Which cuda variant should I install?

[Dave Jennings]

Arrayfire 3.8.3 has three linux install candidates, targeting CUDA 12.0, 11.8 and 10.2.

I have a several machines that I manage with a variety of CUDA installations from 11.4 to 12.1.

Should I:

1. always use the 12.0 version (assuming arrayfire CUDA version has to be at least the same as installed CUDA version)
   
2. use 12.0 for CUDA installs 12.0+, 11.8 for my 11.8+ installs, and 10.2 otherwise (assuming installed CUDA version has to be at least the same as arrayfire CUDA version)
   
3. something else

Or to put it another way, when should I choose (say) the 11.8 version?

Thanks for any guidance.

Dave.

[Umar Arshad]

If you are downloading the binaries from our website you do not need to install CUDA. The binaries will include all libraries and runtimes necessary to run ArrayFire code and all you need to install is the driver.

The reason you may want to choose an older version is to support older hardware. The new versions of CUDA have dropped support for older compute architectures so those devices will not be supported in the new versions.

Umar

[Dave Jennings]

Thanks Umar, that's very helpful.

So are you saying that Arrayfire includes CUDA (of the given version) in the binary installer?

I have to support systems that run a variety of software, generally with CUDA already installed for some other piece of software, and occasionally requiring specific versions of CUDA as they have been thoroughly tested.

Cheers, Dave.

[John Melonakos]

Yes, ArrayFire installs the necessary CUDA components, except for the driver.

[Dave Jennings]

Just to check my understanding, we have pretty recent hardware and use recent versions of the NVIDIA driver (e.g. 525). We also install the CUDA toolkit for our own development (currently 11.4 on most machines).

From your answers and looking at the files installed by the Arrayfire installer is it correct to say that our version of CUDA toolkit is irrelevant, since Arrayfire bundles it's own libraries? So I can safely install Arrayfire 3.8.3 (CUDA 12.0 version) on a system that already has CUDA toolkit 11.4.

Cheers, Dave

[uma...@gmail.com]

Yes, if you are using our binary installers, then you don't need to install the CUDA toolkit if you are only using the GPU through the ArrayFire library. If you are building a CUDA application outside of the ArrayFire library then you will need to install the CUDA toolkit that you are going to be targeting. You should try to make this the same version as the one provided by ArrayFire to avoid using additional resources(usually insignificant) but its not a necessity.