video_OCP2026_OIF_448G電氣介面_20261015

Okay. So, now we're going to go into a session where we'll have uh three speakers talking about the OAF's 448 uh gig electrical specification developments. So I I showed this chart earlier uh during the introduction and the key area we're going to focus on now is in the lower left corner and this is a common electrical IO work where the OAF has a history of uh developing um industry specifications that then become the basis for many industry applications. So for this session as I mentioned we'll have three speakers. I'm Nathan Tracy. I'm a senior principal engineer at TE Connectivity. I also serve OAF as president. And then Ye Tang will talk to talk to us about the margin challenges that we face with 448 gig. Uh Yei is a principal hardware engineer at Cisco. And then finally, our third speaker on 448 gig electrical will be Kathy Louu, a distinguished engineer at Broadcom. and she'll talk about the uh different techniques, the tools in the toolbox that we'll use to uh make 448 gig attainable. So, jumping into my slides, I'll provide a little bit of over overview of the electrical work that we do at the OAF and um that the how we've uh developed a process for that. get my timer going so I don't go over time. So, as this slide shows, dating back to the year 2000, OAF has a long history of developing different industry uh specifications. I'm not sure who's controlling that laser pointer. Anyways, it's gone for now. Um >> ah okay. So it maybe someone else has a uh a Bluetooth pointer that is on the same frequency as us. Who knows? Um yeah. So in in on the far right column it shows the broad range of industry applications that have adopted in whole or in part uh portions of the OAF electrical specification. So uh starting back to um you know 1.6 gigabits per second all the way up to the most recently uh uh work that we've done it's almost ready for publication is 224 gigabits per second. Um and uh you can see the uh over the over the history of how that uh work has gone on. Whoops, wrong direction. So the way that we do this and and here I'm using the the example of the work that the OAF is is finishing up right now on 224 Gbit per second is we define an industry uh uh channel specification that is optimized for different electrical reaches and this enables us to provide a power optimized solution for different applications to the industry. So looking at this example, we start at the top with the XSR extra short reach where we aimed at a 50 millimeter channel. So think of that as to two chips within a package uh an MCM type of application could be um a co-acked optical engine as well. Applications like that where short reach and we're really trying to optimize for energy efficiency. And then across the next bar we see the VSSR which is our chip to module specification. And this is where we're aiming for a 200 millimeter channel uh with another 20 mm in in the module and with one connector in the channel. So we're optimizing for that critical um from the from the host silicon to the face plate. MR is the next uh one there and that's where we're aiming for a half a meter of reach. This is uh potentially within a single line card where we have a mezzanine board or something like that, some type of a a uh co-planer uh type of architecture and we're optimized for just delivering that performance for that reach. And then finally the uh the LR specification which we could think of that as an external cable or a backplane um type of application where we're trying to go from line card to line card within a rack. And then finally at the very bottom is the new work that the OAF has taken up which is uh uh writing an electrical specification to support linear optical channels. Um in this work we recently published a 100 gigabit per second performance and now in process on 224 Gbits per second. So the methodology that we commonly use at the OAF when we come to a new inflection point in the industry, a new data rate is to gather all the industry inputs that we can and put together a framework document or a white paper uh that discusses the various different um requirements that the industry is looking for. We look at things like reach and architecture. Um what type of power budget might we have? The what's the media that we're going to consider? For example, transitioning from print circuit board traces to twin ax cables, things like that. What are the modulation options that are uh desired? What are the benefits of those modulations? The drawbacks. Um we talk about channel simulations that can be brought forward early in the discussion and what are some of the test and measurement channels that we're that we'll face and we look at what are some of those challenges that are associated with each one of those different uh directions and we come up with a white paper that we bring to market. And the value of this white paper is it gets the whole industry thinking on a common language and thinking about the problems together as opposed to each going off and solving problems on their own. And then we can talk from that we can talk about what are the different uh reaches or applications that we may be needing to support in the industry. So we kicked off this effort a couple years ago for 448 gigabits per second. We finished up the work uh last uh November. It's been published uh uh uh about a little over a year ago or yeah a little over a year ago and you can access that document uh on the OAF's website. So some of the key things that we talked about in that white paper are the architectures that we need to support and what are going to be the requirements for those different types of app uh applications. So we looked up we looked at this the unique needs that would be required for scale up and scale out for front-end networks and also for scale across networks as well. And uh we uh set about understanding what are the different key performance requirements that were going to be required for that and for