I think the title explains my main concern: CUDA 9 final release will probably be out very close in time to Fedora 25’s End-Of-Life. While there are other options (mainly CentOS/Ubuntu LTS/OpenSUSE), my main system is and will remain Fedora on multiple machines and CUDA 9 is going to officially support Fedora 25. Now, I have no problem compiling GCC 6.4.0 myself from source. Will this suffice in order to use CUDA 9 if I upgrade to Fedora 26 or even Fedora 27 (which will be released about at the same time)? In case I decide to keep my current F25 installation, will I be able to install new packages (possibly required for working with CUDA) after Fedora’s EOL? Have you got any experience in this regard? It bugs me a bit to see that an unmaintained system is going to be officially supported for a long while.
Generally speaking, that is what I have done. I recently set up CUDA 8 on laptop on fc25, and just added gnu 5.4 toolchain, and have had no issues. I have plenty of other systems running fc20 with CUDA 8 and 9 RC also, with no issues.
It’s not officially a supported setup, so your point is still valid, but if you are willing to consider workarounds, it shouldn’t be overly difficult in my experience. However I’m not suggesting this addresses your main point, which has been a thorn for a long time with CUDA. It is related to the tighter integration of the CUDA toolchain with the host compiler than what you might experience with even similar platforms like OpenCL, and the tight integration brings a QA cost and an associated time lag.
I think I’ll do something similar and hopefully have CUDA 9 work on more recent versions of Fedora. I did use workarounds in the past to get CUDA working on unsupported systems and I did have luck in multiple occasions. The problem is that since I’m working on a paper which will possibly be sent for review by the end of the year, I’m planning to do some intensive testing as soon as CUDA 9 is released, and I also want to do that on an officially supported platform. I’m aware that whenever a certain level of reliability is demanded, the obvious choice should be a long-term supported system such as CentOS or others. However, much of the software I’m developing requires recent versions of quite a few libraries and tools, which are easily available in Fedora 25. Being this distribution officially supported by CUDA 9, I’d like to use the same platform for my whole research, also because my CUDA application needs to talk to the rest of the software I developed and built on F25. This might not be possible on CentOS or Ubuntu LTS… not only because I’m not really used to them (this shouldn’t be such an issue), but also because they come with rather obsolete packages as of now. Plus, I have Fedora already installed on multiple machines. Thanks for your answer.