The EuroHPC Joint Undertaking (JU) serves as Europe’s concerted supercomputing play, currently comprising 32 member states and billions of euros in funding. In June 2019, EuroHPC selected eight supercomputing centers across the European Union to serve as hosts for EuroHPC’s first eight systems.
Today, Nvidia is announcing that it will be powering at least four of those eight systems: Leonardo, hosted by CINECA in Italy; Meluxina, hosted by LuxConnect in Luxembourg; EURO IT4I, hosted by the IT4Innovations National Supercomputing Center in the Czech Republic; and Vega, hosted by IZUM in Slovenia. All four systems will utilize Nvidia Ampere GPUs and Nvidia HDR InfiniBand networking, and Atos will serve as the prime contractor for all but EURO IT4I (which will be built by HPE).
Meluxina is expected to deliver over 18 peak petaflops; EURO IT4I, 15.2; and Vega, 6.8 – all very respectable targets that should earn some recognition from the Top500 list. CINECA’s Leonardo system, however, is the belle of the ball: a computing juggernaut expected to serve up over 200 peak petaflops and cost up to €240 million (half from EuroHPC and half from the Italian Ministry of University and Research). Leonardo is one of three planned pre-exascale systems announced by the EuroHPC JU.
The specs and the siting
Leonardo will contain a whopping ~14,000 Nvidia A100 GPUs distributed among ~3,500 Atos Sequana nodes, each equipped with four A100s and a single Intel CPU slotted on Nvidia’s HGX baseboard. The water-cooled nodes will use Nvidia Mellanox HDR 200Gb/s InfiniBand networking, which CINECA says is boosted by smart in-network computing acceleration that enables low latency and high data throughput. Per Nvidia, this acceleration yields a 7X speedup in reduction operations, a 2X speedup in MPI performance and a 10X speedup in data movement among remote GPUs. (The other Atos-built systems are expected to use very similar configurations, apart from obvious discrepancies in scale; the HPE-built system, meanwhile, will be based on HPE’s Apollo 6500 configuration.)
CINECA – one of the leading supercomputing organizations in Europe – will host Leonardo in a new, energy-efficient datacenter that is being built in Bologna Science Park. Subject to the timely construction of the datacenter, Leonardo is scheduled to begin deployment in 2021 and become fully operational sometime in 2022. Once operational, CINECA will use Leonardo for a wide variety of research activities.
What will Leonardo be doing?
“Like any other national-level supercomputer, [Leonardo will] be used, really, for a broad range of research,” said Marc Hamilton, vice president of solution architecture and engineering at Nvidia. CINECA, for its part, highlighted possible uses in drug discovery, space exploration, weather modeling, material sciences and climate simulations. Indeed, just a few days ago, the European Commission announced that one of the three pre-exascale EuroHPC supercomputers would host its moonshot “Destination Earth” project, which aims to create a massive, cloud-based simulation platform to represent Earth’s interconnected ecological and social systems.
“CINECA plays a critical part in evolving both the research and industrial community in accelerated HPC application development,” said Sanzio Bassini, director of the HPC department at CINECA. “The Leonardo supercomputer is the result of our long-term commitment to pushing the boundaries of what a modern exascale supercomputer can be.”
Leonardo will also see use by CINI, an AI-focused coalition of universities across Italy that is currently one of the top users of CINECA’s top system (the 21.6 Linpack petaflops Marconi-100, which ranked 9th on the most recent Top500 list). CINECA expects that much of this current research on Marconi-100 will continue virtually without delay once Leonardo is installed, as the new system will run the same CUDA software stack as its predecessor. Leonardo will also almost certainly continue CINECA’s research in another crucial area – one being tackled by nearly every research supercomputer in the world.
“In today’s day and age, one has to mention the work of supercomputers on COVID-19,” Hamilton said. “The current CINECA Top500 system, like so many others in the world, is being used for COVID-19 research. Interestingly, CINECA has been doing (of course) academic research, but they’re also very closely teamed … with some of the pharmaceutical companies in Italy, working on cures for COVID-19, vaccines for COVID-19 and other research.” (More on CINECA’s battle against the coronavirus.)
The game of (AI) thrones
Nvidia anticipates that Leonardo will deliver 10 exaflops of half-precision AI performance (with the A100’s structural sparsity feature enabled), and everyone involved is eager to crown the new system as the world’s fastest AI supercomputer. While results from HPL-AI and MLPerf benchmarking will need to wait for a real-world system, Nvidia expressed confidence that Leonardo will perform well and stressed the necessity of strong AI performance in modern supercomputing.
“Supercomputers have been measured, historically, by their ability to perform simulation,” said Ian Buck, vice president and general manager of accelerated computing at Nvidia. “But with the advent of AI, we now have a new metric for measuring supercomputers. … And as a result, the performance of our supercomputers has exploded as the computational power of them has increased exponentially with the introduction of AI. Today’s modern supercomputers are AI supercomputers – they must be, to be an essential tool for science.”
In terms of the Top500 list, Hamilton said that if Leonardo became operational today, it would rank as the fastest system in Europe. Still, he cautioned that estimating a future ranking is “more of an art than a science” and that “we’ll have to wait and see” where Leonardo lands when it becomes operational.
Looking to the future
The back half of the eight announced EuroHPC systems include the four-petaflops PetaSC system, hosted by Sofia Tech Park in Bulgaria; the 200-petaflops LUMI system (update!), hosted by CSC in Finland; the ten-petaflops Deucalion system, hosted by the Minho Advanced Computing Center in Portugal; and the 200-petaflops MareNostrum 5 system, hosted by the Barcelona Supercomputing Center in Spain. These systems are still undergoing procurement, with vendors likely to be announced by the end of the year.
Overall, Nvidia hardware continues to pervade the Top500 list, with 333 Top500 systems using Nvidia GPUs and/or Nvidia InfiniBand networking. Amid challenges to this ubiquity – such as AMD’s overall resurgence and prominence in exascale plans, Fugaku’s chart-topping, Nvidia-less performance or Intel’s forthcoming GPUs for the exascale Aurora system – Nvidia is moving fast to maintain and expand its favorable position. This is perhaps most visible in Nvidia’s planned $40 billion acquisition of Arm.
If approved, the Arm acquisition would give Nvidia crucial IP for the European Union’s exascale plans. It would also create a formal relationship between Nvidia and SiPearl, the company producing an Arm-based microprocessor for European supercomputing under the auspices of the European Processor Initiative (EPI). Between the Arm acquisition and these new EuroHPC wins, the message is clear: Nvidia’s foothold in European supercomputing isn’t going anywhere anytime soon.