those different unique applications. scale up, scale out, scale across have different requirements in terms of reach, in terms of energy efficiency, latency, uh different attributes like that. And so we want to capture that as we start this industry discussion on a next generation data rate. [clears throat] Some of the key metrics um that we looked at as we started this work were what are the number of nodes that we're going to need? We're talking about in in the case of scale up uh GPU to GPU links uh what are the the range of uh applications that we expect to be challenged with within the time frame of this new data rate uh what are the critical network properties that we're looking for such as lossless and low latency and what what are these interconnects primarily going to be used what does that application look like and so we did that for scale up we did that for scale out and here you see some the um data that came out from that in terms of where the industry was at at the time that we did the work and what were the projections that we had for that. And so from an electrical uh channel standpoint, what we talked about is the fact that we need to push copper interconnects um to to achieve these um requirements. But we're going to face a number of challenges. We're going to be bandwidth challenged, signal integrity challenged, where we have to understand every element of the electrical channel and what impairments it brings to the channel and what techniques can we bring forward to address those. And we'll talk more about that after I finish and uh ye and Kathy talk about it, the the progress that the industry is making. We also talked about some of the trade-offs of different modulations that might help us to solve those problems. But then what are the challenges that those different choices bring to us? What are the drawbacks? What are the benefits? We talked about power consumption. Increasingly sustainability and power consumption are critical needs and um uh lightning rod uh discussion points in the industry. And so how do we think about these architectures? How do we think about uh the channel requirements and and what can we do to minimize the energy consumption uh to lower the the power consumption of the links um while still delivering the performance that's going to be required. And then finally tail latency. When we talk about um a scale up or even a scale out network, it's the tail latency. It's the latency of the of the slowest uh link in in the fabric that has the the hugest impact on and how do we think about that and what are the things that we need to be conscious of as we do this holistic consideration to um develop solutions for this next generation data rate. What are the key factors that we're going to have to address as we make decisions going forward? And so based on that we came up with a number of applications that are likely to be need to be supported by industry. So these are potential target areas for the OAF to consider starting projects and creating um implementation agreements or specifications for that. So we looked at applications such as die to die in a package where we might have 25 millimeters or so of channel die to optical engine um as much as 50 mm of electrical channel to support the various different uh applications for a die to optical engine chip to nearby optical engine such as NPO potentially CPO where we might see a channel as long as 150 millimeters a pluggable module uh you Can the industry still find a way to to pave a path for uh chip to pluggable? And that would be in the order of 250 millimeters or so. Uh chip to chip within a printed circuit board uh 50 cm. This is where we're within one sled or one line card. Um uh PCB across a back plane, which is a critical application in the the fabrics that are being deployed today. And so we want to make sure that we're going to be able to address that. and then chassis to chassis within a rack. So this might be a face plate or a back plane uh application. So based on all that at at this point in time the OAF has started two projects so far. Uh one is uh for our VSSR or chip to module application and one is for LR and Kathy and E will talk further about that. From looking at this from um an implementation cartoon here, we show that these electrical channels are all centered around the the primary uh silicon device that's going to be on the host board or on the sled. And then how do we get from that uh silicon to the the other end of the link that we need to support? And here you can see that we're looking at uh co-ackage copper solutions um co-ackage optics with uh fiber optic links inside the box. All different applications that are going to need to be supported to create the the range of different industry specifications that are going to be required to enable energy efficiency. So comparing to the 224 gig road map that we looked at a few minutes ago, right now we've started those two projects uh the 448 gig VSSR and LR. Um and and then I expect these other uh applications are likely to be required as well and that we'll see more project starts there. So it's a fundamental shift in in system design. When we think about 448 gig, it brings a number of challenges. Uh, and the industry has really been working hard at that and making good progress so far, but it really is a significantly larger challenge than it has been in the past. We're trying to take a holistic approach in the OAF of understanding all of the different factors that can mutually impact uh that work and then the fact that it's it's really about collaboration more than it ever has been before. 