which gpus will nvfortran run on?
eg GeForce GTX 960
While a simple question, the answer is going to be a bit more complex, so apologies in advance!
Under the hood, nvfortran is using the back-end tools from the CUDA installation to generate the device code. Hence what target devices are supported depends on what CUDA version and CUDA driver you’re using. nvfortran can adjust which CUDA version to use, either via the “-gpu=cudaXX.YY” flag for CUDA installs that ship with the NVHPC SDK, or the environment variable “NVHPC_CUDA_HOME”, to set it to use a separate CUDA SDK install.
In our release notes, we list the minimum CUDA Driver version. This is the determination of what is the oldest device that can be supported for this particular release. For our latest 25.1 release, the minimum CUDA Driver is 450.36.06 which corresponds to CUDA 11.0.
A GeForce GTX 960 is Maxwell (cc52) based device which I believe is still supported by CUDA 11.x. Though In the CUDA 12.8 release notes, they do say:
- Architecture support for Maxwell, Pascal, and Volta is considered feature-complete and will be frozen in an upcoming release.
This is a bit different language than I’ve seen in the past when architectures were dropped. I interpret this to mean that code generation for these devices will still be available, but they wont be worked on any longer. Though I’m not sure what issues you may encounter.
Though if you do ever have issues with newer releases, you can always download an older version of the NVHPC SDK from our archives page.
Thanks that is a useful reply, and i appreciate the time taken to give a thorough response. I was looking at buying a cheaper used GPU, and was just checking if i will be able to run CFD code on it.
The consumer cards (GTX) have very limited double precision floating point performance. Hence depending on how much your code use DP, it will run but may be slow.