video_OCP2026_Microsoft_NPO延續可插拔生態_20261015

All right. Uh good afternoon everyone. Uh my name is uh Yabei Yin, coming from Microsoft. Um So, the talk today will be focusing on how we can make a MPO uh continue the ecosystem success. I I I've seen the previous speakers and many focus on why MPO. There's many benefits, uh low power, lower cost, and potentially more reliability. Uh here I I want to change the focus a little bit to focus on how, right? So, we have to first observe there's a industry ecosystem success with pluggables. It's been there. This is one one one of the reason today why today is a such a prosperous uh ecosystem. Um and then there's many reasons. I'm listing some of them. And then we'll talk about the challenges we're facing, why we cannot continue uh with pluggables. And then MPO, what are the enabling technologies? And some of call to action, uh which is uh what are the OCP forum is really asking. Um so, yeah, so I think a lot of this knowledge is a common. I won't be just reading the slides line by line, but uh we have a kind of a In summary, we can summarize categorize as a four pillars, right? So, the PMD standards, there's a First, we have to see the PMDs defining the lane speed electrical and optical, and also the reach uh is enabling we have a such a different type of optical transceivers and uh applying at a different places in the data data centers. Form factors is another thing that is bring all of us together. Industry is many suppliers uh building the different form factors. Uh there's uh Don't forget that there's also mechanical designs of those uh form factors trying to make the uh transceiver assembly process uh smooth, right? So, and the the electrical interface standards uh also the management interface standards the same is CI those call call those are there so to make the also the switch and the optical module interactions standardized. Also there's many multi-vendor specifications so we us as a hyperscaler can choose different components coming from different suppliers still making them feel and looks the same inside a switch. So this uh something that enables today's uh ecosystem of the pluggable. And then but we are facing some challenges otherwise there won't be change right so people will just continue the past dependency will just move the whole industry forward but we are facing indeed a facing challenges first summarize that is three domains thermal envelope so face plate density that's number one we are keep pushing the lane speed to higher and higher eventually we are the limited by the densities of the face plate in the switch front we are first limited by the face plate density. And then there may be a future people may be thinking like there's also the chip level pitch front density but we are first limited by the face plate density therefore the transceivers a lot of the time when we talk with suppliers they say back to us if you have to use the OSP module then you have to make the switch to our use where are you type of size right so that's the form factor on the transceivers brings the industry standard but also is limiting the face plate face plate densities. And then the module power concentrations we are also pushing more and more lasers modulators inside of the module uh to have a fast faster and a narrow type of a lens. Uh as the switch radius grows, uh we also need to put in more modules in there. So, those uh added together is counting the power uh density and also the thermal challenges. Uh and then the electrical reach resistance uh where uh once the electrical speed goes higher, signal integrity is becoming more challenging to bring the signals from the switching ASIC to the front plate front plate is uh according a lot of uh increasingly more difficult. Um and then the yeah, so bear with me for this uh you can see these AI-generated pictures uh to avoid uh copyright issues. Well, the trick to actually ask AI to generate pictures. So, the idea is close, but the uh the those the picture are not uh pretty in my eyes, but Uh so, it tells the story. So, we are trying to get the optical engines closer and closer to the switching ASIC. Uh eventually, uh I think the end goal of the industry, many people would agree to have the optical engine co-packaged with the electrical engine. Uh that's kind of a holy grail of thing. The closest can get is actually get them on the same substrate or interposer to have a co-packaged uh chip together. Uh but before we get there, there's a lot of uh challenges on the CPU, but this forum is about a MPO. So, I will not deep dive into the CPU challenges. But before we go to CPU, um getting the as close as possible, right? So, to put it in in either in a substrate interposer maybe uh it's also like even in the PCB on-board optics. Those are the options of getting the optical engines closer to the switching ASIC. Uh each of them gives uh also pros and cons, but the goal is the same. We get we get closer. Uh we have a better signal integrity to control uh and we can therefore keep going faster and then narrower. Um And then the DSP today uh uh is there because we