A conference on numerical analysis and scientific computing
for graduate students and postdocs from the Mid-Atlantic region.
The first US exascale capable systems capable of executing a quintillion floating point operations per second will be installed at Department of Energy leadership compute facilities as early as 2021. The current leadership systems use graphics processing units to deliver the majority of their floating point operations and already have theoretical peak performance rated at approximately 0.2 exaflops.
In this talk I will review the computational challenges for finite element flow calculations that must be overcome to achieve a respectable percentage of theoretical peak performance on the thousands of GPUs of these large systems. I will highlight the central roles that high-order approximation, physical flow model selection, and multigrid preconditioning play in reaching this goal. I will also discuss the progress that is being made within the Center for Efficient Exascale Discretizations to target exascale capable systems.
This work is funded in part by the John K. Costain Faculty Chair in the College of Science at Virginia Tech together with funding from the DOE Exascale Computing Project via the Center for Efficient Exascale Discretizations at Lawrence Livermore and Argonne National Labs.