[NLPL Task Force (A)] [uninett.no #202600] updates of CUDA Toolkit and cuDNN on Saga

[Vegard Eide via RT]

ma. 27. jan. 2020 11.26.29 skrev oe at ifi.uio.no:

    > We have installed
    >
    > CUDA/10.1.243-GCC-8.3.0
    > cuDNN/7.6.4.38-gcccuda-2019b

    many thanks for the quick follow-up, vegard!

    i am wondering about the dependencies on specific GCCcore versions,
    are they actually necessary in this case?  i am guessing these modules
    provide the precompiled libraries and binaries distributed by NVDIA,
    which in principle should be compatible with a bit of a range of
    versions for GCC, binutils, and friends, or?  for RHEL 7.6, for
    example, the CUDA Toolkit pages seem to only require a minimum GCC
    version of 4.8.5.

    i am asking because these kinds of dependencies can get in the way of
    'mixing and matching' of modules, so if there were no actual
    dependency on one specific version of GCCcore, it might make things
    easier for us to not require GCCcore 8.3, so that one could for
    example do something like:

    $ module purge; module load Python/3.7.2-GCCcore-8.2.0
    cuDNN/7.6.4.38-gcccuda-2019b

    could one imagine rewriting the CUDA modules in a manner that would
    inspect the current environment of active modules and just accept any
    suitable version of GCC and binutils, if preloaded, and only bring in
    GCC 8.3 if needed?

    with thanks in advance, oe



Hi,

The GCC dependency is there because they are part of toolchains (gcccuda,
fosscuda). But I agree, we should in addition have versions not tied to any
specific GCC version (i.e. toolchain version). We have installed
 

CUDA/10.1.243
cuDNN/7.6.4.38-CUDA-10.1.243


Regards
Vegard