 Mae'r cyfnodol yn ymddangos, yn ymddangos i'w ddweud. Felly, y University of Bristol, y GW4, yn ymdweud â'r maesyn gyda Simon Macintosh Smith, yn ymdweud â'r maesyn. Felly, rwy'n gweithio'n gweithio'r ddweud. A'r gweithio'r gweithio'r gweithio'r gweithio. Yma, yma, yma, yma, yn ymddangos, mae'n gweithio'r system yma, yma yma, yma yma, yma yma yma yma, yma. BSc, barcelona supercomputer, sy'n ddod o'r marinosrwm 4 sy'n gweithio'r cyflog, mae'n ddod o'r nodu o post-k o'r ddod. Rwy'n gwneud i'n ddod o'r cyflog yn y masynrym, yn ymddiadol. A'r project y Comanche, ychydig o HPE, gael eich ar bardzo i'w peirio ar y gweld, gyda'r bynnag o'r cyflog yn ddod o er mwyn. Ond mae'n gwneud i'r par料 o'r gwneud y trafnol, wedi cael ei ddod o'r cyflog yn ddod o'r cyflog. Ychydig o'r cyflog yn ddod o'r gweithio, yn ymwyaf meddwl i'r technologi ystod. Ychydig o'r gwneud, gyda IEU, gan amser rydym gylai, a'n gŵl ymddir iawn, gallwn cael ei ddod o'r gwneud, yng Nghymru, yw eu, os yw'r ysgol y Mont Blanc, yw'r ysgol yng nghyrch, ac mae'r Mont Blanc 3, 2020, ac ym Gwroexa sy'n rhoi'r projekty yw'r ysgol yng Nghymru yw'r ystafell yw'r ystafell, yw'r ystafell yw'r ymgyrchun hefyd. Yn Japan, mae'n gweithio ar gyfer post-k o'r cyffredin, ac mae'r ysgol yw'r ystafell yw'r a Chino wedi bod yn bod yn anon yw'r arddangos oes yn un argynwysau'rёмau. So, mae'n anon nhw'n mynd i, mae'n anon nhw'n mynd i, a nhw'n mynd i'r arddangos oes ym Mhwyth. Rwy'n bwysig, mae'n anon nhw'n mynd i chi, rwyf wedi gwneud yn anon nhw'n mynd i'w rwyf wedi gwneud. Felly, y FF2, yr anon nhw, yn cael ei cyfrannu, yn anon nhw'n pwysig. today they've had their press releases going out about this. And this is their second generation SOC. One of the exciting things we'll be hearing later about this machine in particular is its very large memory bandwidth, you can see the 8DDR channels and we're talking a very large number of cores. So 32 cores per chip and you're having two sockets in the system is one of the configurations that you're able to get hold of. Also announced last week was the Qualcomm Centric 2400. Ac mae'n ddweud, y cyfnodd arnaf, ddyn ni'n ddiddordeb, ychydig ar y cyfnodd, ymmerch, ymrhaeg, yw 48 ar y cyfnodd a maen nhw'n ddiddordeb a maen nhw'n ddiddordeb. Mae'n ddweud o'ch gweithio'r cyflodd, wrth gwrth gwrs, â'r ddweud yw'r clerwyr yn y cynghwystech Cymru. Mae'n rhan o'r cyflodd, mae'n 10 nanomethau mewn cyfnodd. One of the things that our partners are doing is constantly innovating to try and get energy efficiency and performance out of their workflows. So some of the big announcements that we've had this past 24 hours, about two o'clock this morning I think it was, Cray announced that they have announced full production systems, so XC50 systems of ARM V8 chips with Cabium Thunder X2 technology are in there. And also HPE announced at 7.45 this morning that they have now got a fully productised version of the Thunder X2 as well called the Apollo 70. And you'll be able to hear more about that if you go and find out more about from some of the partners as we talk later on this afternoon. And this one's been a project involving various people, so we've got ARM, Cabium, Red Hat, all involved in that sort of thing. Atos Bull, they announced at ISC earlier in the year, they've got Thunder X2 systems as part of their Sequana project coming out later next year as well. And that's what the Mont Blanc III prototype is going to be all about. So one of the things about being in ARM is that for the past however many years people say, oh it's great you've got all this hardware coming, but what about your software ecosystem? And when we think about the mobile space for example, ARM worked very hard to make sure that the software ecosystem was there to enable all partners to succeed. And it's not just a case of ARM doing the work, all the partners themselves contribute together to make sure that this sort of effort to get all the state of play for all the codes that people want to run on ARM systems comes up together. So whilst you're doing work which benefits your microarchitecture, it's also benefiting your partners, and so people have choice at the end of the day. If we've got one SOC targeting one thing and a slightly different one targeting something else, you can make sure that you're picking the right system for the sorts of workloads you're interested in. So open source work is a very important part of all the areas that we do efforts in. We work on both GCC and LLVM within ARM, we look at trying to make sure that the compilers in open source are as high performing as possible. So I got our compiler teams to put together this slide so it probably talks about lots of things that you and I probably don't care about. But what's quite interesting in this column on the left, when it talks about spec int as the benchmarks it was doing, the sorts of work that they've been doing have actually been getting big benefits for these benchmarks but actually by optimising things which are used by everybody. So in particular there I'm picking out the G-Lib C optimisations which have had