight the power saving as a main value proposition for us for considering MPO. It's because that if you think about the overall network stack and we are going down the trend of removing or reducing the reliance on DSP. But at the same time we still need to have the very reliable link to connect different accelerators together and also to make sure the network is robust. So going down that trend um I list out the the power comparison for the typical 800 gig optics on the right. So on the top you can see that normal power for 800 gig module is about 16 watts and it can go down to almost the 10 watts each for the MPO or CPU solutions. And would that be enough? We don't think so. But the key thing is we identify the 30s is the key power driver behind the scene. So it's not about other like supporting architectures. As long as we can remove the reliance on 30s we can drop the power down to a certain level that can satisfy our needs. And to meet that I think MPO is just the like a way to to roll like that. The alternative may including the hybrid tandem optics and also the co-packaged optics. And also another thing I want to highlight here is in most of the MPO talks you might not see that co-design work going on between the switch vendor and the MPO partners. And this is a significantly challenging for going to the next generation architecture. And because I'm working at the sourcing so the biggest challenge on my side is that the optics vendors always care about optical module performance but when the optical module being designed in a switch system and the switch partner usually come back with some challenges regarding its integration or signal integrity or link reliability. So, in the um pluggable world, uh I think this is manageable because we have a pretty good standard to define the boundary conditions and to make sure the pluggable can be interoperable. But, moving to MPO, um I think as the Jiping mentioned earlier, we don't have that standard yet. So, the the market and also the customers demand a close collaboration between the switch vendor and the optical partners to work together to optimize the link. Especially um when you think about how to optimize the the power, it's not only the the optical vendor's job. It's the system optimization work need to be done. And in order to achieve that, we need to make sure the linear drive can power enough uh reach from the ASIC to the front port and also to um meet the standard IEEE standards. Um and on the the other side, I think I got a lot of question on the sourcing side when when I uh go into a meeting with the suppliers, and they usually ask me, "Hey, what's your um like expectation uh between different technology choice? Is that a pluggable or half-re-timed or MPO or even CPO or COI?" And my typical answer is that there's no kind of either-or choice. So, um each technology here has its specific use case, and we don't really think they are competing with each other. Because let's say for scale-up um application, and uh the MPO and CPO and also OCI uh it's not mature enough for us to completely replace the copper backplane. But, for um the scale-out network, because we have the bank up there, and we can take a little bit forward and take some risk. Um and as long as we have the backup plan available. This is why Meta has made a decision to roll out MPO to scale-up uh scale-out network and as a pilot platform. Um and also the other thing um we are on a journey to reduce the deep uh dependency on the 30s. And at this moment, I think for the LPO, MPO, or CPO, we can remove the 30s on the DSP for sure, but still have the relies on the DSP in the ASIC. But moving forward, as we established the OCI with the industry partners, so the North Star for us is to completely remove the 30s in the optical link and use optical fiber for the high-speed interconnect. So in a nutshell, we believe MPO is the right solution for the next generation scale out architecture and just as a start and in considering balancing the power efficiency, the open ecosystem benefits, and also the data and trade-offs. And we encourage all the ecosystem partners to support that direction and happy to reach out if you have any questions. Yeah, that's really wrap up my talk. Thank you.