Efficient and scalable software is in urgent demand in all of science and industry.  At the same time, we are facing an unprecedented change of computer architectures, driven by the end of Moore's Law and the market value of data-intensive Artificial Intelligence. Only solutions which are 1) inherently parallel, 2) data-local at all levels, and 3) robust in terms of reduced precision arithmetic will prevail.

Using extreme-scale seismic simulations as an example, I will illustrate a transferable and interdisciplinary approach to computational physics. Here, best-in-class performance is obtained by solving diverse challenges in the entire modeling and simulation pipeline. The effectiveness of this approach has been confirmed by a collection of high-impact milestones.

This includes the concurrent utilization of over 1.5 million Tianhe-2 cores, petascale local time stepping on SuperMUC-2, and a sustained full-application performance of 10.4 DP-PFLOPS on Cori. I will conclude the presentation by discussing the potential of public cloud offerings. Baseline is a recent petaflop simulation on 768 AWS c5.18xlarge instances, having over 5 SP-PFLOPS of theoretical computational power (21% of Theta).