It is now pretty much assured that the first three pre-exascale systems being installed in the European Union will not be powered by processors being developed under the European Processor Initiative (EPI).
The three upcoming pre-exascale machines – LUMI at CSC in Finland, Leonardo at CINECA in Italy, and MareNostrum 5 at BSC in Spain – are scheduled to begin installation in 2020 and be operational in late 2020 or in 2021. The first-generation Arm and RISC-V processors from the EPI effort, which were originally aimed at the pre-exascale supercomputers, are scheduled to be available in 2021, which means the timelines no longer match up.
According to the EPI people we spoke with at the SC19 supercomputing conference in Denver, their schedule has not changed; rather, it was decided that the first crop of pre-exascale systems should be deployed sooner rather than later. Which makes sense. Waiting for the first EPI processors would likely have pushed deployment of the pre-exascale systems out to 2022, significantly delaying access to those flops for European users. The first European exascale machines, which will start to appear in 2023, are still expected to use the second-generation EPI processors.
All of this is being managed under the EuroHPC Joint Undertaking (EuroHPC-JU), whose mission is to bring European supercomputers on par with that of the other major HPC powers (the United States, China, and Japan), as well as develop domestic high performance computing technologies for the EU market. That means the initial pre-exascale systems will only address the first goal. All three of those machines are expected to deliver more than 200 petaflops (peak), which, in the 2021 operational timeframe, should place them among the top 10 systems in the world.
At this point though, we only have rough idea of what processors will be going into the pre-exascale systems. For example, the LUMI system for CSC in Finland is going to be accelerated with GPUs in “a large number of nodes,” according to CSC’s LUMI webpage. But these GPUs could be supplied by Nvidia, AMD, or even Intel.
The choice of CPUs is even wider since X86 (Intel and AMD), Arm (Marvell, Fujitsu, and Ampere), and Power (IBM) are all viable options. LUMI will also include 60 PB of storage, with a large flash partition. Vendors are expected to be selected in the spring of 2020, followed by installation later in the year. LUMI is slated to be operational sometime in 2021.
We may have a better idea of what will be going into the Leonardo supercomputer for CINECA, since a new partition for the existing Marconi system is apparently going to be a forerunner of the pre-exascale machine. According to the CINECA website, the upcoming Marconi partition will be based on IBM Power9 and Nvidia V100 GPUs, which “opens the way to the pre-exascale Leonardo supercomputer expected to be installed in 2021.” Keep in mind that by 2020, the IBM Power10 CPU is scheduled to be launched and a GPU upgrade from the V100 is certainly not out of the question. On the other hand, they may opt for an all-AMD CPU-GPU combo as a number of big HPC centers have done.
According to different sources, including some here at SC19, Leonardo will deliver between 220 and 250 peak petaflops, spread over 5,000 compute nodes and laced together with a 200 Gb/sec interconnect. The system will also include 150 PB of storage and the entire machine is expected to consume 9 megawatts.
No such detail is currently available for the upcoming pre-exascale machine for BSC. As we reported in June, the MareNostrum 5 system is going to be a heterogenous architecture, although the specific processor types are unknown. The 200 petaflops system is scheduled to be first European pre-exascale machine to come online, with an operational date of December 2020. It is also going to be the only one of the three systems to incorporate an “experimental” platform, which apparently will be powered by some sort of European technology. Since the EPI processors definitely won’t be ready in 2020, that leaves us with something of a mystery.
The mystery might be solved by a recent bit of news. Earlier this month, BSC announced it had launched Laboratory for Open Computer Architecture (LOCA), an initiative to develop processors based exclusively on open architectures – RISC-V, OpenPOWER, and MIPS. While one year is far too short a time to come up with a homegrown implementation of any of these architectures, we can imagine BSC could cobble together a cluster equipped with existing off-the-shelf chips of these architectures and fold then into MareNostrum 5.
It remains to be seen what will become of the first-generation EPI chips that missed their date with the pre-exascale trio. There are five petascale supercomputers (of 4 petaflops or more) that the EU is funding in conjunction with the three pre-exascale machines. And if the timing is right to match up these smaller systems with the first-generation EPI silicon, perhaps these petascale systems could serve as testbeds for the future exascale supercomputers. Stay tuned.