European supercomputing centers are known for deploying innovative architectures and working on cutting-edge applications, but the sheer number of these systems lags rather far behind the U.S. and China.
Europe currently only has one top ten machine, the Piz Daint supercomputer in Switzerland.
While the number of new centers or machine counts might not change much in coming years, the European Commission’s Horizon 2020 research budget has set aside almost a billion dollars to provide two grand-scale exascale systems by 2023.
The project, which is set to begin in 2019, is managed by Luxembourg-based EuroHPC, which will get 486 million Euro from the Horizon 2020 organization and the other roughly half from the EU’s member states. The Horizon 2020 effort has other investments planned for post-exascale systems, including early evaluation of quantum computing, among other architectures.
According to EuroHPC, these financial resources will be used to launch activities structured around two main pillars: an Infrastructure pillar, including activities for the acquisition, deployment, interconnection, operation and access time management of world-class supercomputing and data infrastructures; and a Research and Innovation (R&I) pillar, including activities to establish an innovation ecosystem addressing hardware and software supercomputing technologies and their integration into exascale supercomputing systems, advanced applications, services and tools, skills and know-how.
The goals of the funding are two-fold. First, to establish Europe as an HPC leader with two top-ranked systems and at least two other smaller systems that would perform in the top 25 on the Top500 list. These are for use by “Europe’s private and public users, scientific and industrial users, and for use in more than 800 scientific and industrial application fields.”
The group also hopes this will bolster European high tech industries. The goal is to “develop a European supercomputing ecosystem, stimulating a technology supply industry, and make supercomputing resources in many application areas available to a large number of public and private users, including small and medium-sized enterprises.”
The first set of goals to rank Europe higher are achievable with enough funding, which appears to be in place. By 2023, these two new largest systems might be among the fastest, but it is not likely they will be the first to reach exaflop capability. The U.S. and China are on track for their own systems with that peak potential, but all of this comes at a time where more supercomputing centers are questioning the value of reaching exascale for its own sake. The real development required to keep scaling the largest HPC systems is needed on the software side and with power budgets in the many millions of dollars required to feed these beasts, the race is on elsewhere (Europe included) to find more efficient architectures.
The second is quite a bit more complicated since there are very few companies that make HPC systems in Europe. The first that comes to mind is Bull/Atos, but all of their chips must come from the U.S. or China (and Japan with Fujitsu). There is, of course, Arm, which is based in the U.K., but any use of Arm in European supercomputers will need to come through a licensee or supplier of Arm chips (Qualcomm, for example).
So while Europe might achieve some leadership in the upper tier of the Top 500, the chip and OEM ecosystem it relies on is rooted in the two places Europe wants to compete with more fiercely—the U.S. and China.
It is not likely that Europe can cobble together a highly competitive native ecosystem of chip vendors and OEMs in time for a full-scale system deployment in 2023. And for any new company to take on the massive upfront investment to deliver a supercomputer that won’t pass acceptance for several months to a year? That is hard enough for established HPC OEMs to do, as anyone who follows companies like Cray well understands.
It is also worth noting that the exascale goal that is much-touted worldwide might not have staying power as a reasonable concept beyond a few systems (and only a few problems can take advantage of that scale anyway). While Europe might be able to place a few more higher ranked systems at the top of the Top 500 charts that can be termed “exascale capable”, by the time 2023 rolls around that FLOPs-based metric will have lost steam for most users entirely. That is already happening, as evidenced by interest in companion benchmarks like HPCG which provide a balanced performance, efficiency, and data movement metric for large systems and more important, for real-world applications.
What really matters for European supercomputing is a continued emphasis on doing some of the leading-edge work on new platforms so that they can pave the way for new directions in HPC architectures. They have ample incentive, given the lack of native vendor ecosystems, after all. That is perhaps not such a drawback since it frees European centers to pursue other routes and build new ideas internally without waiting on the ever-slower tick tocks the U.S. does. But it all goes back to the problem of risk-taking as a vendor since this is at the most extreme scale possible.
As programs in the U.S. like PathForward (among many others) have shown, however, is that with enough government gusto, a robust tech ecosystem can be built (or at least bolstered) to support the goals at hand. It will be interesting to see how and where Europe invests in ways that trickle into the small but vibrant European hardware and software vendor ecosystems.
To date, the following European countries have committed to joining the Joint Undertaking: Austria, Belgium, Bulgaria, Croatia, Czech Republic, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Ireland, Italy, Latvia, Lithuania, Luxembourg, the Netherlands, Norway, Poland, Portugal, Romania, Slovakia, Slovenia and Spain. More here.