The EuroHPC Joint Undertaking (JU) was formalized in 2018, beginning a new era of European supercomputing that began to bear fruit this year with the launch of several of the first EuroHPC systems. The undertaking, however, has not been without its speed bumps, and the Union faces an uphill battle as it aims to establish both supercomputing leadership and supply chain independence in a competitive landscape. At the virtual HPC User Forum this week, Leonardo Flores Añover, senior HPC expert for the European Commission, updated the community on the present and future of the ambitious initiative.
“This presentation is very timely, because the European Union recently approved the continuation of EuroHPC for the next period,” Añover said. Indeed, the JU – which now comprises 33 member states, including all of the nations in the European Union – was just budgeted for the next seven years, spanning 2021 through 2027.
Successes and setbacks
The renewal comes as no surprise. Several years ago, EuroHPC worked to procure five petascale supercomputers and three pre-exascale supercomputers, representing an acquisition investment of some €510 million by the European Commission (which was supplemented by investments from the host nations). “This is probably the most outstanding achievement of EuroHPC,” Añover said. “The procurement of eight world-class supercomputers in a very short period of time.”
The JU has now successfully launched four of those petascale systems, which all made their debut on the Top500 in June. Those include the Bulgaria-hosted Discoverer system (six peak petaflops); the Slovenia-hosted Vega system (6.8 peak petaflops); Czechia-hosted Karolina system (15.2 peak petaflops); and the Luxembourg-hosted Meluxina system (18 peak petaflops). The Portugal-hosted Deucalion system (10 peak petaflops), meanwhile, “is expected to be installed in December of this year,” Añover said.
The pre-exascale systems have faced a somewhat more troubled road. Lumi, a ~375 Linpack petaflops system to be hosted by Finland, is leading the pack. “The expected installation date for the CPU module [of Lumi] will be November this year,” Añover said. “Then the accelerated module … will be installed in March 2022.” While Lumi will still be the first of the pre-exascale systems to launch, this is a bit of a slip from the timeline outlined six months ago: there, Lumi’s CPU module was targeted for a Q3 launch.
Italy’s ~250 Linpack petaflops Leonardo system also appears to be slightly behind schedule: its booster module is now slated for January of next year (versus a prior December target), but the remainder of its modules are still slated for May – well within the Q2 2022 target outlined in March.
Spain’s MareNostrum 5 system, however, remains the true problem child of the eight systems. “As you know, the procurement for the MareNostrum 5 system was canceled,” Añover said, addressing the disappointing development that emerged amid rumors of contentious political disagreements between representatives from Spain (who allegedly wanted to prioritize MareNostrum 5’s power) and representatives from the European Commission (who allegedly wanted to prioritize the use of European-made components in the system).
Despite more details of the long-anticipated system seemingly emerging in June, after the cancelation of the procurement, Añover said that the tender remains canceled and that they “expect that it will be reopened in autumn,” leading to a selection “early in 2022.” “Unfortunately, I cannot give you more details on what would be the requirements on the MareNostrum 5 tender yet,” he said, “so please stay tuned for more information in the following weeks.”
Establishing supercomputing sovereignty
If, indeed, the MareNostrum 5 tender was canceled over a disagreement on the use of homegrown technology, it’s no wonder: as the rest of Añover’s talk illustrated, the European Commission has a fervent and growing desire to establish digital sovereignty.
“The most important development [for EuroHPC] has been the definition of European digital strategy,” Añover said. This strategy, he explained, spanned everything from improving digital literacy to producing at least 20 percent of the world’s cutting-edge semiconductors.
“We have to focus on European sovereignty by reducing our dependency on other regions of the world for critical technologies and infrastructures,” he said. “Although the situation is better in the last couple of years thanks to EuroHPC, Europe is still underinvesting in supercomputing.” Añover pointed out that the EU only hosts around 18.6 percent of the supercomputers on the Top500, while the U.S. hosts nearly 25 percent. Further, he said, “no European player produces general purpose processors or accelerators, and this is reflected in the fact that no EU processor technology is present in the Top500 supercomputers.”
“This production [of innovative HPC systems in Europe],” he said, “has to be based on a European supply chain of components, technologies and knowledge.”
The increasingly defined European digital strategy has, therefore, translated to a bold vision of EuroHPC’s future in the coming seven years. Powered by an expected budget of around €7 billion (combined public and private investment), the JU’s new mission includes “all the elements that were already present in the previous mission,” Añover said, alongside a series of new elements: most notably, quantum computing.
EuroHPC enters the quantum realm
“Europe aims to have its first quantum-accelerated computer at the end of the decade,” Añover said. “The intention is to create a European hybrid classical and quantum infrastructure. … At the later stage, we hope that the maturity of the technology will allow the deployment of prototype quantum computers with error correction and robust qubits.”
Not all of this will fall under EuroHPC, Añover hedged. Research and innovation activities will fall under the Horizon Europe program that birthed EuroHPC, while efforts to acquire and operate the quantum technologies will, in turn, be EuroHPC’s domain. A pilot project, Añover said, is working toward deploying European quantum simulation infrastructure “of around 100 plus interactive quantum units” and making that infrastructure non-commercially cloud-accessible.
These efforts, of course, complement EuroHPC’s primary and ongoing objective to deliver powerful, homegrown supercomputers. According to Añover’s “very rough approximation” of the JU’s plans for the next seven years, EuroHPC plans to deliver additional pre-exascale systems and two full exascale systems between 2021 and 2024, with France and Germany each intending to host one of the exascale systems. From 2025 to 2027, the JU plans to generally deliver exascale and post-exascale systems.
Further, Añover said, the renewal also leaves room for EuroHPC to deliver “industrial-grade” supercomputers. “What we mean by industrial-grade supercomputers are mid-range supercomputers that are specifically designed with requirements for industrial use, such as security, confidentiality and data integrity,” he said. Añover also referenced ongoing pilot projects, including a project to test the in-development Arm-based Rhea CPU with “a wide range of commercial accelerators,” and a second pilot working on developing and testing European RISC-V accelerators.