big influences on making performance gains overall. Looking at single thread performance within Lib C as well and also some LLVM contributions as well. We have to work hard to make sure that we are optimising for all the calls that are out there. So architecture support for ARMv8, every year we have a point release so we make sure that all the new features that get into there are also fully supported within the compiler. And so adding new calls optimisations in there and also systems like the ThunderX2 and the Qualcomm Falcor calls make sure that all these calls are supported in the open source. And there is extra research work going on involved in there in terms of even more performance optimisation so I'm not going to go into those details at the minute. The HPE Comanche programme has been running for since about the middle of last year as far as we've been concerned. So this was what I was talking about in terms of a collaboration effort. So HPE, ARM, Cavium, Mellanox, Red Hat and AMD have all been involved in this one big project to take the ThunderX2 technology and put it into an advanced technology programme to make sure that we're putting it into real people's hands to see how well it can work and what are the speed bumps that need to be resolved. So it's been an amazing collaboration programme where people have actually got together every week from these different organisations and been looking at things like MPI performance. Are we getting the right sorts of things going in the drivers? Are there things that need to be fixed? And there's been lots of increases in performance and bug fixing going on because of this programme. It's not just those companies involved, we've also been engaging with customers as well. So as well as ARM on some of our non-research EHPC side, we've also been working with the national labs. So Argonne, Lovensalamos, Lovenslivamore, Oak Ridge and Sandia, all of which are taking delivery of systems, all are going to be contributing work back to this project to find out how well ARM is working for their technology and making sure that the stuff we need for EHPC to be a success on ARM is actually being developed at the right time. Working with these applications, we don't any of us claim to be an expert on every single EHPC code. That would be a crazy situation. We know that the likes of you are probably developers of codes or users of codes that know the right sorts of benchmarks that would really be useful for us. So we're trying to engage with codes. We've got to make sure that we can compile them up to start with, with both GCC compilers and ARM compilers. I'm sure Cray are doing similar work with their own compilers. The one thing that we've found is that most codes kind of just work. You might have to change the occasional compiler flag to get it to work, but getting it to work to start with is almost always pretty straightforward. The next step is going from there to making sure it's optimised. And we're doing lots of work to try and make sure we are picking out any intrinsics from other architectures that can therefore be replicated in our intrinsics, for example, making sure that the right sorts of balance between open MP and MPI is all understood. So what we've got is we've got a GitLab page where we are collecting all this information together. So this is all part of our community work that we're trying to do. We're trying to make sure that user groups like this, the HPC user group, the Going ARM initiative, making sure that they all come together to bring the likes of you here to be interested in what's going on. But then to engage beyond that, to join the Google groups thing we've got, which is ARM-HPC, which doesn't show up brilliantly on this projector, and also these GitLab pages. So gitlab.com slash ARM-HPC, that's the place where we're being able to put together all this information. As I say, you can find access to the data we've got there already. We've got a whole load of applications that we've not been able to test yet because you ask anybody, they'll come up with a different list of the 10 most important ones they want to get optimised. But we've not really hit any showstoppers yet in terms of getting something working. These GitLab pages, as I say, they're community-driven. They're not being run by ARM, for ARM. We want you all to be able to contribute your experiences, say this works, put the instructions to make it work up there, put your best optimisation, maybe even include information on which test cases would be good to be running and what performance you've been able to get on different systems. And it's tracking everything from the underlying libraries all the way through to the full applications, the mini microbenchmarks that people have been running as well. On top of those sorts of work, ARM have also been involved in doing various open-sourcing work of their own to try and help users such as yourselves get the best experience out of what's going on. I've highlighted three GitHub repositories we've got there. The first one, ARM Optimised Routines, is a set