The four supercomputers will be production supercomputers that replace (or upgrade) existing supercomputers at the sites. A production supercomputer is a system that has a main focus to solve scientific problems. It is not a research exascale machine or prototype. The supercomputers will be part of the pan-European supercomputing infrastructure (PRACE), so basically researchers from all over Europe can get access.
Procuring (buying) a new supercomputer is a complex process. The supercomputer centres want to acquire the newest technology, an innovative system, and get the best architecture for their users. But, because it is a production machine, it has to be very stable, easy to manage and secure. A procurement process can easily take a year or so and has to follow European rules for public tenders.
If it is already so difficult to procure one supercomputer, why would you then procure four systems in four countries in one procurement? One might expect that the four systems are more or less the same, and that this would be the reason for working together. However, that is not the case, the systems will be quite different. The scientific user base is different, and each centre has a different focus and expertise.
Actually, on a European scale, it makes sense the new supercomputers differ and each focus on a separate class of applications and users. This way, more scientific user groups can get access to a supercomputer that is suited for them. The European Commission also sees the benefits of such a combined approach with four machines, each with a different focus. That is why from the total investments for the four supercomputers - 73 million euro - the European Union funding will be 26 million euro, which is a considerable amount.
The investment - less than 20 million euro per system - shows it are not the exascale supercomputers Europe is working on. Actually, the goal is to have something like 50 Petaflop/s performance per system, which is a modest performance for the 2019 time frame, although it is a challenge to build them for the available amount of money. (The budget includes maintenance and support.)
Because there are 4 systems to be procured in four different countries, it legally, could become complicated. The overall procurement is lead by GENCI/CEA and governed by French law. Each machine procurement is governed by the law of the country the procuring centre is located in. So the Juelich machine is procured under German law. Actually, the final contracts are signed by the individual centres: it is only the procurement process which is done jointly.
If this type of joint public procurement is a success, it will be used in the next steps towards European exascale systems too. So large-scale exacale prototypes, pre-exascale systems, and an exascale computer could use the same procedure.
The systems should be installed in the 2019-2020 time frame, so the actual tender will be published next year. The tender will be open to European companies. However, this does not exclude American or Chinese companies like Cray or Lenovo. As long as "foreign" companies have a substantial research and development endeavour in Europe, they can participate.
The new supercomputer at CINECA should replace the MARCONI-KNL supercomputer currently installed. The new system should have a peak of about 50 Petaflop/s. CINECA is especially interested in power and performance: efficient power monitoring and management. The architecture of the new system should be a solution towards exascale: the network, sockets or nodes should provide path(s) towards exascale.
At Juelich, the new supercomputer should replace the JURECA Cluster system. The new supercomputer should be faster (6-8 times) than JURECA, but especially have an increased memory size (about 20 times bigger) of both volatile and non-volatile memory. Juelich is very interested in features for data-intense science and data-analytics. The JURECA Booster Module (developed in the DEEP-ER project) will be integrated in the new system.
The new supercomputer at GENCI will complement the current CURIE2 supercomputer. It should be a balanced architecture, not only delivering flop/s but also having a good network performance, memory foot print and I/O performance. The system will be used in the context of the current convergence between HPC, Data intensive computing and AI.
The Barcelona new system is actually a data infrastructure that should provide storage to the existing BS supercomputer MareNostrum4 and HPC clusters. In addition, it should contain nodes that are suited for HPDA: High Performance Data Analytics. The new system should include new storage and memory technologies and permit new data analytics paradigms.
As a next step, the PPI4HPC procurement team will talk privately under NDA with individual vendors to assess the possibilities in the market and know what is on the technology horizon. After the consultation phase, the tender will be drafted for publication early next year.