Page 1 of 1

Issues with cuInit() with VASP 6.4.0 with GPU support

Posted: Wed Jan 24, 2024 4:54 pm
by mrhermes
Posting this message from one of the admins of our cluster:
I have built VASP 6.4.0 with nvhpc 22.11, 23.3 and 23.11, which are
packaged with cuda/11.8, cuda/12.0 and cuda/12.3, respectively. However,
when I run an example script with vasp_std with 1 single MPI proc,
the run crashes with the error message::::

Failing in Thread:0
call to cuInit returned error 803::: Other

This error code indicates the mismatch errors with the GPU drivers. The GPU
driver on the compute is now 520.61.05, which comes with CUDA 11.8. We have
reverted the driver from 535.x (from CUDA 12.x), which also led to the same
issues.

Could you please help me with the issue? What is the recommended
combination of the GPU driver and the nvhpc version?

Re: Issues with cuInit() with VASP 6.4.0 with GPU support

Posted: Thu Jan 25, 2024 3:11 pm
by alexey.tal
Dear mrhermes,

I haven't seen this error, but it looks to me that it has to do with the CUDA version you are using.
Which CUDA version do you set in makefile.include?

We have not updated to 535.x driver, so I can't tell if this could be the culprit.
One of our machines is equipped with A30 with 525.105.17 driver and I am compiling VASP with CUDA 11.8, so this combination works fine.