video_OCP2026_Oracle_NPO部署_20261015

Good afternoon everyone. So today I'm going to uh give you uh our view of uh NPO deployments more from a fabric reliability and stability perspective what it means. Uh so uh this is actually a interesting chart that I thought will set set the uh nomenclature a little bit. uh uh it was presented at uh OAF by Chris and Jeff Hutchkins. So it's like front front pluggable optics, near package optics, co- package optics, the different variants of the nomomenclature floating around. So I I thought I'll just calibrate everybody into this. Uh one of the interesting thing is you see OB OBO's onboard optics. So uh the precursor to today's NPO is actually from about 12 years back uh which Oracle we actually developed uh what's called a midboard optical module which an onboard optical package uh it's actually a socketed solution that we deployed uh and pushed into uh high volume production and uh we also actually deployed the very first expanded beam optics as part of this solution and uh there are many lessons learned from that uh experience uh deploying this in high volume about 12 years ago and so I'll share some of that experience what it means for current AI fabric reliability and stability and uh so wanted to acknowledge some of our partners from way back uh TE and Amphenol FCI who developed the midboard optical modules for us uh back then And so a lot of those folks are still uh relevant to this conversation today. So uh NPO adoption uh it's actually a very interesting ecosystem uh evolving. Uh we have a lot of uh suppliers coming into this ecosystem which is actually great uh for the industry. Uh when you compare to uh CPO uh the NPO ecosystem is much broader. uh we actually want to support that also. Uh we have electrical connectors, external or internal lasers, polarization maintaining fibers, fiber attached units and then there's a whole conversation around uh linear drive. Uh so we expect broad adoption of NPOS's in scale up, scale out and scale across. uh and uh also there is a significant uh power advantage here. So scale up target if you people are talking about slow and wide could be a microLEDD or a micro pixel it's about 1 to2 picojoule per bit uh with a caveat that that some of the power benefit has to be given away because there is the reverse gears box uh that uh eats into some of that power but a scale up conventional scale up or scale out uh it's uh 5 to 10 pico per bit which is actually pretty Good. So one of the uh foundational requirement is the NPO must demonstrate reliability and power efficiency uh should be on par with soldered CPO. So uh as long as that is uh met I think we'll be uh great. Uh so the NPO adoption the benefits uh many people have talked about it. Uh reliability uh availability I'll get into a little bit of the uh details here. the system fit is actually going expected to be uh much lower than the uh pluggables. uh transceiver mishandling contamination is a significant part of the failure modes that we are seeing and um the NPO and CPO in that context removes human handling and from deployment and operations and so uh that once that failure mode uh a big paro of that uh is actually uh removed uh the system fit is actually lower than equal and pluggable optics of course power savings and cost reductions. Uh several people have talked about it already. Now in the NPO there are two uh different uh streams that we want to talk about. So there are two connector types uh in in the uh MSAs that are coming up. So connect type one connector that has uh separated out the highspeed from the management and power and then type one which is uh type two that has integrated that. Now along with that there are two variants of the lasers also. One is an external pluggable laser. The other is an integrated laser. As you can see in in this uh cartoon uh image courtesy of eoptolink uh the size uh the footprint of that is uh different depending on which connector is used and whether it's an ILS or an ELS uh type laser. Now speaking of connectors, because we have a lot of experience from way back uh deploying socketed type solutions, we have to actually focus on a lot on failure modes associated with these connectors. And I just put a list of things that we have encountered in the past. And then there is also the uh additional things that we can look at from a type one uh RF tomp uh relative position error and so on. So these are like mating, misalignment, uh signal integrity, degradation at the high-speed contact field or uh retention and uh lateral load failure and uh the thermomechanical dis uh distortion from cold plate integration. So these connectors uh are uh the the cop planarity and the ability to mount with a cold plate and everything uh it can introduce some some failure mode. So uh back when we deployed we were deploying like tens of thousands of units of uh infiniband switches but now we are talking millions of units here. So the scale is pretty large. So the failure boards associated that with the large numbers I