of math.h functions that we have found from our own commercial compiler, had better algorithmic performance than standard ones on the system. So what we've done is we've open-sourced those so people can be able to pick them up and use them as they see fit. These are highly accurate. There's a test suite in there as well to make sure you can see exactly the accuracy of those compared to the G-Lib C work as well. So we've only included the ones which actually were performing better in there at the minute, but it might grow over time as we put extra work in there. The Sleaf Library is another interesting open-source project, not one that ARM owns, but again, they provide math.h functions, and this is work to look at vectorisation across multiple different architectures. But what we've actually done is been able to use this in our commercial compiler, and that might be something we hear a little bit more about later. And finally, on the right-hand side, it's the Perflibs tools trying to understand which math functions from BLAS and LA-PAC you're actually using within your codes. And when you've got a piece of software that uses hundreds of thousands of lines, you probably don't know which BLAS or LA-PAC functions you're calling. And so we equally need to learn that sort of information. So getting hold of runtime information as to which calls you made, how many times you called them, and what sort of input size. If you're doing very tall skinny matrices, for example, that might be a very different sort of optimisation case from short fat ones. And if it's trying to understand this is, right, you call this code, this function a million times in your code, but it only takes a fraction of a percent of the overall runtime, that's not one that you probably worry about the optimisation of much. But if you find you're using 80% of your time there, and maybe this isn't the right function to be using, then take that information and look again at how your performance is behaving. OpenHPC is another one of these software projects that's been going on for the past couple of years. It was actually announced two years ago at SC in Austin. It's now up to its 1.3.3 release, and I'm very proud to announce that ARM is now no longer in tech preview status. They've taken that flag off. Just about all the packages that build for OpenHPC for ARM on both Centos and Suzy, so you can just go and download those. What's been really good as well is ARM used to be on the technical steering committee for OpenHPC, and we've been able to move that responsibility onto an organisation called the NARO, a kind of an organisation that is above ARM and our partners working together to get the best for the ARM ecosystem, and they've also got a special interest group for HPC now. More than that, other partners have been contributing to the work going on within OpenHPC, so we'll hear probably later on about some of the work Fujitsu has been doing to make sure extra packages have gone in there that are relevant to their users. There's a lot going on this week, certainly from the ARM point of view. In terms of official SC events, we've got the Exhibitor Forum presentations, so Fujitsu, Cavium and ARM have all got talks to do, and there's a panel session led by CJ Newburn from NVIDIA on the ARM software ecosystem. Then there's another buff, and on Thursday afternoon back in here, we have a session about SVE. SVE is our Scalable Vector Extension. This is the future technology coming in however many years' time, as I say, that's going into the post-K architecture. What we want to do there is get some early experience from people who've been using things like the compilers and the simulation tools. We've also got various booths as you go around the exhibition floor. ARM's actually got two booths this year, which you can see there with a whole load of demos on our tools, but also some of the technology from our partners, so Cavium and Qualcomm both have systems up there. Also Fujitsu of doing some SVE compiler stuff, and Singularity are talking about their containerised stuff. Other booths of interest will be the Cavium, the Cray, the Fujitsu, and HPE have actually got a special booth dedicated to the Comanche project. So booth 494 is all about the Comanche project. They've got a whole rack of the systems there, so you can have a look and see what it looks like. It looks like a server rack. You can go and have a look and find out more information about some of the work that's been going on there. In a month's time, there's going to be another ARM HPC event. This time, organised by Leonardo in Tokyo, the ARM HPC workshop, and there's all sorts of details about it on the Leonardo webpages. I think we've probably got time for a question or two, have we? One question. Five minutes? Oh, cool. No questions at all. Right, as we say, we've got a whole load of interesting talks coming later. These people will go into far more detail on what's coming up than I've done. So please enjoy, please think about the questions you've got, and I look forward to hearing more from you. Thanks.