New Frontiers in Earth-System Modelling

The gradual progress in global numerical weather prediction includes a systematic approach
to assess and quantify the associated forecast uncertainty by means of high-resolution
ensembles of assimilation and forecasts. This involves simulations with billions of
gridpoints, the continuous assimilation of billions of observations, rigorous verification,
validation and uncertainty quantification, and it involves increasing model complexity
through completing the descriptions of the global water and carbon cycles. A particular
challenge arises from ensuring energy efficiency for these extreme-scale applications. This
talk will comprehensively describe the steps taken towards preparing complex numerical
weather prediction systems for potentially disruptive technology changes. This includes
adaptation to heterogeneous architectures, accelerators and special compute units, adaptation
to hierarchical memory layouts, increasing flexibility to use different numerical techniques
with fundamentally different communication and computational patterns, frontier research
on algorithm development for extreme-scale parallelism in time and in space,
and minimizing both time- and energy-to-solution. For example, a significant step towards
further savings both in terms of throughput and speed-up is provided by the impact on
simulations if numerical precision is selectively reduced in high resolution simulations.