No supercomputer in the European Union in the top 3 of the world? For BSC that is a choice, says Mateo Valero

23 Jun 2018 Barcelona - No doubt, Europe is lagging behind in supercomputing compared to China and the USA, and perhaps even compared to Japan. But the often cited criterium that there is no supercomputer in the European Union may not be that correct after all. In a presentation at the RISC-V workshop in Barcelona, Barcelona Supercomputer Center (BSC) director Mateo Valero said that "with the money we use in the Mare Nostrum, we could be number 3. But we do not like to be number 3. We like to have hardware that our users can use in an easy way." So the Mare Nostrum could have been at the number 3 position in the TOP500.

Perhaps the criterium of having a machine in the top 3 is not a good one and scientists like Mateo Valero should have told the politicians, it is a bit more complicated and find criteria that make more sense.

There is a difference between buying the best machines for the European users and having those machines based on European technology. For the scientific and industrial users it does not really matter whether the machine is made in China or made in Europe. So giving European users access to the fastest machines in the world, and to have some of the fastest machines in the world based on European Union based technology are two different goals. These sometimes seem to get lost in the current discussions.

There are different grades of "European technology". A supercomputer with 100% European Union based technology is impossible. See for instance the interview with HLRS director Michael Resch in the Deutsche Wirtschaftsnachrichten earlier this month .

Much of the technology focus for EuroHPC is on the European Processor Initiative (EPI), a framework agreement in which the partners decide what happens, and there is less control from the European Union as a funding organisation. Of course, a processor is just one part of the technology needed. In general, when talking about processors one has the basic processor and an accelerator. In EPI, at least one of them will be implemented in RISC-V. To many this came as a surprise: The Horizon 2020 programme heavily invested in ARM technology. For instance, in the Mont-Blanc projects in most of which BSC played a leading role, some 50 million euro was invested in 10 years’ time. After a decade, the first HPC systems that provided a useful alternative to Intel-based machines came in sight. Building the ARM HPC software ecosystem was the most important challenge. However, ARM-based supercomputers are not anywhere near the top supercomputers in the world yet. Writing and testing software just takes time. Defining and implementing the right HPC extensions to ARM for HPC just takes time. But it will look like ARM-based supercomputers could be a reality in the near future thanks also to the European Union investments in ARM HPC.

So why throw all this ARM-based experience in Europe away? Mateo Valero knows the answer: Brexit and the Softbank acquisition, he says during his presentation. ARM was not considered European anymore. So that is why the EPI consortium looked at RISC-V. We would have liked to discuss this a bit more with Mateo Valero. ARM is not Intel, and RISC-V is not ARM. With Intel, an American company, you buy the chip and hence use the Intel Intellectual Property. With ARM you license the basic IP and can extend that IP and you have to produce the chip yourself. ARM is now British/Japanese. With RISC-V you license the IP from an American society, and then the same as with ARM. The difference is that license from the RISC-V society is an "Open Source" license and ARM's license is not. But would that make a difference if the US for one reason or the other decides no American technology may be used in EU supercomputers? Or if the UK decides the same? In reality it will be the same.

From a technological point of view, starting more or less from scratch with a new technology - RISC-V - seems a recipe for not getting the processor ready in time for a European technology-based exascale supercomputer in 2023. Not because of the amount of money (although that is quite modest) but because of the lack of time. And that is not good news for the users of supercomputers in Europe. Perhaps a scenario where there are still considerable investments in ARM technology alongside RISC-V would be worth investigating. In 2021 we could then look how far each technology has been developed and decide to put more effort in one or the other scenario. And let us consider: what are the odds that the UK or Japan would embargo technology for use in European Union-based supercomputers? Having the UK a member of EuroHPC would even lower that risk, and setting up an intense collaboration with Japan on ARM HPC technology - the next Japanese supercomputer will be ARM-based - could lower the risk there too.

Mateo Valero seems to agree: "Here is what we call the European Processor Initiative. The target expressed by the European politicians was to develop a real microprocessor, with the owner being a European company. This mean they exclude Intel, IBM, AMD, NVIDIA. For a person at BSC, who is collaborating with everybody in the world, this is not good, but these were the rules. For the first time in a European project, the owner of the hardware should be a European company."

So EPI is there to develop "European" microprocessor technology. But will the newly developed technology stay in Europe? Currently the consortium only counts European partners. But Mateo Valero wants to change this. He says the EPI consortium is open to new partners, even American ones it seems. "BSC is trying to get more partners in this consortium. Open to American companies, open to Japanese companies, to anybody who likes to collaborate with us." This would mean that American and Chinese companies, would also get early access to EPI-developed technology. Perhaps that is not a bad thing, but it is not clear how that would help to attain European leadership.

Ad Emmen