video_OCP2026_NVIDIA_SpectrumX多平面光學_20261015

Dennis, are you still here? I wish you hadn't mentioned you were the 10th speaker cuz that made me very self-conscious that I'm the last speaker. But, in Major League Baseball, the last speaker is the cleanup hitter. And I feel that when it comes to NPO, there is a few things that need to be cleaned up. So, in our environment, workloads are getting more complicated. AI factories are getting much bigger. The network and optics determines performance and scale. The data center is the computer and the network defines the data center. I want to set the stage of what we're doing and a gentic AI is the most complicated workload that we've seen. It pushes um all of our hardware to the limit and it pushes computing to the limit. It's the most complex computing workload in history. Um Nvidia's AI platform and Nvidia's AI flagship AI platform is Vera Rubin. Vera Rubin platform brings together the Vera Rubin CPU tray, the um the Vera Rubin GPU tray and switch tray, our um NVL 72 compute tray, uh our um Bluefield 4 DPU tray, our Spectrum 6 switch tray, and then the newest thing that we're bringing to the lineup is what we got our language processing unit from the Grok team as our LP 30 tray. Uh we put all these together into a Vera Rubin pod. These components are all designed to operate uh seamlessly together to create a one incredible AI supercomputer. This Vera Rubin pod is 40 racks. Of the 40 racks, 10 of them are our Spectrum X uh switch. Our Spectrum X switch is an Ethernet switch that has co-packaged uh optics that was purposely built for AI factories. Our CPU chip is based on micro-ring modulators uh and it's currently in production. It uses 3D stacked silicon photonic engine that's based on the TSMC CoWoS process. We use high-power, high-efficiency uh external lasers, and we use detachable fiber connectors. By going to CPO, we're using 1/4 of uh or 1/4 of the discrete components. We have 5x lower power consumption, and we have 10x higher uh mean time between interruptions. But, co-packaged optics is not the only thing that we're going to be offering our customers. We're still using a lot of pluggables, and in particular uh we're using uh OSFPs, which are still the workhorse of our industry. In our Vera Rubin NV72 compute tray, uh which I'm showing here for example, we uh have eight uh OSFP modules. Uh the other thing that we're doing is uh we are also offering our Spectrum 6 switch in a fully pluggable uh configuration. So, here you see the Spectrum the version uh that has uh fully pluggable uh OSFP modules. But, what succeeds OSFP? Well, one thing is we're going to look at stuff that's 400 gig capable. We're going to look at things that have more channels in a pluggable form factor. And so you probably all saw yesterday's announcement that Cisco and Nvidia announced the USFP, the ultra dense SFP form factor. Um they're we're co-designing co-defining the form factor and we're driving the MSA. We feel this will bring four times the density in a pluggable form factor and reduce the power consumption by at least 30%. It's going to be both 200 gig per lane and 400 gig per lane capable. It'll initially [snorts] start with 16 lanes. So it's a 6.4 T module. Uh and then later we'll have a version with 32 lanes that will support uh 12.8 T. Our partners for the connectors in this are Amphenol and TE. So one of the things that I get from people from time to time is this fear over doing CPO uh because it's not socketed. There's some notion that if everything were socketed uh it would somehow be better. And I think that's a fallacy. Um one of the things I want to point out is Nvidia built this Quantum InfiniBand switch. This was socketed CPO. So we had six of these things. They were pigtailed, they were socketed. We're currently building the Spectrum-X switch. It's got 32 CPO engines. They're all soldered down. Okay, so we sort of been there, done that. There's good reasons for doing sockets. [snorts] I mean, I think they're good for initial prototyping. Um Was it Zping from uh Alibaba pointed out that actually sockets are helpful in manufacturing. If you have a defect, you can quickly change it in manufacturing. But I think that's as about as far as it goes. There's a lot of negatives for sockets. They add cost. They add area. They impact your signal integrity. They impact your thermal solutions. Soldering is a better choice, in my opinion. It's great for signal integrity. It minimizes your footprint. It's the lowest cost solution. It's also the hardest to implement. And limited rework is kind of a euphemism for no rework. Um but uh but these things um apply to LPO as well. Um And one thing that socketed does not mean, socketed does not mean field replaceable. If you think socketed means field replaceable, you're kidding yourself. You don't know how these boxes are put together. And this is the same for NPO. So, NPO um I think it's near packaged optics or it's also no pluggable option. Um What Nvidia sees is many of our customers are exploring NPO. And we will work with um our ODMs to enable NPO. We will also support uh NPO through our uh NV fusion NVLink fusion uh efforts. I do feel that um CPO pluggables and NPO will likely coexist uh for scale out. But I will ask you, what problem are you really trying to solve with NPO? Okay. So, if you have no face plate, uh and you need you need more density, no no pluggable option, and you want to move stuff inside, I I get that. If you're trying to lower power because your flyover cables and connectors are too lossy, and you want to attempt to do a linear solution or a TRO solution, I get that. I wish you luck. It's harder than it looks. Um but if you're doing it because you want multiple suppliers, I think you're kidding yourself. If you look around today at all the prototypes, they're all different. If you look at the MSAs, there's four of them. They're all brand new. I think the only way that LightCounting's projections come true is if a whole bunch of people are going to go out the door single-sourced, okay? I find the multi-source aspect to be uh questionable. And then the other point I want to make, um serviceability. I'll say it again, I said it with sockets. I think the notion of serviceability for NPO is dubious at best. You you're not going to take these things apart in the field, especially when they're assembled in the factory, likely in a clean room. You're going to pack them up, you're going to ship them back. It's going to be a very expensive repair job if you have pigtail optics, and they're all woven in together, that's going to be a very hard repair job. If you have shuffles, that's almost impossible. So, I think um I think we need to be honest with ourselves about about several aspects of NPO, about the multi-source and the serviceability aspects. So, back to where I started, workloads are getting more complicated, AI factories are getting much bigger, Network is critical. Choose your optics wisely for scale, reliability, and efficiency. Thank you. >> [applause]