What is the recommended card for running Gromacs?
This is probably not the right forum to ask this, as this forum seems to be designed for Cuda programming.
There are gromacs specific forums maintained by the gromacs community (not NVIDIA).
[url]http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List[/url]
Before you ask this question anywhere else, you might want to clarify the specific definition of “best” on which any recommendations should be based.
Do you have cost constraints? Electrical power constraints? Can you utilize more than one GPU? Does the size of molecular systems you need to simulate require a specific minimum amount of memory? Does your research require a particular minimum throughput (ns/day)? Etc, etc, etc.
Thank you for your responses. I was unable to find the info I need on the Gromacs forums, so I appreciate your help here.
My specifications are as follows:
-I would like to stay <5K for the cost.
-I will be using a Dell 5820 Tower workstation which supports up to 600W of graphics power.
-I would like to start with one GPU, and upgrade later, but at least two should be supported by the workstation architecture. The limitation is that Gromacs performs some computation on the CPU and some on the GPU, so CPU may be a limiting factor. I would like to benchmark this before getting a second GPU.
-Gromacs generally has modest memory use. For my largest systems (~1,000,000 atoms), the usage is <1GB/core.
-I am hoping to get speedups of 100’s of ns/day with the GPU. Of course it depends on system size…
I know that the Tesla P100 has been well-benchmarked and works very well for speeding up Gromacs. Gromacs documentation does recommend both the Tesla and Quadro series, however I have not seen any benchmarks for Quadro, so I am hesitant to get it.
I understand the Quadro series is for visualization, but can it also be used for computing?
If the performance speedup is similar, I would like to get the Quadro P6000, for both computing and visualization tasks. Also, it is cheaper than the Tesla P100 (5K compared to 6K).
Checking the messages for the month of August 2018, I don’t see where your question was asked on the Gromacs mailing list:
https://mailman-1.sys.kth.se/pipermail/gromacs.org_gmx-users/2018-August/thread.html
What response did you get on the mailing list when you asked for a hardware recommendation? Does the Gromacs documentation include any hardware recommendations?
I have not posted on the Gromacs mailing list.
Gromacs 2018 documentation recommends both Tesla and Quadro. So I assume Quadro P6000 is ok. I hesitate because I haven’t seen any benchmarking for it.
The benchmark numbers on the GROMACS website seem to be a bit out of date. The most recent numbers I could find appear to date to 2015. You could always ask on the GROMACS user mailing list about the performance other users see with the Quadro P6000 or similar hardware.
Quadros can certainly be used for computation: I am using a Quadro P2000 to run CUDA applications at this very moment. Whether the Quadro P6000 is a preferred choice for running GPU-accelerated GROMACS I cannot say, as I don’t use GROMACS.
Thanks very much for your input.
Thank You very much @njuffa for your comments. Helped a lot.