About two years ago, the EuroHPC Joint Undertaking (JU) selected eight host countries for its first eight systems. Now, those trees are bearing fruit – Slovenia’s 6.8 peak petaflops Vega system, for instance, is already operational. The stars of the show, though, are the three pre-exascale systems: Finland’s 375 Linpack petaflops Lumi system, Italy’s 249 Linpack petaflops Leonardo system and Spain’s still-to-be-detailed MareNostrum 5. At the HPC User Forum, representatives from CSC and CINECA provided updates on their respective pre-exascale systems as installation and operation dates draw ever nearer.
Sanzio Bassini, director of CINECA’s Supercomputing Applications & Innovation Department, provided the update on Leonardo, which is expected to cost up to €240 million. Bassini said the roughly 3,456 Atos Sequana nodes in Leonardo’s 3,456-node “booster module” will deliver 240.5 Linpack petaflops, while the “data-centric and general purpose module” – consisting of 1,536 nodes – will deliver 8.97 Linpack petaflops. The hardware in these Atos Sequana XH2000 nodes, detailed in prior releases, will include roughly 14,000 Nvidia A100 GPUs, Intel Ice Lake CPUs, Nvidia HGX baseboards and Nvidia InfiniBand 200Gb/s networking.
Bassini also shared that the nodes will be “95 percent” direct liquid cooled with warm water, and that the system will be accompanied by an “exabyte facility” targeting a capability of processing a terabyte of data per second.
Leonardo is to be housed in a new datacenter located in the newly constructed Tecnopolo di Bologna a couple miles away from Casalecchio di Reno, the current headquarters of CINECA. This, Bassini said, posed some challenges for efficient cooling: the system, he explained, would be “quite in the south of Europe, and with a summertime which is quite warm.” As a result, the datacenter – which will also house supercomputing resources for the European Centre for Medium-Range Weather Forecasts – is forecasting an eventual power footprint of 40MW. Still, Bassini said, the Leonardo team is aiming to deliver the system with a power usage effectiveness (PUE) of less than 1.1.
Evangelos Floros, a program officer for the JU, had recently said that CINECA was planning to install the booster module by August 2021 and complete configuration and testing by 2021; the data-centric module, meanwhile, was slated for installation in Q1 2022 and finalization in Q2 2022. According to Bassini, that’s still the case: “more or less eight, nine months from now,” he said of the booster module, followed by January 2022 for the data-centric module.
Pekka Manninen, director of the Lumi Leadership Computing Facility, offered some more details on that system’s installation process. CSC, the Finnish national supercomputing center, had previously elaborated on Lumi’s specs in fairly extensive detail. An HPE Cray EX-based system, Lumi will include AMD Epyc Milan CPUs and Instinct GPUs, with its primary GPU partition (Lumi-G) delivering 550 peak petaflops and supplemented by an array of partitions and storage systems: Lumi-C, a supplementary CPU partition with ~200,000 Milan cores; Lumi-D, a data analytics partition with 32 terabytes of memory and additional GPUs; and many petabytes of storage through Lumi-P, Lumi-O and Lumi-F. And, Manninen said in the HPC User Forum, “we do have aspirations of adding some emerging technology capacities for users to explore.”
With most of the node internals laid bare in advance, Manninen instead offered more details on Lumi’s impressive datacenter. “It’s in a former paper mill, [and] paper production [had] gone already to Latin America quite a while ago,” he said. “But the factory remains, and a capability of hosting supercomputers all the way to, say, 200 megawatts. So the paper machines used to consume a lot of electricity and this is still on the site and we can use it for IT capacity now – and then we can run this IT load with carbon neutral electricity.”
In contrast to Leonardo’s inconveniently warm summers, Finland delivers a decidedly cool climate around the year, offering free cooling for the entire system year-round, Manninen said. The power that the system itself uses will be 100 percent hydroelectric thanks to nearby hydro plants (though, Manninen said, the system could still tap into the grid if needed). He also touted the reliability of the local grid: in the last 38 years, he said, it had only experienced one two-minute outage.
Even the waste heat from Lumi won’t be wasted. “Lumi is not wasting the heat by transmitting it out, but collecting it and heating the surrounding city of Kajaani,” Manninen said. “In fact, 20 percent of the city will be heated up by Lumi’s excess heat. Somebody’s waste is somebody’s treasure.” All these factors combined, he added, make Lumi not only carbon neutral, but carbon negative, with an estimated negative CO2 footprint of 13,500 tons per year and a PUE of 1.03. (And, of course, the center will make money on the sales of the waste heat.)
Despite the supercomputer’s somewhat distant location – Kajaani is a solid day’s drive from the comparably balmy Helsinki – Manninen assured the audience that the system would be hooked into the backbone of Nordic networking, with multiple gigabit connections to GÉANT that Manninen said could easily be upgraded to terabit connections. “Kajaani may be remote on a map,” he said, “but it’s very well-connected.”
Per a prior session, Lumi-C (and much of the rest of the system) are slated for deployment by Q3 of this year; the much more powerful Lumi-G module, however, will have to wait until Q4 2021 or Q1 2022. Lumi will be hosting pilot use periods for select project areas (such as data-intensive computing and high-throughput computing) beginning in early September (for the first phase) and late December (for the second phase).
Two down, one to go. The mysterious MareNostrum 5, to be hosted by Spain’s Barcelona Supercomputing Center, is the third and (for now) final pre-exascale system commissioned through EuroHPC – but unlike the other seven systems, vendor details have yet to be shared. All we know so far is that as of 2019, MareNostrum 5 was planned as a heterogeneous system with a peak performance around 200 petaflops that would include an experimental platform aimed at the development of new technologies for future supercomputers.
Anticipation for the reveal is definitely building, with “coming soon” teases tracing back to late last year.