call it the tyranny of large numbers uh that actually crops up. Uh yeah so the next uh topic I want to transition to is the lasers for NPO. So uh most of the laser work uh has been done at OAF in the context of the ELSFP. Uh so there are two uh variants the very high power 100 matt laser which is 20 dBm and the ultra high power lasers which is the 23 to 26 dBm. Now uh the power efficiency uh the wall plug efficiency is about 20% is the target less less than one pico per bit. But there is an interesting chart that was presented at ECO uh recently by Rangchen of Terrahop uh where they were saying oh the uh the laser the CW laser the fit is less than 0.1 fit uh however that is actually not applesto apples comparison in the sense that that is actually OSFP based uh uh uh uh laser power but here for the context of NPO or even CPO for that matter uh it's actually ultra high power lasers. So, uh the fit rate associated with that we expect it to be higher and uh actually that has to be also validated by uh the the laser uh providers. Uh but we can also tell that the laser fit will be substantially lower than a module fit. uh but the problem associated with the laser one of the significant components of a laser failure mode uh that we have seen in our deployments uh laser mode hopping that problem is not going to go away. Uh so whether it is uh NPO or CPO we have to solve uh and screen out for laser mode hopping. So we see that and then uh thermal management associated with the laser uh either air cooling or TEC or liquid cooling. Uh so those those uh things are very relevant to the context of what lasers are doing. So I want to compare now uh what's the uh differentiators between external and uh internal laser because in the NPO context you have both uh uh ELS and ILS. Now uh certainly uh ELS provides for higher optical power options uh internal lasers we can go in a lower power regime. However the blast radius associated with ELS uh is actually low because it's actually a pluggable right if there is a failure in an ILS uh it's actually inside the box. the NPO box needs to be field returned and there's better scalability flexibility because it's a pluggable on the ELS side field upgrades are possible whereas upgrades cannot be done without field return on the ILS and uh the there is with the ILS that there is a challenging thermal management uh uh it's a uh we have to rely on the liquid cooling loop and ELS one of the major uh pain points that we are starting to see is with regard to the polarization maintaining fiber. Uh there's lots of uh stress factors associated with the polarization maintaining fiber, the epoxy stuff and all of that. So it actually adds to the fiber uh and cabling complexity whereas ILS doesn't need the polarization maintaining fiber and the performance actually in that case is dictated by the laser itself on the engine. So uh ELS therefore has additional interfaces and inter insertion losses and complexities to uh deal with simpler interfaces on the uh ILS side. Uh obviously cost is higher on uh the ELS side a lower bomb and system cost on the uh integrated uh laser side. Now transitioning to a different uh topic which came up in a couple of uh presenters previous presenters. uh so uh NPO basically we are in the linear regime but I want to calibrate a little bit on uh what we have accomplished uh so uh at Oracle we have deployed the largest uh LPO uh of any any fabric ever anywhere so the LPO performance is great when it is actually LPO to LPO uh FR to FRO uh compared to FRO to FRO O uh the actual uh fabric uh the link flaps is actually lower in the LPO context. So the LPO uh is actually uh good from that perspective. So the number of transitions about one and a half 1.75 times less transitions uh than the FRO. Now we are moving towards LRO at 1.6 60 mainly because the the the signal integrity uh is not there uh for uh 1.60. So LRO is actually LRO or TTRO uh half time solution is what we have uh working on. uh interop has been demonstrated interrupt is a big part of this right uh so interoper has been demonstrated across multiple suppliers uh with LRO and uh it it boards well for NPO uh because linear drive correlates well with stable resilient network and it's it's it's actually pretty good uh very uh highly promising but what is actually the missing part uh in the NPO context we have uh just one one point here is uh more details on the uh linear drive is going to be presented by my colleague uh Mark Filer in tomorrow's session uh uh and then uh here uh another colleague uh Steven Manley he's going to be talking about the disappearing DSP so in the NPO context what is happening is we have DSP P either completely being eliminated or DSP in a half free time mode right TRO or LRO right so but all of the functionality associated with the DSP cannot go away we need that uh we have that uh rettimer equalizer the effect uh BR uh everything whether how we are keeping the diagnostics and everything so today we have a big challenge in our network which is I don't have