[CP2K-user] GPU testruns fails with: CUDA driver version is insufficient for CUDA runtime version

[Jörg Saßmannshausen]

Dear all,

over the weekend I had access to a GPU test machine so I could compare the 
performance between a RTX-6000 and a A100 GPU card. 

I have installed the following CUDA drivers:
$ nvidia-smi
NVIDIA-SMI 450.80.02    Driver Version: 450.80.02    CUDA Version: 11.0

I am using GCC-9.3.0 with CUDA-11.0.2

I am using the toolchain with the following commands:

$ ./install_cp2k_toolchain.sh -j 12 --no-check-certificate --libint-lmax=6 --
with-cmake=system  --with-libxc --with-libint --with-fftw=system --with-
openblas=system --with-libxsmm  --with-parmetis --with-superlu --with-quip --
with-plumed --with-gsl --with-libvdwxc --with-spglib --with-hdf5 --with-spfft 
--with-scalapack=system --gpu-ver=V100 --with-elpa=no --with-sirius=no

(I had problems with the ELPA and SIRIUS build, hence I disabled them).

Running the test jobs, *all* of them fail with:

"CUDA driver version is insufficient for CUDA runtime version"

The box has both cards installed and I am masking one with the
CUDA_VISIBLE_DEVICES
environment variable. 

I am puzzled of where I am getting wrong. Yesterday I tried it with both, the 
A100 and the RTX-6000 card using most of the above modules being installed on 
the system via EasyBuild, today I only tested the RTX-6000 card. 

Does anybody has a good idea? I don't have the box for much longer, as I said, 
it is only a test environment I got access granted to. 

Thanks for your help!

All the best from a mild London

Jörg