After the workshop on International Cooperation following ISC 2016 in Frankfurt, Germany, we had the opportunity to talk with Thomas Schulthess from the Swiss National Supercomputing Center (CSCS) and from the Swiss Federal Institute of Technology - ETH Zurich about Piz Daint, Europe's fastest supercomputer ranked at no. 8 in the June 2016 edition of the TOP500, which is located at CSCS. Piz Daint is part of the bigger initiative that is being run in Switzerland, both in application software development investments and in vendor collaborations, like NVIDIA in this particular case. This allowed CSCS to build the system in 2013. Thomas Schulthess is happily surprised that Piz Daint is still in the top 10 and in the top spot of European supercomputers for such a long time. In the fall, Piz Daint will be upgraded.
Piz Daint is a package of both very aggressive application development - a lot of applications have been ported on GPUs in multi-processor architectures - and innovative system investment. With the modest investments that research has in Switzerland, the result is that Switzerland can still keep up with the leading systems in Europe.
Piz Daint gets its performance, both in terms of floating point operations and bandwidth, from the GPU accelerators and the high memory bandwidth of the memory that is on the accelerators.
From the beginning Piz Daint has been used in chemistry codes and climate applications. The climate applications in particular are interesting because these allowed the researchers to use COSMO, the regional climate model that is also used for weather forecasting by CSCS's partners, MeteoSwiss, the German Weather Service DWD and a number of other weather services. The development of Piz Daint, together with the climate codes, allowed the researchers to use the same technologies - GPUs and the COSMO implementation that they wrote - to design an appliance for weather forecasting. This is the machine that was put in operation last fall. The model that goes with it at MeteoSwiss is now in operation since April 2016 and is probably one of the - by far - most efficient and most performant weather forecasting systems in terms of cost-efficiency and energy-efficiency, that is running on GPU accelerators.
The COSMO model is running at high, kilometre-scale resolution, showing how the atmosphere runs over the Alpine region. It allows the weather service to make forecasts for the next couple of days. When you visit the MeteoSwiss website, you can get the forecast for the following day in Zurich. COSMO also keeps the airplanes flying in and out of Zurich and Geneva airport. The weather forecasts are derived from the simulations. You need the high resolution because of the many mountains in Switzerland. You have to resolve the valleys. In Switzerland there are, for example, the Föhn winds. In order to predict those, in order to predict a change from good to bad weather, you have to be able to predict these winds in the valleys. Therefore you need the high resolution.
Within all the communities in Switzerland, be it material sciences or climate, there are collaborations at the European scale and at the global scale. This is very simple. Switzerland is a small country with very high ambitions in technology leadership but you have to realize that there is no Swiss science. There is only science with a Swiss participation. This is basically driving the collaboration. The smaller the country, the more you have to collaborate with the others.