HPC YEARBOOK 2021/22
down to – do we see a realistic way to go, a genuine and significant
>
processors coming out optimised for AI and machine learning, such as those from Graphcore and Cerebras, the latter of which builds a wafer-scale chip – it’s one wafer, that’s a single gigantic system. And we then make these new technologies available to the whole ExCALIBUR programme, so that everyone can try them out and discover how they might change the way they’re developing their algorithms.
How do you evaluate new technologies? Most of these technologies are so unique, they almost need to come with their own set of success criteria. Where possible, we evaluate new technologies using the benchmarks that are most relevant to that space. So, if there’s a new kind of processor,
and it can run a particular benchmark 10 times faster than anything else, that will look really exciting. If it runs but at a similar speed to existing technology, such as a GPU for example, then you might say it was good to try, but it looks like that technology won’t be really a breakthrough for ExCALIBUR. If a vendor has a new kind of network which is programmable in some way, then we’ll work with them to try and demonstrate that the programmability yields a significant benefit of some kind. For example, it might enable a new kind of highly optimised collective communication operation, which might mean that some codes are able to go a lot faster. If we can’t think of a really good use
case and get a significant benefit from that new feature, then maybe that was interesting to try, but it’s not going to revolutionise what we’re doing in the future. In each case what we’re looking for is a route to a significant benefit from a new technology for exascale
12
computing and exascale science codes in the UK. That’s ultimately what it comes down to. For FPGAs, if they could get you a 10 times speed up, but you’ve got to rewrite all your code into register transfer language (RTL) and you’re basically generating hardware, that’s probably a non-starter. For new kinds of processors, it is especially important to evaluate ease of use. So on an FPGA, you can now actually write code in a high-level language that you might want to use anyway – for example, there are some people looking at whether you can use SYCL, which is a sort of high level C++ parallel abstraction, which you can also use with GPUs and multi-core CPUs. If you could use an approach like
SYCL and from that and generate efficient code for FPGAs as well as CPUs and GPUs, that would be powerful. But if you have to rewrite all your code for a specific platform, that’s probably not going to happen. Ease of use and ease porting are key metrics in any kind of new technology we’re considering within ExCALIBUR.
Are there any early success stories from these early evaluations of new hardware? We have got some FPGAs in the ‘Hardware and Enabling Software’ program but those are some of the more recent projects that are starting up. We have some projects evaluating the Bluefield technology from Mellanox, now Nvidia, and that’s been useful to get a feel for what these programmable network technologies can and can’t do. We’ve had several projects evaluating
different kinds of GPUs. Over the last 10 years, Nvidia has owned most of the GPU market in HPC, but other GPUs are becoming important now too.
benefit from those technologies for exascale in the UK?
“
AMD GPUs in particular are being used in many of the first wave of exascale machines in the USA. Intel has some exciting GPU technology coming to HPC soon in the form of their Ponte Vecchio GPUs.
So we’ve had a couple of projects looking at AMD GPUs, and getting things running on those as well as Nvidia GPUs, and that’s been quite successful. We’ll evaluate Intel’s Ponte Vecchio GPUs when they become available too. These efforts will ensure we have much more agility in UK science codes so that they can use any of the technologies in the future when they turn out to be successful.
This means that we need UK science
codes to run on whichever GPUs turn out to give us the most science per pound, whether those are AMD, Intel or Nvidia GPUs, and ideally, our codes will be able to run well on all of them. This is a key goal that ExCALIBUR is trying to achieve. We have got some of the AI and machine learning hardware available to the project as well. For example, we have some of the Graphcore technologies available both in Bristol and at UCL. This is a nice story because the Graphcore processor was designed here in Bristol, so it’s nice to have a local link there. Then we also have the Cerebras
technology which is from the US. One of these systems has recently been installed in Edinburgh as part of the ExCALIBUR hardware and enabling software programme. This is one of the first of its kind anywhere in the world, so it’s quite exciting to be able to make these new technologies available to AI and machine learning users in the UK. l
www.scientific-computing.com
That’s ultimately what it comes
“
Novikov Aleksey/shutterstock
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42