visibility into what is happening on the remote side once the fabric is delivered to a customer uh it's blind to us so we are operating blind so I can see what's happening on the switch side but not on the host side so uh in the OCP workstream there is a work reliability workstream where uh remote and telemetry is actually being developed uh so that is actually will continue on as long as pluggables are available. But once pluggables go away into the NPO context or CPO context, we are losing that capability. So that capability needs to get sucked into uh the switch side of uh things. So uh the DSP may disappear in NPO but its operational responsibilities actually is a mustave. So for more on this listen to uh Steven Manley Manley's talk tomorrow. Uh he will be going into the details of some of this. So the last uh uh message I want to uh leave uh here is uh what can we expect for link flaps with NPOS's. So just because we are removing pluggables uh doesn't mean that uh we don't uh see link flaps. link flaps uh are there. There are two categories of link flaps. One is called a persistent link flap. It points to a catastrophic failure in uh failure mode and then there is a transient link flap. Transient link flaps are those mysterious link flaps that come in, pop up and then vanish mysteriously. And so we we see that in a pluggable system uh uh basically dust contamination uh impact is high that actually causes uh uh link flaps because uh of uh things like uh MPI multipath interference and so on. Now uh the contamination factor goes down quite a bit in the NPO context, CPO context. uh there will be minimal impact to engines from uh contamination uh and so that failure mode is gone. Uh however uh we have a few other uh uh connection related to the uh the fiber itself uh the polarization maintaining fiber the fiber attach unit and so on. So any any possibility for contamination there has to be eliminated completely. Now DSP failure modes are also prevalent. Uh uh loss of lock. We see examples of loss of lock. Uh and now because DSPs are uh going away or uh half DSPS uh some of those uh impact will go down uh quite a bit uh the flaps associated with those will go down. uh laser mode hopping whether it is uh ILS ELS uh that problem will persist. I think the the only solution for that is it has to be screened out at manufacturing uh all the way from D level uh to chip on carrier uh to the optical engine or uh uh the the ELS uh module itself. So uh there is no no uh uh way around it other than screening it out for this. So we we do have uh uh excursions that we are seeing laser modop excursions. It could be a combination of factors uh environmental induced or things like that. But we have we have we are seeing that now thermal and environmental uh variations uh in the NPO context it's better controlled uh because there is uh a cold loop liquid cooling loop. So that that actually better that will go down. uh now I did mention about electrical connectors. So there are failure modes and socketed connectors. Those will all be uh uh uh need to be uh factored in uh uh the connectors uh will cause SI deviation. So we we have to so we are talking about uh hundreds of thousands to millions of connectors there. So those uh what works well in the factory when it's shipped uh when it comes to the data hall uh actually can be problematic depending on how good how robust the shipping and other things the vibration sensitivities and so on are there. Uh now in the scaleup context we are seeing a lot more uh fiber density. So the failure modes associated with the fiber density in in the uh pluggable system we have the flat fat finger issue during servicing and repairs that will cause link flaps in in uh the fiber density in scale up we have uh doing that and then uh uh the the fiber contamination on the face plate side is actually the panel side is also uh remaining. So uh the industry is moving towards expanded beam uh optics. We are uh co-chairing the ebo MSA with 3M and all that. Now the two other points there is uh latent defects the we are talking a large number of uh parts uh surface mount components uh uh all of that. So SMT micro cracks are something that we are seeing uh everywhere. ESD uh dam damage uh latent defects. So these are all latent defects that can manifest 3 four months after def uh deployment. Uh some of the RCA analysis have point pointed to uh these kinds of defects. And then if it is a repeat offender, it's very easy to uh pull a pluggable out. Whereas in the context of NPO, if it is a repeat offender uh related to the NPO uh engine, uh we uh actually have to send the box out for repair. So that's basically what it is. So just because uh NPO's uh you know uh the engines are inside it's uh developed in a controlled environment doesn't mean that the the flap problem will go away. It will considerably reduce but it's still we we still have to uh pay attention to that. So uh I'll just stop here. Uh these are some of the takeaways of basically what I just uh discussed. Thank you.