Increasingly packages and libraries are written against the latest CUDA version, CUDA 13. However, CUDA 1.3 requires a recent version of the NVIDIA kernel drivers, which are not yet available on OrangeGrid.
Until the drivers can be upgraded there is a compatibility library that will allow code needing CUDA 1.3 to run, however only on "Tesla" models GPUs or later, which as of time of writing means the following GPUs in OrangeGrid:
- NVIDIA A100 80GB PCIe
- NVIDIA A40
- NVIDIA L40S
It is first necessary to have a Conda environment set up, see the Python documentation for information on how to install Conda.
Once Conda is installed, create a environment for CUDA 1.3 and the compat library:
eval "$(/home/$(whoami)/miniconda3/bin/conda shell.bash hook)"
conda create -n cuda31
conda activate cuda31
conda install cuda
conda install cuda-compatYou will also need to set an environment variable to ensure that the compatibility layer is found:
export LD_LIBRARY_PATH=/home/$(whoami)/miniconda3/envs/cuda31/cuda-compat
After that it should be possible to run packages needing CUDA 1.3 as well as
compile them using nvcc.
Adding the following line to submit files will ensure that code is run on a compatible GPU:
requirements = CUDACapability >= 8.0
Please email any questions or comments about this document to Research Computing at researchcomputing@syr.edu.