AI at the edge

(Unfortunately, this presentation is currently only available in German.)

In naher Zukunft wird KI zu unserem Alltag gehören. So werden selbstoptimierende Maschinen, Spracherkennung sowie diagnostizierende Medizingeräte uns täglich unterstützen. Die zugrunde liegenden Algorithmen sind extrem leistungsfähig aber auch leistungshungrig. Was steckt wirklich hinter der KI? Wie kann diese «at the edge», also eingebettet im Endgerät laufen, ohne Daten in der Cloud zu rechnen? In diesem Talk erklären David Gschwend und Florentin Marty, wie Ingenieure von SCS für Ihre Kunden erfolgreich KI in eingebettete Systeme packen.

Dieser Vortrag wurde am 9. Juni 2020 im Rahmen der Embedded Computing Conference 2020 mit dem “Audience Choice Award” ausgezeichnet.

You have an idea, a project or a problem you’d like to discuss? Get in touch with me now!

    Contact David Gschwend

    Protected by reCAPTCHA. The Google Privacy Policy and Terms of Service apply.

    Graphic Processors (GPUs) offer better performance at lower costs and energy demand compared to regular processors (CPUs). A consortium linked to the Swiss National Supercomputing Centere (CSCS) in Lugano, the Federal Office of Meteorology and Climatology MeteoSwiss, the Centre for Climate Systems Modelling (C2SM) of the ETH Zurich as well as other partners has established the basis for the application of these advantages in the field of weather forecast and climate simulation.

    The huge market for computer games and the corresponding consoles has rapidly promoted the development of graphic processors. Instead of 2 or 4 cores, as they are found in common desktop processors, a modern graphics chip contains up to 2496 energy efficient cores. For years, scientists have been working on possibilities to make use of this computing power for scientific and technological applications. The following hurdles have to be overcome:

    • In order to achieve the required system performance, a large number of these graphic processors are necessary. Like in a classical supercomputer or cluster, the processors need to be closely assembled in rack systems all while being tightly linked, providing a high reliability. The CSCS, being one of the first purchasers of such a system, has essentially influenced its design and development. Since the commissioning of the machine named ‘Piz Daint’ at the CSCS, the fastest computer of Europe is now located in Switzerland.
    • The powerful GPUs are always used alongside generally applicable CPUs. This combination of different kinds of processors is called ‘hybrid computing’ and requires software adjustments, so that the various processing steps can be performed on the most suitable processor (GPU or CPU). Due to their design, consisting of many small processors, GPUs can only be fully used when thousands of tasks (‘threads’) are being executed simultaneously. The splitting of the computing task into many independent parts (parallelization) requires an exhaustive revision of the codes and can greatly increase its complexity. With the successful porting of the weather model ‘COSMO’, the consortium has realized calculating weather forecasts on hybrid computers for the first time. The ‘Domain Specific Language’ (DSL), developed by SCS, allows weather researchers to formulate their physical models in a most straightforward way, without the need of considering the architecture of the processor being used. The automated translation of this code, using our back end solution, generates a highly efficient code for CPU and GPU.

    Thanks to the progress mentioned above, the weather forecast in Switzerland’s small and demanding terrain can be calculated even more precisely in the future and exceptional events like strong thunderstorms are thus more predictable.



    The weather forecast by MeteoSwiss and many other European meteorological services are based on COSMO, the numerical weather prediction system. The ‘dynamic core’ of this model was reprogrammed by SCS. This was done in co-operation with the following key partners: MeteoSwiss, ETH Zurich and the Swiss National Supercomputing Centre (CSCS) in Lugano as well as Nvidia.

    The new implementation is based on a Stencil Library developed by SCS. This uses ‘Domain Specific Embedded Language’ (DSEL), which allows it to use very diverse processors such as Intel and AMD x86 as well as multi-core architecture such as ‘General-Purpose Graphics Processing Units’ (GPGPU) with maximum performance (performance portability). The performance speed of the weather model could be enhanced many times over thanks to this approach. Unfortunately, this won’t make the weather better, but at least it can be forecasted with a greater degree of accuracy.