Hallo Du!
Bevor du loslegst den Talk zu transkribieren, sieh dir bitte noch einmal unseren Style Guide an: https://wiki.c3subtitles.de/de:styleguide. Solltest du Fragen haben, dann kannst du uns gerne direkt fragen oder unter https://webirc.hackint.org/#irc://hackint.org/#subtitles oder https://rocket.events.ccc.de/channel/subtitles erreichen.
Bitte vergiss nicht deinen Fortschritt im Fortschrittsbalken auf der Seite des Talks einzutragen.
Vielen Dank für dein Engagement!

Hey you!
Prior to transcribing, please look at your style guide: https://wiki.c3subtitles.de/en:styleguide. If you have some questions you can either ask us personally or write us at https://webirc.hackint.org/#irc://hackint.org/#subtitles or https://rocket.events.ccc.de/channel/subtitles .
Please don't forget to mark your progress in the progress bar at the talk's website.
Thank you very much for your commitment!

======================================================================


With me on stage, it's a wonderful person who will talk to us about CPU development, power, open source and how this all fits together and how this will hopefully influence of future CPU architecture and development. So please give a very warm round of applause to Matthew Michel. OK. Thanks for having me here today. And I apologize at the beginning for my strong German accent mixture of German and English. So sorry for that because I'm coming from second year actually living in Kemet, still born in chemist's, grew up in Kemet, so starts whoever knows about that. Um, that's why the accent is still a mixture of sex and German, but still want to show that are still smart people living in Saxony, not just people in Western waving some strange flecks all the time. Um, I want to say thanks. I just want to give a preproduction today about how I got inspired. I think two years ago by a talk by David Caplan, who was talking about Sebou development altogether from scratch through the final product and all these steps for it. I also want to touch that briefly today, but I want to go more into open power and just tell you something about this because I'm not really sure if everybody knows about that. Uh, yeah, foundation. And um, might also be a little off topic in this conference here. But uh, maybe it's still of interest. Um, this is my agenda for the day. So first of all, uh, like to just talk a little bit of Barbie then uh, as mentioned about this, if you would open flow at all and y or power, uh, myself, as I mentioned, um, still living in cabinets. But I'm um, my desk is located here. This is Steve bobbling of IBM Germany. Um, I'm there in the verification department of CPU development. Um, it's not just located there, it's uh, built by a distributed team, virtual teams. Um, I'm looking to get to with people in Bangalore, US, uh, France and so on. So will you? Uh, yeah, all the candidates are included. Um, my main focus is verification. Um, I'm also not an expert of a pow
er. Um, but I just still involved in some topics of it. Uh, that's why I think I can at least talk about some steps to you, and that's all about me. Let's go into the technical agenda. Um, this is the CPU development for s we are still looking at today at IBM. Um, so we basically it's um, you know, a simple overview. You start at the left, uh, by either coding or generating or somehow specifying some hardware description. Um, maybe it should maybe be a log or whatever you think might work. Um, this is also, um, this, uh uh, large arrow pointing to it. Um, this also I'm using usually for verification. Um, there are also multiple different versions of each year that are possible. But just assume it's basically one thing that's busy specifying the hardware how it should work. And uh, yeah, there are some kind of two or three that body converts. Um, all the different things to the final product, um. Right? Um, so out of the HL, uh, you will get the list with all those logic components being generated. Uh, that also some other specific application teams working on that one. Um, then you can generate the schematics out of it. Uh, yeah, you will see some resistors in there. Um, yeah, maybe. Uh, forming those transistors into layout and layout, you basically get to define a product at the end of the application process. Um, I will also have some numbers later today, but uh, assume this process. Uh, you can go through hell on a daily basis, but uh, from a first line of uh, yes, code until the final product. Uh, I just assume like two or four years until you have a real working product. Um, now there's some nasty, uh, uh, flowchart from some of my colleagues, but somehow it just should illustrate how the overall flow is somehow working. So you just saw the diversion coming from, uh, hot weather scripting language into the product. Now this is the workflow inside the company itself. You have basically some kind of product. Um, you believe CEO targets or the previous version. Y
