Skip to content

Phoebe: CUDA software from EESSI works on gpu1 and gpu2

Since 4 October 2026, GPU-enabled software from EESSI can use the GPUs on the rocky10 nodes gpu1 and gpu2, and EESSI's CUDA and cuDNN modules work too. Until then GPU software from EESSI failed because it could not find the NVIDIA driver.

How to use it

EESSI is not loaded by default. In a job or an interactive session on the rocky10 partition, add its modules and pick a version:

module use /cvmfs/software.eessi.io/init/modules
module load EESSI/2025.06
module avail

module use takes the directory, not the EESSI/<version>.lua file inside it. On a GPU node EESSI also lists its CUDA builds. A quick test:

module load CUDA-Samples
deviceQuery

What works

  • EESSI 2023.06, 2025.06 and 2026.06: prebuilt CUDA software (for example CUDA-Samples) runs on the A100s.

  • The CUDA and cuDNN modules load as well. NVIDIA's licences do not allow EESSI to ship them, so they were installed on the nodes:

    EESSI CUDA cuDNN
    2023.06 12.1.1, 12.4.0 8.9.2.26-CUDA-12.1.1
    2025.06 12.6.0, 12.8.0, 12.9.1 9.5.0.50-CUDA-12.6.0, 9.10.1.4-CUDA-12.8.0, 9.15.0.57-CUDA-12.9.1
    2026.06 13.3.0 9.23.0.39-CUDA-13.3.0

  • EESSI modules do not set LD_LIBRARY_PATH. If you compile your own code against an EESSI library such as cuDNN, link it with an RPATH so that it finds the library at run time:

    module load CUDA/12.9.1 cuDNN/9.15.0.57-CUDA-12.9.1
nvcc -o app app.cu -lcudnn \
    -Xlinker --disable-new-dtags -Xlinker -rpath=$EBROOTCUDNN/lib:$EBROOTCUDA/lib64