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

[oe@ifi.uio.no via RT]

hi again, vegard,

> 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

thanks, once more, for working with us, and for those additional modules!

please forgive me for pushing this point: but regarding cuDNN, i
suspect the above is still overly specific.  we would like to combine
cuDNN 7.6 with CUDA 10.0 (because a third-party module identifies that
specific combination as what they support, and in the CUDA space i
tend to not take chances regarding backwards compatibility :-).

for all i can see, cuDNN supports a relatively broad range of CUDA
versions, i.e. should work with both 9.0, 9.2, 10.0, 10.1, and the
latest and greatest 10.2:

https://docs.nvidia.com/deeplearning/sdk/cudnn-support-matrix/index.html#cudnn-cuda-hardware-versions

assuming the above is correct, could you make the
toolchain-independent version of cuDNN more general too?  i believe
getting these basic 'general-purpose' modules in place will be worth
it: in the NLPL community we are building our own ecosystem of
discipline-specific modules on top of these, and version dependencies
can cross-multiply quickly, in the worst case leading to incompatible
constraints higher up in the module stack.

best wishes, oe