NWSC-3 to bring significant GPU capability to Earth system modeling at NCAR

UPDATE 1/20/2021: CISL has announced the details of the next NCAR supercomputer.

NCAR is entering the final phases of selecting the system that will succeed the Cheyenne system in early 2022. While we can’t share details yet, we can say that the NWSC-3 system – the third petascale system at the NCAR-Wyoming Supercomputing Center (NWSC) – will be getting 20% of its peak computing capability via graphics processing units (GPUs).

The 20% GPU capability reflects a balance in system architecture that will allow NWSC-3 to run the current generation of Earth system models on familiar CPU-based nodes while also serving as a pathway to future NCAR systems. These future systems are likely to obtain much more of their computing capability from GPUs, and NCAR will be working over the lifespan of NWSC-3 to help the community migrate their codes to GPU architectures.

Aerial view of the NCAR-Wyoming Supercomputing Center

Aerial view of the NCAR-Wyoming Supercomputing Center in Cheyenne, Wyoming. A new system to be in place here in 2022 will get 20% of its peak computing capability via GPUs. Image: UCAR.

A recent example is NCAR’s effort, in partnership with The Weather Company, to port the Model for Prediction Across Scales (MPAS) to run on GPUs.

Why GPUs? In addition to being the predominant technology in near-term exascale systems, including Summit at Oak Ridge National Laboratory, GPUs represent a significant potential increase in energy efficiency for Earth system models as measured by flops/watt – that is, the number of floating-point operations per unit of electricity. This means eventually converting the NCAR HPC workload to run primarily on GPUs will allow us to continue procuring more powerful HPC systems while controlling the NWSC utilities budget as well as the carbon footprint of our modeling enterprise.

Over the coming year, NCAR will be providing more details on the NWSC-3 system; offering training courses to help the community with the transition to GPUs; and pursuing other efforts to ensure that our user community can make the most of the NWSC-3 GPU partition.

Users who want to get a head start with GPUs can request access to the Casper cluster, which has a small number of nodes with a total of 64 NVIDIA V100 GPUs.