ou have some customer that somehow has some need for certain things to run on that and typically gives some kind of specification. You have some teams working on high level design and then this is basically your input source. Uh, sometimes very specific, sometimes very loose. Um, and then you have different teams working in parallel, uh, trying to to get this product put together. So we have some kind of system component teams that are looking at the boards and all the electrical components on it. You have a team that's looking at this black box. Uh, we gotta to put it at the end, you have premier teams, you have basically a design team that writing that visual code or very low code and then you have basically at the end, uh, verification team that is busy making sure that this also works at the end before you go through it and get it produced. And then you have also other teams looking at timing and looking at circuit design integration. So that's all going into physics, which is really important that the process design, because you want to have the maximum performance coming out of your product. Um, yeah, when you're done Besley, you push the button, get it fabricated. Then there are some tests going on. And if that goes well, then you have a good machine. Um, sometimes you find bugs. That's why this small red bellows is coming in. Um, if you find bugs, you basically want to find them as early as possible to basically turn and turn around your logic until it really works before you get to the FAP. Um, if you missed that point, um, do you have to spend much more dollars on it? And then you, yeah, um, have to spend multiple cycles on that. Um, just be sure you don't want to be that person being responsible for having to respond, uh, through the FAP, because that process going to that is like two months, 12 months. Uh, and you don't want to be that person, first of all, spending the money and also causing the delay for the customers at the end. And if you are a reall
y bad person, then um, at the end you get into manufacturing. All these machines are being put to the customer sites, they push a button and nothing works. And um, yeah, let's don't talk about that. Um, but I didn't see that in my life so far. Um, this is not the picture somehow trying to illustrate the same thing again, but now trying to decide to the left more to design verification on the right side, more to, uh um, prevent physical stuff that is being done. So you see a physical science into this placement routing stuff that you know needs to be done. But I personally don't want to be that we are. It's always critical and you'll always be at the end of the process. Um, everybody's waiting for you to push the button. That's why I chose the left side of being willing to do verification front. Um, also dealing very close 50 design teams. Um, and you see, it's it's two circles circling because you are always some hole in the same source, different times and that's basically what you do all the time. You are going into a kind of continuous integration. OK. And here are the numbers that I promised um, some would try to compare the numbers from building a larger ship compared to a processor. And that's quite interesting because somehow at the end, when comparing you for both things you need for the whole system, about 5000 people. Um, also about the construction type. It's similar. So we're like three years with the ship, like a four to five years for the whole protozoan system. It's not four to five years for just one shot. It's basically a multiple turns of the processor until the machine is basically ready to ship to the customer. And also, you see it because at the end, so it's more than one billion dollars that needs to be spent for one generation of those systems. Just to mention, there are two architectures currently still being worked on at IBM. It's the power architecture and the sea architecture. Those numbers are for this processor. But you can compare those
 pretty similar from a cost point of view. So that's already it from the um, CPU development side. So let's jump into Apollo. Um, and you are looking at that. Um, that's my view of white power because, um, sometimes it looks strange. Um, why should something change in that kind of popular stuff? It's always be there. Uh, so you see you look on your PC, you see somehow getting smaller, you have your phones, but somehow a similar kind of stuff in there, also from the server part of it. Um, it's somehow in some building in the cellar and uh, it's just work. Um, but there is some commercial component to it. Um, over years, this kind of process worked very well for players like IBM or others. Um, but all the time the market shifted drastically. Um, so you have a Moore's law, which basically, uh, provided over years, uh, a very constant flow of, uh, performance improvements. So we didn't have to do much. Uh, you just had to move to the next generation of your technology. And then surprisingly, you got, uh, ten twenty whatever percent of new performance or changing much in your design. Nevertheless, you added more features to it, uh, to satisfy customer needs, but that it's, uh, more more, uh, never true anymore. Um, also on the other side, you have, uh, other trends like, uh, new kind of workloads are coming up. Uh, just think about, uh, some kind of bitcoin mining, which is, uh, somehow trending today. Uh, or these days, but also a um, kind of, uh, consumption of, um, those server farms are changing. In previous years, it was like, uh, somebody from some company called the I need new machines and the customer. They moved there and, uh, yeah, uh, sold something of which was staying there for the next couple of years. Uh, but these days, um, cloud is something, uh, very drastically changing, uh, the operation model of our sellers at IBM. And that also was something that we felt because at the end, um, just less and less products were sold. Um, the cloud infrastructure was 
