nVidia CUDA with the wrong video card

In trying to build my first crypto rig I wanted to be able to get it functional with the GPU instead of the CPU.

I had originally installed an nVidia GeForce GT 710 but came to find out later that its not a supported GPU for use with CUDA per developer.nvidia.com.

Lucky for me I had a video card to “Upgrade” to, which was an nVidia GeForce GT 630, which supports Version 2.1 of CUDA-Enabled Compute Capability.

I did have to download and install the nVidia Linux X64 driver, and without knowing that the install required kernel-devel so it could built the driver correctly.

The only thing blocking me is it seemed my build of libxmrig-cuda.so was built with the CUDA 11.6 API and the Linux X64 driver does not support that. In fact, running nvidia-smi tells me that the driver supports CUDA version 11.4, so looks like i’ll have to rebuilt my driver with 11.4.

Add/Remove CUDA 11.6 -> 11.4

Now I have to undo what I’ve done and redo what I need to do.

First I need to “properly name” my build folder in xmrig-cuda. If I ever get a newer driver I won’t have to rebuild this driver.

$ mv build build-11.6

Now I have to remove the CUDA sdk:

sudo dnf remove cuda
sudo dnf clean all

I had to navigate to an older CUDA download, specifically for 11.4. The install instructions are pretty much the same, just with a different RPM:

wget https://developer.download.nvidia.com/compute/cuda/11.4.0/local_installers/cuda-repo-rhel8-11-4-local-11.4.0_470.42.01-1.x86_64.rpm
sudo rpm -i cuda-repo-rhel8-11-4-local-11.4.0_470.42.01-1.x86_64.rpm
sudo dnf clean all
sudo dnf -y install cuda-11.4.0-1

I was able to verify that the new cuda library was installed from symlinks created in /usr/local/cuda to point to /usr/local/cuda-11.4

I next had to reinstall the nVidia 470.*-series Linux Driver so that nvidia-smi could talk with it.

Finally I rebuilt and re-symlinked xmrig-cuda with the 11.4 CUDA library.

Leave a Reply