I believe that one link might not be working properly. Whereas I have no problem with https://developer.download.nvidia.com/hpc-sdk/20.9/nvhpc_2020_209_Linux_x86_64_cuda_11.0.tar.gz, the multiple CUDA installation download https://developer.download.nvidia.com/hpc-dk/20.9/nvhpc_2020_209_Linux_x86_64_cuda_multi.tar.gz doesn’t seem to be working this morning. Thanks.
You have a typo in your link, “hpc-dk” should be “hpc-sdk”.
My bad. Somehow while copying and pasting the ‘s’ got erased and I didn’t notice (even after checking twice). Thanks.
I have another question about the installation. [Please let me know if you want me to change the title to the topic -as the old one was caused by my own sloppiness- or even if you prefer me to open a different topic]. The issue concerns the ‘libcuda’ library. The standard CUDA installation usually (as far as I’ve seen) places this library at
/usr/lib64. However, nvhpc is placing it at a new subdirectory
/opt/nvidia/hpc_sdk/Linux_x86_64/20.9/cuda/11.0/targets/x86_64-linux/lib/stubs that has the following components:
-rwxr-xr-x. 1 root root 48992 Sep 30 05:20 libcuda.so
lrwxrwxrwx. 1 root root 10 Nov 11 00:04 libcuda.so.1 → libcuda.so
-rwxr-xr-x. 1 root root 38680 Sep 30 05:21 libnvidia-ml.so
lrwxrwxrwx. 1 root root 15 Nov 11 00:04 libnvidia-ml.so.1 → libnvidia-ml.so
-rwxr-xr-x. 1 root root 3840 Sep 30 05:21 libnvrtc.so
lrwxrwxrwx. 1 root root 11 Nov 11 00:04 libnvrtc.so.11.0 → libnvrtc.so
The linker is not picking up the cuda library and even the command
ldconfig -p returns libcuda.so.1 but not libcuda.so (using LDFLAGS doesn’t make any difference). The thing is a bit unorthodox and I was wondering if there’s any procedure so that these libraries end up at
/usr/lib64 or any other suggestion.
“libcuda.so” is the CUDA Driver runtime and will be installed in /usr/lib64 when you install the driver. The version under the “stub” library is simply a stub library only there to resolve the symbols on systems without a CUDA driver installed. The stub libraries should not be used at runtime.
Note that the NVHPC Compilers don’t by default link against libcuda.so. Instead, it’s dynamically loaded at runtime.
Hi Mat. Thanks. I wasn’t trying to use the stub libraries but to figure out what and what is not working, and how it differs vs installing everything via the CUDA Toolkit. From your words, it sounds like the installation of the CUDA drivers is a separate task (unlike the Toolkit that installs everything if you click on that option). I haven’t found the time today to try it but, hopefully, it will happen tomorrow.
Yes, the CUDA Driver is separate and distinct. You can download the CUDA Driver from: Official Drivers | NVIDIA