not up and running from our point of view because, uh, we didn't bet on that. Other companies like Amazon were very fast on bringing it up. Um, so as you can see today, uh, most of the cloud infrastructure is x86 based. And uh, I think it's hard to find, uh, power servers for the open public cloud providers or others that you have seen. Also, the software ecosystem changed a little so the UM two more open and mature open providers, um, that's one part of it. And on the other side, you have some kind of strategy changes that you need to do now. Um, but more to that later on. Um, just back to the, uh, to Moore's law. Um, on that chart, you see, uh, it's basically complete performance per dollar. That's something um, the customers look at, um, they they look at how much money they need to spend for which kind of performance. And that's the number they compare between the different providers. And then only if you can, uh, match or be below the the other companies that you are able to to sell stuff to them. Um, and as I mentioned before, that was working quite well, uh, for the past couple of years. But now obviously, Moore's law is slowing down and doesn't help and more. So what you are now trying to do is, um, you need to somewhat enhance the performance by other means. So what could it be? Um, so you could plug something to your system so you can't move everything onto the processor because you just can't fabricated so you you need to do something close to the post so you can plug some extra radios on it. You can somehow. On the storage systems to be faster, you can look at the I o. If you can move to a new generation there or you can look at the memories. If they can provide you something. Um, but yet everything comes together at a data point. So data is that thing that somehow is supposed to be the issue over the next couple of years because of different trends you have, in other things, providing much or sometimes too much data and you just somehow need to handle i
t. So every device is somehow creating data. At the end, you need to have some backbone working on it and basically use some computer centers that can deal with all those data and to see what kind of data the greatest of those charts basically envisioned coming from different fields. Um, coming to power, um, and typically, um, the Moore's law is something I've already talked about. But there will be two other things that forced this kind of creation of this open power community. The second thing was Google. Um, Google was basically known to be a large provider of something like search or other stuff. Um, and somehow they were not really happy with their x86 infrastructure. So some of the big talks, um, in other companies and also with IBM, and they basically ran some stuff on the back then power aids machines and uh, surprised, uh, form different patterns that were surprisingly good. And um, on the other side, they also, um, basically were looking right in the eye of a monopoly of intel. Um, so that was the second thing. Uh, Google was not being afraid, was afraid of uh, and basically, um, seeing that this was creating a very costly infrastructure, basically being bound to a monopolist um and then not open to any movement, something else. That's why the better you felt to to do something else and also, um, at the end, um, I don't really know, but at the end they want to build their own, uh, computers, and that's why they did the search for some partnering. Um, and they didn't came to IBM and uh, yeah, looked at us if we can do something together as well as the third point is extend the reach of power architecture. Um, so as mentioned before, in the past there were PowerPC years architecture was somehow a little, uh, stronger before, um, how architecture is, uh, getting market share or gaining market share in the last couple of years. But the overall Unix market is decreasing that fast, that it's basically not really working out well for keeping that architecture jus
t how it is. So you need to find other means to to strengthen your whole infrastructure and the architecture of those boundaries and see how you can work with other parties to get into other different markets. And that was basically the root cause to basically get this open power initiative being done in 2013. Um, if you look at the other part of this chart, it back then it looked like all those companies somehow losing market share of all right to come together and found something. But basically looking at this, you base, it somehow looks like a good move. And let me show you how that worked out. Um, the first steps that were done, um, were basically done by ABM, um, first of all, basically, uh, investing a lot of money into a Linux, uh, Linux, being able to run on power. Um, that was a complaint even done by engineers internally over the last decades, basically that even our own applications were not finding really good on our own machines because we built our own computer pool was, uh, mainly, uh, based on x86 machines because we were selling them all before. Um, that was a strong complaint, but it didn't help, uh, to complain internally about, uh, the customers, uh, basically, uh, showed us that the machines we weren't really, uh, able to, uh, to beat the market. So there was some kind of mind shift going on. And uh, basically, uh, resulted in investing a lot of money in, uh, structuring the code around Linux and basically even the little engine big engine complications to sort them out and get them resolved at the end. The second thing is this kind of off source model for hardware. Um, I think I will talk about that in a minute. And basically, it's do you feel step to release, uh, lots of firmware code to really make it available? I myself was even surprised. Uh, sometimes I just, uh, googled my name and, uh, some kind of, uh, word. Um, and surprisingly, I, uh, one of the first five matches was, uh, just something that I wrote myself with a colleague. Um, so ju
