Cuda version issue

Install, Configure and Update
1

Hey everybody.

I have a workstation set up with RTX A5000s in them. The driver they seem to be accepting is nvidia 525. However, the corresponding CUDA version is 12.0
Until a few months ago, this system was running Cuda 11.7 with 515 driver which got removed when we performed an overhaul and now if I try installing the same driver (using sudo apt install nvidia-driver-515), 525 gets installed by default which doesn’t seem work with CUDA 11.7.

Is there a way to get around this issue, or is there any cryosparc version that runs with CUDA 12?

Thanks for the help

2

One needs to distinguish between the nvidia-driver’s associated CUDA version and the version of the CUDA toolkit that is specified with --cudapath during installation or CRYOSPARC_CUDA_PATH inside cryosparc_worker/config.sh. I have run CryoSPARC jobs with an nvidia driver as recent as 535, but a CUDA v11.x toolkit. You may want to ensure that the toolkit to which CRYOSPARC_CUDA_PATH points is not modified inadvertently during OS updates. If the toolkit at CRYOSPARC_CUDA_PATH did get modified, ensure you install a CryoSPARC-compatible version of the toolkit at a path that is unaffected by system updates and run

cryosparcw newcuda /path/to/compatible/cuda

(guide).
You should not run the newcuda command if you have installed 3D Flex dependencies.

1 Like
4

(In response to a post that has since been deleted)
Take, for example, a fresh, fairly “minimal” ubuntu-22.04 installation. Then you could

sudo apt update
sudo apt upgrade
sudo apt install nvidia-driver-535

Then reboot computer, and run the following commands as a non-privileged user.

# _not_ as root
nvidia-smi # ensure your GPU(s) is (are) "found"
cd /home/my_home/
mkdir sw
cd sw/
wget https://developer.download.nvidia.com/compute/cuda/11.8.0/local_installers/cuda_11.8.0_520.61.05_linux.run
bash cuda_11.8.0_520.61.05_linux.run --silent --toolkit \
   --toolkitpath=/home/my_home/sw/cuda-11.8 \
   --defaultroot=/home/my_home/sw/cuda-11.8 \
   --tmpdir=/home/my_home/sw/

(details).

2 Likes
5

@wtempel Thank you for replying.

I had tried installing 11.7 anyway, but the installation showed an error complaining about the driver mismatch. If that is not fixable, I will try installing 11.8 at a different path location and try again.

Also, what corresponding changes will have to be made to get Topaz running again if I decide to install 11.8 separately?

Thanks again.

6

If you installed topaz inside a dedicated conda environment and are using a wrapper script, no changes should need to be made, as topaz should not depend on the newly installed v11.8 toolkit.

7

@wtempel

Thank you so much. I installed cuda 11.8 separately and Cryosparc is working now.

1 Like