The session started, of course, with congratulations to Los Alamos National Laboratory, represented by Andrew White, and IBM, represented by Don Grice, for the achievement. As they explained it took about USD 120 million to develop. IBM won a procurement in 2006 to deliver a petascale computer. The system gets most of its speed from the CELL processor.Most of the software is build on open source, apart from an Software Development kit from IBM. Also the operating system is open source: Fedora Linux.

The system was operational on May 23 2008. Just three days later, on Monday May 2006, it successfully ran two Linpack benchmarks at 1.025 and 1.026 Pflop/s.

The system also did run some real application codes already. The VPIC, particle-in-cell code ran at 374 Tflop/s. The SPaSM Molecualr dynamics code ran at 361 TFlop/s: impressive achievements too.

But what about Japan? Satoshi Matsuoka who is participating in the construction of a smaller supercomputer the TSUBAME, pointed out that both the US and Japan have programmes in place to produce 10 Pflop/s machines. Japan has a development programme of close to USD 1 billion to produce such a supercomputer. So yes 1 Pflop/s is an achievement, but he expects that Japan could be the first to have a 10 Pflop/s machine, although he expects the USA to be not too far behind.

Michael Resch was asked to represent the European View. As a true European, he said he could not do that but he was willing to give his own opinion. In Europe, the road taken is different from Japan. The focus is on creating a full HPC eco system that supports the real supercomputers. So there is the HPC_Europe projects that gives hundreds of European Researchers the chance to work on the fastest supercomputers on the continent. There is the DEISA project that tightly links large existing supercomputing centres in Europe and there is PRACE. PRACE aims at providing the technical and legal (very important when a lot of countries are involved) basis for tier 0 top supercomputers in Europe, tier 1 supercomputer in countries and tier 2 supercomputers in regions. So the HPC ecosystem should take care there are researchers that can use the tier 0 supercomputers, and supercomputer centre that can provide the support. Michael Resch thinks this is a more sustainable approach than trying to take a one shot at the number 1 position in the TOP500.

Hort Simon pointed out that also in the US, the challenge is now to chance this success of reaching 1 Pflop/s on one machine into usable Petaflop/s for the masses: i.e. create many Pflop/s computers, port hundreds of applications to them and take care it can be used by thousands of researchers.

To close the session, Jack Dongarra provided details on a workshop on creating Exaflop/s machines. They should be around by 2019. How to build them? How to operate them? How to program them? Nobody nows yet. But at least a kind of road map and issues to be solved are emerging.