This week we delved into a wide variety of European projects that address future needs of systems in research and industry at HiPEAC in Valencia, Spain.
HiPEAC (more on its vision here) is one of those conferences that does not tend to get as much press as the major supercomputing shows given its European focus, but it does do one heck of a job of getting over 500 folks on the architecture and software sides of the Euro HPC house under one roof. And with some deep technical sessions to balance the industry and academic partner material. While not everything is HPC-centric, there’s enough—and in deep enough detail—to inform where research is heading.
As it turns out, the focal points for the future of computing have a different flavor than what we see in the U.S. (just as Asia Pacific also emphasizes some research endeavors over others).
The expectation, at least based on the other Euro events and projects we follow (long list here) was that we would see much about ARM, a lot about new GPU supercomputing in Europe at the research cluster level, and learn more about Europe’s mission to build exascale supercomputers in the near term over focus on longer-range tech bets like quantum computing.
And we were wrong about all of those things.
That’s not to say those aren’t important focal points in Europe, but at HiPEAC, it was the inverse of all the above.
ARM, whose licensees have thrust truly competitive chips on the market with major HPC OEMs as the vehicle, was not mentioned much, nor were the range of ARM-focused server efforts in Europe, including those focused on homegrown exascale.
Technical sessions on GPUs were few and far between with the main one focused on users and development. Instead, FPGAs were the star of the European accelerator show—a surprise because there seems to be more adoption (or at least prototype/interest) for research HPC than in the U.S..
And as for near-term exascale ambitions, those are still there, of course, but when it comes to near-ish term goals, quantum computing garnered more attention than the next step in floating point development.
So what does this tell us about the state of European HPC and distributed computing research? Beyond what Horizon2020 and other funding news tells us?
Well, to be fair, this is just one conference, so this might not be broadly representative of what is happening across Europe. And before you furiously pound out the kind of indignant, exquisite hate mail that only our academic readers can craft, note, this is not to say that the things we expected to see are not alive and well. But this conference was on research horizons and it was clear from sessions and conversations that the emphasis of HPC development in Europe is changing. Not that it’s a bad thing, either.
The topical focus points were fed by real action on the research front. HiPEAC has been quite wide-ranging, well-attended, and packed with a who’s who in European research computing.
The real takeaway from this vantage point is that Europe is trying to differentiate in a few unique ways. One is certainly by investing in the hairy software stack that underpins accelerators in general and FPGAs in particular. We sat through a number of sessions presenting research endeavors that targeted everything from compilers and languages to interfaces for clouds on up the stack. Another is by being first to show off what new hardware, especially home-built prototypes, can bring to the table for users.
The thing is, so many projects have been funded by the EU to build hardware and while these are all interesting, they are often one-offs that don’t emerge to Top 500-level systems. The most notable exception to this statement is the Mont Blanc project, which we have followed diligently over the years. This is ARM focused and built with native European integrator, Atos/Bull, which showed that Europe could indeed roll its own systems with homegrown tech. But perhaps the changing of hands with ARM has shifted the sentiment, that is hard to tell. We did talk to countless folks whose resumes include solid careers at ARM until relatively recently.
Another theme that was immediately clear, however, is that the partnerships between academic centers and industry are much tighter-knit, collaborative, and open than in the States. Not that it’s news, but nearly everyone we spoke with at HiPEAC could rattle off companies they had worked or talked with during development without shushing because of NDAs or pure-commercial partnerships. This is important because it means even the most obscure research endeavor can find actual users to test on and the sense of urgency to build something that works in real production (versus gets VC funding) changes.
On that note about different “flavors” of research focus, when it comes to solving large-scale computing challenges a few clear differences between where researchers and startups focus their attention in Europe versus the U.S., at least based on our travels to conferences like this. Most notable is a focus on the software infrastructure to fit various non-standard (X86) platforms, which is not surprising given the strength of ARM in Europe. What also surprised us was the focus on FPGAs as the development platform for acceleration, particularly in academic centers in Europe.
The takeaway, at least from the large number of workshops and technical presentations, is that Europe is putting much work into building the software foundation for next-generation applications in HPC and AI based largely on accelerators, including GPUs and FPGAs. As noted, FPGAs seemed to get quite a bit more attention than their graphics card cousins—everything from the build-out of prototype boards to the sharing of experiences of using cloud-based FPGAs (AWS F1 was mentioned in several talks).
