I have found some Linux software that uses CUDA to do calculations on the GPUs. The software is available as binaries and as source code.
I notice that the CUDA Toolkit is very fussy about which Ubuntu version and which kernel I would need, the first question I have is: Do I need the toolkit to run software that uses CUDA if I have the binaries?
If I DO need the toolkit software, is there an .iso installation file I can get that has everything setup already? Because the current Ubuntu server version 18.04.3 i think has kernel 5.0 but the toolkit says it needs kernel 4.15.0, so what am I meant to do, downgrade the kernel somehow? It’s all a pain. I need to run Ubuntu 18.04, that is my only requirement, the GPUs are 2080ti cards.
Generally, you do not or should not need a full CUDA toolkit install, if the only thing you need to do is run compiled CUDA binaries that have been properly built and packaged.
In that case, the only thing you should need is the GPU driver.
This doesn’t really mean that this is the best option for you. For example, it may be that if the “binaries” that are available to you were built with e.g. RHEL instead of Ubuntu, you may not be able to run them on your Ubuntu setup. The general reasons for this have nothing to do with CUDA. However if the binaries have been properly built and packaged for your particular OS/distro, then the addition of CUDA should only require that you have proper GPU drivers installed.
It’s possible for application developers to make build/packaging mistakes, of course, meaning that libraries that can be statically linked are dynamically linked, or that needed dynamically linked libraries are not included with the binary, but these topics don’t negate anything I’ve said above. In such situations, you might need to provide those libaries, and if they were CUDA libraries, the easiest way might be to install the CUDA toolkit.