Local refinement runs out of memory with box size 320 on workstation with RAM of 64 Gb. Ran successfully to the end once with box size 416 on the same machine, now consistently fails. Tried to use 384, 356, and the latest 320 box size. See snapshots from the latest run.
Can you please let us know how much GPU RAM (VRAM) you have available per GPU on the workstation?
I have 24 Gb of GPU memory on that station.
One more error in local refinement with 320 box size. This time I ran the job without NU refinement option, previous runs were attempted with that checkbox checked. And just an extreme case with resolution statement for GSFSC is attached as well.
I’m having what looks like the same problem. Local refinement crashes repeatedly with out-of-memory error at the local decomposition step (see below). However, memory usage never went above 20% of 125GB RAM (see below), and my box size is not particularly large (340 pixels).
This problem has come up before (3 posts on it already). I was wondering if there’s any fresh advice on how to fix this?
We are re-working the local refinement implementation as well as non-uniform refinement to try and minimize memory usage but unfortunately at the moment there does not appear to be a clear workaround.
It’s worth noting that in all the bug reports on this to date, it is GPU memory rather than CPU memory that runs out.