448 gig. We need to bring everyone together into the room. We need to understand what technologies and what performance can be achieved in the package in in the die in the die fan out in in the media that we're going to be using the interconnect technologies. Um these are all going to be very critical as we think about how to solve the 448 gig challenges. And so just to summarize, the OAF has used the same process that we used at 50 gig, at 100 gig, at 200 gig, and we've started the process of developing work on 400 gig. uh and we've done that by developing a framework project by bringing the whole industry together to contribute to that white paper which you uh can access on the website and by understanding the different and unique requirements and applications that are going to be required for uh hypers scale and AI uh unique needs and this has led to us launching our first two projects and actually getting into the work of understanding the tradeoffs that are going to be required and that's where I'll stop and let E and Kathy talk further about that. But before I turn it over to them, does anyone have any questions? My timekeeper said I don't need to take any questions. I see someone going to the microphone though. Yell out a quick question. Okay, quick question. Bob Wheeler from Light Counting. Um, so there are a couple proprietary solutions out there as sort of a bridge to 448 and I'm wondering if there's been any consideration in the OAF for either uh 224 with simultaneous birectional signaling or a 300 gig per lane um solution. So within the OAF, we've had discussions around these uh possible alternative paths and at this point in time we did not identify that those were broad industrywide demands and so we focused more on doing the traditional doubling. It seems like that's where more of the industry traction and wants and needs are are focused right now. Uh certainly we're aware of some of those other I'll call them uh interim solutions that are uh possibly being considered but uh so far we have not uh identified any start uh activities there. >> Thanks. >> Thanks Bob. Okay so then next let me invite up uh Y Tang to talk about um where does the channel margin go? the down button advances. >> Uh down button. All right. Uh thanks Nathan. Good afternoon and my name is Eton Francisco. I also serve as OAF board member and um as electrical vice chair uh of the PL working group. And thanks for joining us today. So uh Nissan has set the table for the broader 44 age development and in this segment uh in the next 10 minutes I will really trying to be on time. Um I want to focus specifically on the electrical channel. Uh the central question here is where does the channel margin go as we move from 224G to 44 AG. All right. So insertion loss has not been sufficient by itself to uh describe channel performance for several generations that we all know but um that's not new for 44 AG but 44 AG pushes that require bandwidth far enough that some of the uh several effects become much more consequential. Um so right now in the screen what you you saw is actually a 22 to 4G uh channel uh model and it's actually BGA to BGA. So um I'm using this just to demonstrate some of the challenge we were saying it's not necessarily representation of 44 age development just so we're all clear there. Um so the trace shows here that these representative 22 to4GVS channel models at 2 to4G pan 4 the Nyquis frequency for 56 GHz those channels show reasonably uh controlled behavior through that region. Um but beyond roughly 80 gigahertz uh we begin to see rapid roll off and resonance. So um by the time we approach 100 GHz the channel response is no longer smooth extension. So what does this mean for 44 AG? If the 44 AG if we look at the pan 6 the Nyquis frequency is about 87 GHz. Now you would need to push a little bit farther to practical implementation. Now 44 Japan 4 um that Nyquis is 112 gigahertz uh that target is beyond the measured data showing here. So um this really shows us that to 4G channel it's not naturally you can just grow into 44G channel. There are technology developments needed. We really need to push the bandwidth uh to the uh to the one that's suitable for 44 AG. Um so one of the things we talk about it's not just lossy matters. Uh at the early generation we like say is loss DB per inch doesn't mean how how far we can go but when we get to very short UI we have very limited margin. What we found is these transition points actually consume a lot of China margin. Um what all this orange arrow shows is some of the transition points. As example escape the ASIC um the um cable transition package to CPC uh cable to module all this transition point uh would bring discontinuity in the um in impedance and there are electrical magnetic field that cannot be absorbed quickly. Uh it cause multiple reflections. um and there are cross talker challenges as well. So what are the characteristics that at play if we look at here what's beyond? So the first thing attenuation signal attenuation that's something we all are very familiar with but there are also reflections cross talk if you look at this uh model again it's very preliminary model at a higher frequency you can see signal attenuates but reflection the cross talk actually goes up increasingly at a higher frequency so the usable bandwidth becomes very limited Um and then there are other things um that less prevalent at early generation like like intraparis skill now um becomes very critical and we have one slide later to talk about that and one of the things uh application I think Nathan just touched upon is that our users are still very interested in pluggable uh optics and One thing that we know it's a limitation is a connector. The connectors has made signative advancement. If you see this diagram, you show different generation of the connectors how their uh bandwidth improves over time. Um but as you can see uh the CPC interconnect that's 2025 measurement we already passed uh 90 gigahertz 100 gigahertz round that enables um the pan 6 but we are still a little bit far away from the pan 4 yet. Uh we do believe with the speed the industry de developing we could enable pan 4. However timing also matters. I think Cassie will spend some time about the impact of these kind of channel constraints in terms modulation and other choices we make uh in the