Scheduled Downtime
On Friday 21 April 2023 @ 5pm MT, this website will be down for maintenance and expected to return online the morning of 24 April 2023 at the latest

GPU support for WRF/MPAs

WRF has some GPU-accelerated capabilities but it's largely decentralized and difficult to track down. Also not officially supported, unlike MPAS which does have official GPU support built in.

There are a number of papers that have found success in offloading individual WRF modules onto the GPU to dramatically speed up computation time. Check out Section G within the Literature Review here for a good overview:

Evaluation of GPU Acceleration for WRF–SFIRE - Benz (2021)
 
As a new WRF user and wannabe developer, I am wondering why there is not an official GPU approach. I am sure this has been studied and there must be reasons for it. Does anybody know why?
 
@ruecuil
It primarily comes down to the fact that we have limited resources to conduct all the coding it would take to make that transition, in addition to the fact that we are no longer actively developing the wrf framework.
 
AceCAST is a fully GPU-accelerated version of WRF developed and supported by a private company called TempoQuest. There are some limitations on supported options and source code access is limited (it is distributed as a pre-built executable) but it is a mature solution at this point that can have significant benefits for certain types of WRF workloads. You can find more information about it at AceCast Online Documentation Homepage — AceCast Docs documentation.
 
Top