Abstract
● U.S. current and upcoming Exascale machines are GPU based.
● Exascale readiness requires porting ESMs to GPUs, involving large efforts: either new code or major refactoring of existing code.
● No dominant programming model to date. Approaches include rewriting in Domain Specific Languages, C++ with template based performance portability libraries and refactoring existing Fortran code to support GPU directives.
● Several atmosphere models are GPU-ready, with some progress on other components. No full ESM is GPU-ready as of 2024.
● Preliminary results from one of the first GCRMs to run on an exascale system show that modern GPU nodes can be ~6x faster than modern CPU nodes, or ~3.5x faster per Watt.
● We did not attempt to estimate GPU performance vs CPU performance on a per cost basis.
● Exascale readiness requires porting ESMs to GPUs, involving large efforts: either new code or major refactoring of existing code.
● No dominant programming model to date. Approaches include rewriting in Domain Specific Languages, C++ with template based performance portability libraries and refactoring existing Fortran code to support GPU directives.
● Several atmosphere models are GPU-ready, with some progress on other components. No full ESM is GPU-ready as of 2024.
● Preliminary results from one of the first GCRMs to run on an exascale system show that modern GPU nodes can be ~6x faster than modern CPU nodes, or ~3.5x faster per Watt.
● We did not attempt to estimate GPU performance vs CPU performance on a per cost basis.
| Original language | American English |
|---|---|
| State | Published - May 1 2025 |