I’m more bullish on GPU availability as specialized TPUs etc continue to advance. On the other hand the fractured GPU coding situation is also bad and doesn’t help. There was a push 5(?) years ago for domain specific languages that transpiled into CUDA, SYCL, etc, but lots of issues there too

Yes, agreed, but the other problem is the awful (non-existent?) automatic test suites for models. Migrating autonomously (and continuously maintaining two codebases) is a lot harder without one. If I had this project, that would be my first step. Bring models into 2010s code development

The other problem is a structural one: not many good options for using the same code across GPU + CPU means you end up with a lot of duplication or near-duplication. Add to that that most models are written in Fortran (and the CUDA Fortran documentation is Bad), leads to a nasty situation

A very small data point to emphasize this point: I wrote a mesoscale model in CUDA for a grad class project; the CUDA version was about 16x faster than the multicore CPU version for the same results. Models should have been rewritten for GPUs a long time ago, but that expertise costs $$$$ and time

Of course, Unidata has many more projects. This selection are just some of the lesser-known ones. I’m privileged to serve on the Unidata Users Committee, and I cherish getting to interact with the staff and am heartbroken.

This funding gap hurts the full weather enterprise, hurting our ability to forecast weather, improve models, and hurts AI competitiveness.

The Unidata Science Gateway is critical for education, providing resources for faculty and students.

science-gateway.unidata.ucar.edu

THREDDS and the LDM are a critically important part of our weather data distribution pipeline. These data are critical for forecasting.

www.unidata.ucar.edu/software/tds/

UDUNITS makes up the backbone of so much software that touches science and engineering

www.unidata.ucar.edu/software/udu...

Unidata builds AWIPS/CAVE for education, enabling universities to train future forecasters on the software NWS uses.

www.unidata.ucar.edu/software/awi...

The AWS Radar Level II repository? That is Unidata. Millions of $$ rely on this feed, including many in the AI space.

registry.opendata.aws/noaa-nexrad/

You were right, SatPy has gotten really easy to use! Thanks for sharing this.