actually have to use use DSP to do the signal re-timing and also try to correct any errors that's on the noise getting to the signal before we get into the optical domain and send it to a longer distance. Uh those things, if we get a closer if the previous goal as we mentioned is trying to get the DSP closer to the chip, uh those problems will be going away. So, the DSP whether it's needed or not, uh most of the most of the MPO solutions that I'm looking at it today, we will assume linear channels. Linear channels should work, right? But there's a still argument about it if we keep continue to drive to 400G play, uh some of the lanes maybe still facing some challenges. So, uh adding some level of re-timing into the MPO modules uh can be considered and that's also being discussed inside the industry. So, even half re-timed can be considered and then maybe not all the lanes will face the same challenges. So, those options can be discussed. And also one one thing I want to point out is the MPO also give us the flexibility of not having to decide on the same signal integrity of same uh human optical engine over the ports. Uh there's use cases we are talking internally where some of the ports maybe facing a different PMD challenge versus the other ports. Uh so, that uh consideration can be um put into the design of the system of MPO. So, uh DSP or not as needed uh is given to the option of MPO. So, one thing to also enable this is uh the because either CPO MPO, uh the topic of the CPO MPO was not new. It's been brought up even 10 years, some people say 20 years ago. Uh but then the challenge is that didn't make it a reality still once we have optical engines inside the west you do not have a front module you can remove when things fail then what is the feeder rate so this to me is one of the most fundamental issues we that's still not a getting the CPU maybe a more close reality is the feeder rate how we can actually get it down so the feeder rate of all the not only the lasers the laser more hopping shows that I heard a previous speaker talking a lot there's also the modulators the optical engines the thermal mechanical liquid cool all those components can fail right so the overall system holistic feeder rate need to be very much improved for us to get this kind of a big box of a big blast release in the network and then without having to replace it on the other bases so that the feeder requirement uh inside this center was not there previously most of my sensor saying like we can replace things when things fail but for the CPU and PO type of ecosystem we as a hyperscaler as a data center operator we'll have to develop a new kind of operation model where we do not have to replace a whole box as often as we are replacing a pluggable transceiver as before so the feeder rate of individual components and also the whole system of feeder rate need to be very much improved my mental model goes it have to be at least 10 times bigger a 10 times smaller in fit and a 10 times bigger in MTBF and then the the we can still try to shoot a target higher target goals and then then when things fail we also need to develop operation model how we can not stop the traffic and not hurt our customer experience but still try to keep the whole network running and the GPU and the CPU servers running um yeah so these are a few enabling technologies we are seeing different optical engines, silicon photonics, CML, Vexels. Uh and also we are seeing the semi-thermal management and telemetry as already there. Uh the key takeaway point of this size would be we would have want to leverage whatever can be leveraged already from a electrical, optical, and a management. Uh but we are not limited into okay, having to follow every single standard that was there before. Like a form factors in this case can be redesigned. It will be redesigned. Just do we need a new standard to standardize or we can allow more flexibility that can be discussed in the industry. I I feel an MPO CPO system will be more like a co-designed system that is a customized to each of the data centers special requirement. Um so, yeah, MPO is a is a is a bridge. It's a bridge to getting to a better future of more power efficiency and a more reliable network for the end users. But it's not a like a just a try to define another smaller pluggable different form factor. Try to get a uh get get this business still running. We We still need the industry to work together, but this one is a I see that not only just a a form factor change. It's kind of paradigm shift from a one generation to one era to another of innovation eventually get to the most reliable and the most power efficient. In other words, planet friendly type of infrastructure for the overall sustainability of the technology growth growth and AI growth. All of this driven by this some money demand of all the AI usages. Those AI we are seeing them indeed helping us improving our daily life. So, this demand I see that it will continue and the MPO is a bridge for us to get to the future of a truly planet friendly type of ecosystem hardware and architecture. Thank you very much.