HPC SDK 22.9 + Cuda toolkit 11.8

The NVHPC SDK 22.9 was recently released, in the release notes, they say that , it includes cuda 10.2,11.0 and 11.7. On my system I have installed the cuda 11.8 driver from the cuda toolkit 11.8. I have a question.
Can I use the NVHPC SDK 22.9 with the cuda toolkit 11.8?

thanks for the help

Using a CUDA 11.8 driver should be fine given it’s compatible with binaries built with earlier versions of the CUDA tool chain.

For the CUDA 11.8 toolkit, unfortunately the time of the release didn’t quite align with NVHPC SDK 22.9’s release, but should be integrated into our next release in a few months.

To use a different CUDA toolkit than the ones that ship with the HPC SDK, you’d set the environment variable NVHPC_CUDA_HOME to the location of this toolkit. Using this method with CUDA 11.8 wouldn’t be supported, but only because it’s not been thoroughly tested as of yet. So if you really do need the 11.8 toolkit, give this a try and if something doesn’t work as expected, let me know and we can try to work through the issue.

-Mat

Hi everyone. I just had the same need to use nvhpc with cuda 11.8 and after some searches & tweaks, made it work.

In my case I have an RTX4090 at hand and firstly tried to run a program with nvhpc and its build-in CUDA 11.7. It gives an error of cuFFT and after some googling it seems like the RTX 4090 only works with CUDA 11.8.

So I re-compiled the program with CUDA-11.8 by changing the option in the makefile.include from cuda11.7 to cuda11.8 — and of course it didn’t work. The reason is that cmake was trying to find 11.8 in /opt/nvidia/hpc_sdk/Linux_x86_64/22.9/cuda/, but only 11.7 is present.

Then I re-installed 11.8 in /opt/nvidia/hpc_sdk/Linux_x86_64/22.9/cuda/11.8 and tried to re-compile again. This time the error says it cannot find nccl, as nccl is provided via nvhpc and cuda 11.7.

The last step would be download and install nccl. Then the magic happens.

Hopefully my experience could be helpful for other people.

We now support an environment variable, NVHPC_CUDA_HOME, that you can use to point to a different CUDA version for use with the NVHPC SDK compilers.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.