Multi-GPU run

First, thanks for the very cool software.

Designing a sailboat hull and would like to visualize resistance, trim etc.

Question: I see that I can select GPU with -gpu[:id] but how to run across 3 gpu's for a single simulation?

Comments

  • After not finding anything in the documentation and examples, I searched through the code and found some precompiler macros... Guess I'll have to set up vs2015 and try to build this in multigpu mode

  • The code is not ready for multigpu yet.

    Regards

  • Buenos dias Alex

    I saw a lot of literature out there talking about multi-gpu test runs with this software... guess those were custom deployments? if so, how can I get hold of that code, even if not been released yet? I'll adapt it to my servers and make any advances available to the rest.

    Any reason FPGA's weren't used over GPU's? seems like that architecture may be better suited to scaling this up to real-world use.

    Based on the resolution achievable on one GPU card, this doesn't really seem feasible for real-world use until parallel GPU/FPGA or better yet grid compute capable?

  • best resolution I can get to is 3mm on a 33 scale model with a 1080 ti 11gb memory

    not nearly enough resolution for the finer details e.g. foils, need probably 10x that.

    appears GPU memory is the major limitation here, so FPGA may be worse, they do have 40 and 80 gb A100 tesla cards now

    will try one of those on the cloud but the single GPU limit is tough

    is there any have a finer dot mesh in regions? like with CFD where you mesh better resolution in areas. Seem that would help greatly.

  • A100 is a good bet.

    They are working on multi-resolution but don't expect it to come this year as far as I know. Multi-resolution is possible in SPH (but not in DualSPHysics), but adaptive localized refinement is much more tricky in a Lagrangian approach.

    I think you should tone down your language a bit in regards to "real world use". A lot of people have found real benefits in using SPH and specifically DualSPHysics for different cases where conventional CFD methods would lack. For your specific case SPH might not be there yet, but hopefully through your use of the software + increasing number of users from different backgrounds we can help the DualSPHysics Team develop an even better tool, which is to be honest already great considering it is mostly OpenSource and developed by Universities around the world, even with quality check procedures too. A rare thing to find.

    Kind regards

  • >I think you should tone down your language a bit in regards to "real-world use". A lot of people have found real benefits in using SPH and >specifically DualSPHysics for different cases where conventional CFD methods would lack.

    Point taken. I was thinking in terms of my use case. Apologies for the frustration in my tone, a lot of hours invested to hit a dead end.

  • No worries! I had also have had this feelings at times and at the end of the day it also shows passion which is important for this kind of work, especially when it is your own use cases, you care deeply about.

    Regarding your point of why not using FPGA, I believe it was because that GPU architecture was more readily available/common, and would allow everyone to use DualSPHysics, without professional hardware.

    Kind regards

Sign In or Register to comment.