The Barcelona Supercomputer Center will soon be the proud owners of the Europe’s first pre-exascale supercomputer. If all goes according to plan, MareNostrum 5 will almost certainly be the most powerful system in Europe when it goes into production at the end of next year.
The new supercomputer is expected to deliver 200 peak petaflops, a seven-fold increase over the current MareNostrum 4. The big leap in computational capacity means the system will no longer fit in the Torre Girona Chapel, where MareNostrum 4 and all its predecessors have resided. For this fifth-generation machine, some of racks will be placed in the lower floors of BSC’s new corporate building, a few meters away from the chapel.
MareNostrum 5 will be one of three pre-exascale systems of 150 petaflops or more, along with five petascale systems of at least 4 petaflops, to be deployed in Europe over the next couple of years. The new machines are part a coordinated effort by the European Union and the EuroHPC Joint Undertaking (EuroHPC-JU) to bring European supercomputers on par with those in the United States, China, and Japan. Taken together, the eight additional systems will double the HPC capacity currently available to European researchers and industrial users tapping into EU-backed supercomputers
At 200 petaflops, MareNostrum 5 will be more than seven times as powerful as “Piz Daint,” the reigning supercomputing champ in Europe, located at the Swiss National Supercomputing Centre, and will be on par with the top ranked “Summit” system at Oak Ridge National Laboratory in the US. Of course, by the time the new MareNostrum is installed at the end of 2020, China may well have one of its exascale supercomputers up and running. And in 2021, both the US, with “Aurora” at Argonne National Laboratory, and Japan, with the “Fugaku” Post-K system at RIKEN, are scheduled to have their respective exascale systems deployed. Nevertheless, MareNostrum 5 will almost certainly be ranked in the top ten when its initially installed, and quite possibly in the top five.
For the time being, BSC is not offering much in the way of technical details regarding the physical makeup of the machine. According to the press release, it will be “a heterogenous architecture, probably based on two big clusters with different characteristics.” The idea behind having two different clusters is to be able to optimize things such as latency and energy efficiency on different kinds of workloads. But the “probably” reference alludes to the fact that the contract has yet to be awarded, which suggests there is a good deal of leeway on what the system has to conform to.
MareNostrum 5 will also incorporate “an experimental platform” to help develop new supercomputing technologies of European origin. The European supercomputing community, or at least the sub-community backing the EuroHPC-JU effort, is committed to developing core HPC technologies so that Europe no longer has to rely on external sources for its most basic supercomputing hardware – that is, the processing elements.
BSC director Mateo Valero says that the inclusion of this experimental platform will help address that goal and put the region on its own path in supercomputing. “Europe, because of security and sovereignty reasons, cannot proceed with the high level of dependence that it shows regarding computing technologies made in other continents,” explained Valero in a statement put out by BSC.
“We have always thought that Europe should work together in a technological area in which we compete with countries as powerful as the United States, China and Japan and we are very satisfied to be able to say now that Barcelona will have one of the first supercomputers included in the new European supercomputing roadmap,” he added.
At this point, the EU’s effort to develop indigenous supercomputing technology is centered on the European Processor Initiative (EPI), which is tasked to develop implementation of a general-purpose CPU, an accelerator, and special-purpose processors for Europe’s upcoming exascale systems. The initial design work for these chips, based on Arm and RISC-V, is already well underway.
Whether the experimental platform will incorporate some of this EPI work is unknown, but it’s hard to imagine where else it would come from. The only problem with this is that the Arm and RISC-V prototype chips currently under development aren’t scheduled to be ready until 2021, so the best that could be expected for in the 2020 timeframe would be prototypes of prototypes. Alternatively, the experimental platform could be tacked onto the system at a later date, after the Arm and RISC-V work is further along.
MareNostrum 5 will cost €223 million ($249.4 million), which includes the purchase price, installation costs, and five years of operation. Half of this will be funded by the EU, with the other half by four EuroHPC member states that partnered with BSC on their pre-exascale bid: Portugal, Turkey, Croatia, and Ireland. Spain’s Ministry of Science, Innovation and Universities and the Catalan Government are also kicking in an unknown amount money for the system.
That kind of pooling of funds and political will looks to be the way European supercomputing centers will be competing for these big, expensive systems going forward. There are 28 member countries in the EuroHPC-JU, and between them and the EU, €840 million ($939.5 million) has been budgeted for the three pre-exascale systems and five petascale machines. Given that, not every member will get to host its own supercomputer, but procuring these systems using multi-country consortia will broaden the funding base plus help insure all, or at least most of the member states have a machine to call their own.
Beyond the €840 million earmarked for the pre-exascale and petascale acquisitions, the EU has a proposal on the table to allocate €2.7 billion (just over $3 billion) for funding supercomputing in Europe from 2021 through 2027. This money will enable the Joint Undertaking to acquire multiple exascale supercomputers, as well as support the development of software and expertise needed for these machines. That’s probably comparable to what the other major supercomputing powers will spend in that timeframe. Whether it’s enough to put Europe on equal footing with them remains to be seen.