st try to to search for it. Uh, you should find it very easily on the net. No. Um, this is the mission statement, uh, I think nothing to really read through completely, but um, the idea behind this is it's not a new company to fire phone, but um, it's some kind of partnership or some some open community that, um, is to be installed to get people together that want to improve the power architecture, as well as to to work on the whole stack. So that includes basically the basics of the development, as well as the the software development flow and basically use, uh, same tooling use, same infrastructure, same boundaries. That's just a picture to illustrate that on the left. Um, so you have to have the component, um, that basically consists basically of, first of all, the process itself and uh, still the process that doesn't hurt much. You need to have, uh, some connection points and which is interfaces to accelerators and not just those interfaces, but have them rather open that other companies can also use them also roadmaps for both of it. Um, that's what I learned. Also, uh, very early in my work, um, that customers really want to see a roadmap over the next couple of years that they are, uh, we can be sure to invest in your product. And, uh, yeah, can be sure that this continuous over the next, uh, technology that technology cycles, uh, as well as, um, it is not just a of it to make sure that this is basically, uh, common and uh, being used by many, uh, different parties, um, and that it basically gets debunked. The. Um, and that basically gets us to the ecosystem as of today. So the top before there showed you have seen there were five parties, um, funding this community. I know you have a whole stack that is basically, uh, starting with the basics at the bottom. Uh, now with Infinium as well as included with the chips, the bolt manufacturers. I o storage companies. System integration companies. Also software, but as well. You have, uh, many research teams include
d. Um, and that's basically, uh, something that is, um, also a key interest of that group to not just do something with companies and, uh, try to sell stuff, but also to to keep people involved in research and trying to get this, uh, whole method enhanced to trust them up. Uh, like, uh, twenty members, uh, as of today, uh, in thirty two countries, um, you know, these days, I think most of that is still centered around the US because they are basically the headquarters of IBM and the main, uh, labs still located. And that's why also many of the universities, they're joined in those areas. Um, but still basically, um, try to get members over the whole stack. And even those partners, um, also started to bring out first products. Um, but I think I have one chart for that later as well. That's basically, um, just to illustrate also, um, how Opal is somehow Mitrice or the other initiative initiatives. So we have this open compute group that is, I think, more concentrating on mechanical electrical synchronization and specifications. Um, the Opal club somehow is trying to, uh, to fill that gap between whatever is coming from top from, uh, software. Linux, OpenStack, um and open compute at the bottom. Um, open capital is, uh, some a part of that open power initiative, but it's basically, uh, just one fact of it. And as you can see at the screen box in there, uh, limits itself to a protocols and uh, I own the connections internally. Um, so that's how it would fit um or how it's intended to be. Um, and now what's so special about it? Um, just to show you that's the the current road map. I'm not sure if this is even still accurate, but at least it gives you an overview of how that, uh, it's, uh, what's been worked out over the last years. Um, so as a process of manufacture, you try to to have a real good, uh, uh, sense of, uh, regular push of products out into the markets so that different needs. Um, if you look at power, um, there are different circumstances. Uh, because you h
ave a different, um uh, there are markets that you want to attack if you look at, uh, see, architecture. This will look much different. Uh, a C processor will come out each and every time at the same point in time so that it's a very regular pattern by the site that needs to be much more flexibility due to the customer changes or market changes at the end. But as you can see in the next year, um, 10 seven nanometer quality products that need to come out and basically you will be, uh, specified and pushed 50 partners. Um, but that's the really interesting thing about the initiative, um, that you somehow try to to connect at the end. Um uh, as I mentioned before. Um, external accelerators into the system. It's not just to plug them into, but to make them part of the system, to make them part of the the transmission of the processor. So it's, uh, not just enabling, uh, some PCI also, but it's a protocol layer that makes it really fast, uh, compared to other customers. Um, also, uh, with this power eight enhanced product, um, built together with Nvidia um, the building was introduced as in, uh, its own link, uh, built into the CPU, um, to make that, uh, point with, uh, much faster into those chip use. Um, I'm not sure if you heard about this announcement quite some weeks ago. Um, there are those labs in the US which are basically distributed with those new systems. Um, there was some kind of government contract, um one by Nvidia together with IBM to distribute those systems. And basically, they consist with the latest hardware, um of, uh, power processors connected to the latest chip use and basically, uh, built a room full of those servers. Uh, and basically make possible some um computations and use this year to use generation of the willing and uh, that provides, um bentworth that you are not able to get with other CPUs. Um, that's the same thing. Um, but you see that you have um, the interface is not just over Ivy League, you still have PCI generations available to 
