As previously announced, the GPU specification syntax for PBS jobs has changed as of the Nov 4-5 maintenance. The gpu_type resource is no longer specified on the job, but rather on chunks within a PBS select statement. For example, here is how you would now ask for 1 node of 4 A100s on Casper:

#PBS -l select=1:ncpus=128:mpiprocs=4:ngpus=4:gpu_type=a100

Additionally, we have expanded the gpu_type options to include more generic and more specific requests. For example, if you want to get any NVIDIA GPU with 7.0 compute capability or greater (V100, A100, H100 currently), you can specify cc70 as your gpu_type. You will be assigned the first available GPU in that class. Alternatively, if you want to ensure your job runs on 80GB VRAM A100s only, you can now achieve this by giving a gpu_type of a100_80gb. For more details of how GPU selection works, see here, and for a full list of valid gpu_type values on Casper, see here.

We have also deployed a new software stack - ncarenv/25.10 - on both systems. This new environment features the latest versions of compilers, libraries, and tools, including a preview cray-mpich/9.0.0 on Derecho which is built upon MPICH 4. We anticipate this environment becoming the default in early 2026, so please give it a try in this preview period and provide us with your feedback.