improve method for disabling use of GPU
Currently a -DGMX_GPU binary must run with a GPU if one is found at run time. It would be convenient for benchmarking or perhaps running on a node where the GPU is assigned to some other load (or broken) if there was a way to run only on the CPU, without having to recompile a non-GPU binary. I suggest a GMX_DISABLE_GPU environment variable. Thoughts?
#2 Updated by Mark Abraham over 6 years ago
- Subject changed from add option to disable use of GPU to improve method for disabling use of GPU
Yes, you're right, thanks. I tried it out, and the detection of the GPU still takes place, and the .log file doesn't emit the reason why such a GPU isn't being used. That seems to me inconsistent with the things we do try to report. Updated title of this Redmine accordingly
#3 Updated by Szilárd Páll over 6 years ago
- Category set to mdrun
- Status changed from New to Feedback wanted
- Target version set to 4.6.4
In its origal formulation, this is not an issue, GPU-enabled binaries will run on machines with no GPU installed as long as the CUDA libs are available.
Additionally, with Berk's changes in 95d10d3903aed9c700009f6effece9c6f1d60517 with
-nb cpu GPU detection is not done anymore.