get that faster, just to decouple you extensions that you have. And um, here you can get, uh, Xilinx products attached. Uh, also with, uh, Cappy open copy. Um, which is basically this new protocol and, uh, hardware built into the processor. This is just for numbers, not for marketing. Um, basically, they were able to, um uh, Exxon Mobile was able to basically reduce, uh, thousand servers into 30 open power servers just to put together an idea what that makes. At the end, um, just for numbers that's regarding to the community itself. Um, the idea is basically to get uh, or do you have the possibility to get everybody involved? That means people that, uh, write software, people that use software or people at the end that, uh, uh, use the whole systems and uh, try to basically, uh, get those different interest groups together and have a place for them in that overall community. And that's why, um, they created, uh, different groups, um, to get more into technical discussions and basically technical solutions at the end. And that's basically what we currently have today. Um, different different, uh, 15 different work groups available. Um, there are some that, um, I'm even involved with some colleagues. Um? Um, that's basically we're talking about real problems and solutions at the end. Uh, they re trying to solve them at different points of the stack. And that's basically should be the last light regarding what is at the end, um, it's still some something that costs money. That's why companies need to do to to spend money to be part of open power. But if you are coming from academia, then you are able to spend zero dollars and still be part of the community. And there's different levels between them. Um, so it needs to be somehow a little bureaucratic that way, obviously. Um, but you are able to get involved, especially if you are coming from research. Um, I have even checked today, um uh, to you mentioned should be involved. Um, the, uh, Paderborn University. I have se
en the chart, uh, as well today as there are already some German universities joining that, uh uh, community. Uh, but most of them are still located in the U.S.. Um, that was all from my side. Just a brief overview, so it was not intended to be in that, uh, also not complete if you have more questions, uh, just come to me if you also contacts to other people, if you want to talk more details regarding an interface or any contacts in those groups, um, but uh, at first you should check or dot org. Um, that's the website. I bet you can get more information about that. And um, I hope that gave you some inspiration and some information about what is going on in the hardware community. People. Thank you very much, Matthew. We do have time for a few questions and someone already lined up, might one please? Hi, thanks for your talk. As a hacker, how can I get a open power development board of a not like a big data center like Google or someone? Um, I think if you have contacts to universities, it's pretty positive. Look at something there. Um yeah, that would be my advice. I'm not a university, so but I'm sure if you are an hacking activist, you should have contacts in that group. But I think you can even maybe get some some of the stuff in there. Are you all the thoughts coming up yet? It is already the old version. No previous version 2.0 and I just don't work out. Maybe there are some possibilities to get some older machines available. The internet woke up and has a question. Yes. So you talked about companies and academia joining open power. And one user wants to know, are there plans from IBM site to offer affordable systems to consumers so they can join? I think that's the at the end. It's at least an idea for for companies to to get cheap systems available, um, for the end user. So I think it's not the current target since it's also not the market IBM is looking into. So it's more business related issue at the moment. Mike, for Hello? One question about the firmware,
 you said it's 400000 lines of code. Why is that big and why couldn't you just reuse Corbould, which already supports open power the older generations? That is a specific question. I think it's just this tech that is coming out with the machines already, at least from an IBM site. And I think the same is coming from Google. Did you have additional lines of code written for their stuff and just released what they already have? So it would be question, would you feel about that the teams might be going to reuse, as you suspected Mike to a question regarding Moore's law. It's never been about performance. It's been about number of transistors. So do we have some numbers for pin power to give them? Um, I don't have numbers. You need to look that up somewhere, I guess. Um, but basically at the end performance is that what counts for the customer? So they look at performance and that performance is that what you can? And Moore's law was this the thing to improve performance over time? And now maybe it's not that easy anymore. That's why you need to trust other means. Unfortunately, quite a few questions lining up. Unfortunately, we are out of time. I'm very sorry for that, but I think my tell you is still around, so grab him and ask him in person. So please give another round of applause for Matthew. Thank you very much. The.