standard and the in actual application. Um so I mentioned a little bit about intraair skill. I want to mention this here because CPC connector is very attractive for our 44 AG because it bypass uh the BGA uh fun out and it has its big advantage. Um but uh cable also have its challenges. For example, the diagram shows a two to 4G channel um how the uh skill of cable actually is frequency dependent. Again, this does not demonstrate 44 AG capability. This just show us this is something we need to closely monitor that cable supplier needs to control it. industry standard we needed to specify it correctly and also the measurement technology should support it at higher bandwidth and we talk about why the CPC has a advantage and this brings up to the PCB uh escape problem we see at 44 AG this is very particular uh challenging because large geometry transition so Uh this may not become a solution for every applications but there are low rated um applications for example and with uh less crowded escape it's still worth looking into given its maturity and a low cost perspective and so I went through this very quickly but our key takeaway is that when for AG move the Nyquis target into a range where interconnect bandwidth become a force work order constraint. So um later when Cassie talk about the toolbox what we can do it's really based on what currently uh industry can offer and in the short term what it can offer. I think these two things could really connect each other uh better uh once Kathy give her her um presentation. Uh the path forward I do want to say that physical channel uh development is ongoing. there are a lot of advancements in the uh industry and that's why OAF is also trying to consolidate a lot of advancements and give guidance to the industry in terms of a specification development. So um I don't think we have time for questions. I want to okay we have we can get one question if there's any question. Does anyone have a question for Ye? >> Please come to the microphone. and yes and please stay for the presentation. Cassie will give some view there. So um I I think one question what one thing we would say is we think personally pan 4 p 6 they are both on the table. The timing would um really um probably give more guidance but Cassie will talk about that a bit more >> really quick about the intra pair skill. Um is there a direction to actually Um I think that's really a question we still need uh additional study. Uh we think that there is probably uh value to introduce additional um specification or parameters for the entry skill. >> Okay. But that's still open. >> That's still open. >> Thank you. >> Thank you. Thanks everyone. >> Okay. Let's thank you. So, next up is Kathy Lou from Broadcom. >> It is very >> No. >> Are you turned on? >> He I'm still on. I think >> it's always on. Okay. No. Yeah. Thank you, Nathan. NE it just lay out a whole bunch of uh 448 gigs electrical channel and then challenges uh and then drop a lot of tax to me as a thirdis architect. So third is always the last saver for the overall the interconnect. So no doubt moving to 448 gig is a major milestone. it require the whole industry's uh paradigm shift. So I think I want to run faster and because only 10 minutes left for me, right? So this one is uh Nathan already also used this cartoon to lay out uh the the uh landscapes of the fake gig electrical applications. There are so many techn uh technology terms like CPC, CPO, MPO, LRO, AEC, DAC and then LPO. So, but in my presentations here I just want to focus primarily on two type of uh interface. One is chip to module uh VSR. The other one is um uh copper cable uh cable back plane LR type of uh interface because all others I we can view as the special uh application case like the mixed between electrical and optical like LPO uh or the LRO half time optics or as the shortens uh the electrical uh channels um like move the optics more closer to the host die like NPO or CPO if we can uh handle uh the challenge or uh make it a possible for the VSSR or LR and others will follow I consider it's a subset so let's see uh what is the uh the two available for us to solve those channel loss at 448 gigs the top block diagram is uh DSP30s is right you can see is N front end ADC FFE DF MSE and then FAC and in the table below that is map out how the industry adopt the new technology tools from generation to the other generations like we start with NRZ and then we add DF and basic f RX f to the system when we run 11 28 gig and then when we move to 56 and 112 and 224. We um update the modulation to four levels and then we add more uh powerful uh DSP functions like a floating tap DF and MLE and also um stronger resignment code KP4 into the system. when we talk about 448 gigs and then I think it's try to um resolve those challenge we need to reopen the toolbox and then there is two major tools available to us at 448 gigs one is modulation the other is uh interfac so I'm going to talk about this one by one in the following slides so first one modulation choice it will heavily depend on the channel profile and then physibility. So we talk about two type of channels here. Channel one and a channel two. The blue line is channel one. You can see this is improved channel and uh shows a low insertion loss across the frequency even up to 112 gigahertz which serve pan 4 effectively and but the other type of channel two the right lines and it still also existing right coexist with channel one right now and it suffers a severe attenuation and then uh has a high frequency rolloff and beyond the pan four neckquist frequency right for 4 gigs. So here pen six can lower the neckquist frequency and to stay within the physical uh limit and then the modulation choice. Now I have to tell you the bad news there is no single modulation schemes can fit all the topologies we're talking about here at 448 gigs. So let's look at different topologies and then do it independently. First LR passive cable based on the current even with