CUDA 10 vs. CUDA 10.1

The functionality of OpenMM will (eventually) include everything that one would need to run modern molecular simulation.
POST REPLY
User avatar
Thomas Simon
Posts: 3
Joined: Wed Jul 18, 2018 3:10 am

CUDA 10 vs. CUDA 10.1

Post by Thomas Simon » Fri Jun 14, 2019 4:45 am

Hello,

for installation of OpenMM: will this nowadays work with the current CUDA 10.1 (using presumably omnia/label/cuda101), or is still CUDA 10 the way to go? And as followup: will the OpenMM executable name have to reflect the actual CUDA version (so that openmm_7.3.1_py_cuda92_rc_2 is wrong and should contain cuda10 or cuda101 in the name instead)?

Many thanks

Tom S.

User avatar
Peter Eastman
Posts: 1901
Joined: Thu Aug 09, 2007 1:25 pm

Re: CUDA 10 vs. CUDA 10.1

Post by Peter Eastman » Fri Jun 14, 2019 11:10 am

Currently we only have conda builds for versions up to 10.0.
will the OpenMM executable name have to reflect the actual CUDA version (so that openmm_7.3.1_py_cuda92_rc_2 is wrong and should contain cuda10 or cuda101 in the name instead)?
What executable are you talking about? You use OpenMM by importing it into Python code, not by running a fixed executable. (There's the openmm-setup application, of course, but that's in a separate package and its name doesn't depend on which CUDA version you're using.)

User avatar
Thomas Simon
Posts: 3
Joined: Wed Jul 18, 2018 3:10 am

Re: CUDA 10 vs. CUDA 10.1

Post by Thomas Simon » Mon Jun 17, 2019 3:19 am

Hi, thanks for the quick response. Any confusion about correct terminology is my bad. I meant what is probably the display of package names in the course of the conda-based installation. I had prepared a CUDA 10.1 installion and experimentally tried using the label cuda101 for openmm-installation. There was one package, I think fft-related, which contained cuda101 in its name (sorry, no exact notes kept). The openmm package was announced as openmm_7.3.1_py_cuda92_rc_2, which appears to be the default-installed 9.2-related version and in turn made me suspicious about correctness of the label cuda101 + therefore my question here. Probably best to trash everything and set up the machine with CUDA 10!

User avatar
Peter Eastman
Posts: 1901
Joined: Thu Aug 09, 2007 1:25 pm

Re: CUDA 10 vs. CUDA 10.1

Post by Peter Eastman » Mon Jun 17, 2019 11:10 am

If conda is listing an OpenMM package with "cuda92" in the name, that means you have that version of OpenMM installed. If you're trying to create your own build, you should first uninstall that one. Having two builds of OpenMM installed at once in the same conda environment will cause problems.

User avatar
Thomas Simon
Posts: 3
Joined: Wed Jul 18, 2018 3:10 am

Re: CUDA 10 vs. CUDA 10.1

Post by Thomas Simon » Sat Jun 29, 2019 7:33 am

Hi,

everything up and running now with CUDA 10.

For what it may be worth: enclosed are results from the benchmarl.py script run on an oldish workstation, but equipped with a Nvidia GeForce RTX 2080 Ti card.

A minor nuisance, and maybe you want to look at it eventually with respect to the distribution mechanism, was that the OpenMM installation did not install all files of the examples/ folder + at least apoa1.pdb was missing (then located from a OpenMM 7.3.1 archive from Zenodo).

Cheers Thomas S.
Attachments
benchmark_results_CPU.txt
(748 Bytes) Downloaded 1 time
benchmark_results_CUDA.txt
(774 Bytes) Downloaded 5 times
benchmark_results_OpenCL.txt
(1.32 KiB) Downloaded 1 time
benchmark_results_Reference.txt
(755 Bytes) Downloaded 1 time
hardware_details_for_benchmark.txt
(443 Bytes) Downloaded 2 times

User avatar
Peter Eastman
Posts: 1901
Joined: Thu Aug 09, 2007 1:25 pm

Re: CUDA 10 vs. CUDA 10.1

Post by Peter Eastman » Mon Jul 01, 2019 8:54 am

at least apoa1.pdb was missing
Thanks for pointing that out! I just fixed it.

POST REPLY