I used the multi-GPU running command with 2,4,6 GPUs.
2 GPUs can run but the results are unreasonable.
4 and 5 GPUs cannot run succesfully.
Single GPU performs well.
Is there any requirement for compilation for multi-GPUs?
For example, the GCC, MPI, CUDA versions?
Best regards,
Joshua
Hello @joshua, thanks for your interest in GPUSPH.
Results in the multi-GPU case should be the same as single-GPU, so if you’re seeing significant difference you’ve either hit a bug or there’s something wrong with your setup. MPI is not necessary for multi-GPU on the same device, and as long as the CUDA version is supported, and the GCC version is supported by your CUDA version, it work fine.
There is however a known bug in Thrust shipped with CUDA 11 that causes issues with the particle sort. The issue is much more apparent in multi-GPU, but can also affect single-GPU runs. This should be fixed in CUDA 11.3, and a workaround for the issue has also been pushed to the next branch of GPUSPH. Can you try either and see if it fixes the issue for you?