improved channel component. We believe um this channel right now hit uh the physical limit and then pan six have to be a good starting point and may remain primary choice for longer times for the passive cable. If we want to run one meter or even longer over passive cable and on the other hand pen 4 if we want to run the longri copper cables it may have to depends on the active copper cable technologies like AEC and ACC for a while. So that is for the LR topology. How about the second one on the top right? VSR chip to module topologies. So for this one people really hope we can uh start with pan 4 because it has a common uh modulation schemes as optical 400 gigs optical link. However, this one I have to say we need to do a very careful study to prove uh pan 4 can satisfy the strict reach and also is can um yield requirements across a massive deployment and another point I want to bring up to the audience is um ecosystem consistency. We know on the whole side uh we already ded or beginning to use pen six as a passive cable copper cables as a modulation schemes and if we want to use pen six as chip to modules of course is based on the rettimed optics uh adding the low uh weight or you uh for the the gearbox functions convert pan six to pan four before go out to the optical links and this one we can uh unified the host mode for pen six to support all three passive copper cable active copper cable and a rettimed optics. However, on the last but not least topology is linear and um Oro half retired. So um we know pan 4 is a requirement we can now use in pan six but this one we have to deal with the co-optimizations between the host service and the optical channel dynamics because uh we don't have the uh DSP in the optical modules to clean up the signals and we have to do harder work or we have to compromise the uh host channel loss. So no and a CPO as two packaging choice can bring the optics closer to the the host. So it make uh possible. So bottom line there's uh when we do the modulation schemes we have to uh evaluate um topologies and um differently. So the second two fact inner fact consider especially for the pan six here we cannot only look at uh coding games because there's three other factors we have to balance together coding game coding overhead and coding latency and coding complexity and the coding game as I would say is u uh we have to use in the net coding game as the metric instead of the raw coding games we need to look as net coding games at FRR post uh alter resignment code as the metric and also we have to look at uh overall the net coding games on the real channel for example um on the top uh button oh no so right button the plot show when we add in the overhead uh through the um faster over uh data rate and this uh net coding games will be sacrificed and if uh we increase the B rate from 165 GB to 180 GB and for the ideal case you will lose point4 DBs net coding game but however for the long rich channels you will lose 1.8 eight dBs net coding games it could be wash out all the coding games you added into the system. So we have to be careful for the longrange channels. Maybe we have to extend the modulation space or using more efficient consolation um mappings to achieve the overhead instead of racing um increasing the B rate. And then of course we have to choose uh the short inner code and parallel decodings in order to reduce the latency. And then also we need to consider like a t equal to one hamming code or t equal to2 b code to uh reduce the decoding complexity and also hard decoding soft decodings uh need to be balanced to achieve the that area and the sis uh DSP proximity and then uh bottom line here we're talking about inner code need to operate inside the 30s, DSP30s or very close to DSP30s. We cannot separate uh FI and then the FACK 30s and FAC designs anymore. So the code design paradigm shift needed uh for the OFCI for gigs to define in the later stage. So to summarize uh to deal with gigs there's two m uh useful tools we can use. One is modulation, one is inner fact and then so the modulation strategies must match the channel realities and intent to market. So pen six can serve uh passive cable uh well and then the pen four serve active cable or um linear uh optics interface well and then the pen six can uh provide the and users uh customer use uh designers unified uh the host mode and can support uh passive cable active cable and rettimed optics um together and the net coding game is the one matrix uh to uh investigate when we uh making the inner co fact and then overall is a paradigm shaped needed for 448 gigs that is um my presentation so any uh do we have time to question >> one quick question for Kathy >> Kathy you are part of the broadcast I mean if you suggesting multiple modulation scheme. You think the ASIC developer are working on sis that can you know one work with PAM 4 another for PAM 6 or PAM 8? >> Uh good questions IP if you want to do the door mode PA six and PAM 4 you can switch easily. Nope not that easy. So we may have to consider two uh development. one is optimized for pen 4, one is optimized for pen 6 depend on what topologies and customer designers want to use. >> Uh but I mean the optics uh is is going to require gearbox or optic technology also need to scale for PAM 6 because they are traditionally PAM 4 for a long time. >> It's a lightweight I would say the gearbox function is lightweight low cost and but it's only work on the retired optics. If you uh right now the first generation of the 400 gigs optics I believe is four generation is retired right it's not linear not half retired yet and for this one is we can add a low uh low cost conversion the gearbox function from pan 6 to pan 4 and then later on if we want to move uh to the linear or har we have to consider using no cpo instead of you know all the way go to the from the host die to the uh ball box edge that is I don't think is a v physibility is proven >> for the linear. >> Okay, let's give Kathy a hand.