Xilinx hosted a workshop on FPGA-based accelerated cloud computing that was quite well attended. A few of the attendees we talked to said that they had already been using FPGAs at their centers on-site and a couple had already taken the F1 instances for a spin. We also spent part of a day at the Workshop on Reconfigurable Computing, where Michaela Blott from Xilinx and Tobias Becker from Maexler provided an overview of use cases, challenges, and opportunities for FPGAs and accelerated workloads in HPC and AI. This was followed by several shorter presentations of various projects in academia building and deploying FPGA systems as well as using the F1 instances. The takeaway from this wide range of experiences shows that experimental installations have yielded success but so far seeing FPGAs make their way onto the broader HPC stage on much larger systems is still some time away—if it happens at all.
Speaking of FPGA development in Europe, the ExaNest project is nearing its conclusion following a mission to prototype novel technologies in networking and storage as well as unique processing engines. We talked about this program when it was new but focused more on the ARM-driven work (and yet more on that here as well) that was one of the included projects. FPGAs were also part of the novel technologies bent and that work has been paying off with functional prototypes as we learned at HiPEAC.
One of the leads for the reconfigurable side, Aggelos Ioannou, presented his team’s experiences bringing up Zynq UltraScale boards for application testing, which yielded some noteworthy performance results. He showed the high-density quad-FPGA daughter board the team developed and shared that the build process is complete with around 40 such boards running in production at their center in Greece with another 20 running at partner institutions and 20 more planned for their own center.
Forgive the photo quality from slides but the following shows the completed boards.
A project spun from the ExaNode program that is nearing its end was also presented, highlighting yet more work with reconfigurable hardware. The ExaNode project was presented by Yann Beilliard, who focused on a multi-chip module based on FPGAs for HPC applications.
This is another project that will produce a usable device with the novel interconnect for multi-FPGA use. Beilliard walked us through the design and prototyping of the device that interconnects 3 dies with the minimum usage of LVDS signal crossings as a key optimization. IBM is a manufacturing partner on the project, which will yield yet another useful device from the ExaNode project beyond those we have covered, mostly focused on ARM.
By the way, the programs that produced some of the more interesting prototypes—ExaNode, ExaNest, and EcoScale are all nearing the end of their lifespans and the talks were to present the near-final results of the effort. Forthcoming investment from Europe is indeed focused on delivering exascale systems as well as supporting development of quantum computing throughout the old world. More on that in yet another piece that got folks riled at yours truly.
Future investments in Europe are wide ranging and we expect that the same event next year might yield entirely different focal points. With truly productized ARM will we see a resurgence in larger-scale system builds, for instance?
This year it certainly came up in conversations but the program did not have much to cover the arrival of 64-bit ARM processors to HPC. Having the center of gravity shift following the acquisition might have something to do with it, but FPGAs, despite the big focus this week, will not be the broad saving grace for even close to as many applications. Open to answers on this and the likely complaint about someone’s single-institution prototype that is supposed to convince us this is not true.
Finally, for those who follow HPC in the States or Asia, this is one of those under-the-radar and very Euro-centric events (unlike ISC, for instance, which feels much more global and focused exclusively on HPC) that is worth a visit next year.
As a final (and slightly more esoteric) statement, and it could be the jet lag and (and outstanding Spanish wine) talking, it is hard not to think of how the same species of mice, for instance, separated on geographically isolated islands, evolve differently. They still look essentially the same but closer inspection reveals certain adaptations that make them uniquely suited to their environment. The same is true when it comes to research and HPC in two geographically distinct regions. Europe and the U.S. (and Asia as a third) build systems that do the same things, but the entire process is different. The funding models, the influence from industry, the ways certain technologies are evaluated and implemented.
And like everything else in the world these days, the differentiation and drawing inward is happening even more. The good news is that it means increased differentiation, which might just mean a better chance for regional survival. In more practical words, a better chance for Europe’s homegrown HPC efforts to thrive and provide greater competition.
Who knows what winds are shifting in European HPC research, perhaps this was just a slight Mediterranean breeze we detected here at HiPEAC in Spain that doesn’t carry the scent of real change, but this was our view from the ground.