Recommendation for the GPU graphic card

I am planning to buy a GPU graphics card for DualSPHysics research. Can anyone suggest a best one in ranges of
1. <$100
2. < $500
3. < $1000
4. < $2000

Thanks

Comments

  • I was searching some of the GPUs. These following GPUs are good. I need your suggestion on which is better
    1. Titan RTX
    2. RTX 2080 ti
    3. Quadro RTX 6000
    or any other one that you would recommend.
  • I think all the necessary infos where given here

    http://dual.sphysics.org/vanilla/discussion/1446/tesla-k80-vs-gtx-1080#latest

    It will help you to compare the performances of different GPUs

    After that its just a question of price, what you can afford and how much you are willing to pay you compute capabilities

    Good Day
  • Thanks for the link. I understand that the flops is the main factor for the DualSPHysics simulations.

    Can you tell me which is better for DualSPHysics, GeForce vs Quadro?

    Quadro is made for the CAD and simulation purpose but GeForce has more GFLOPs.

    I would like to know with is better
  • No general rule ! check the Gflops and price. Then it's up to you !

    Also memory could be a factor. The more memory, the more particles you can simulate.
  • Tesla are more dedicated to execution purposes than Quadro
    Regards
  • Got it. Thanks a lot.
  • Can you suggest a GPU if my budget is $5000.
  • @C_Kesanapalli

    DualSPHysics currently does not support multi-GPU implementation. Therefore remembering that with increase of price, the performance gains usually diminish, it would in my eyes be wiser to use a powerful/suitable GPU at maybe the 1k range and then save money for later on, if multi-GPU gets implemented.

    Maybe some day someone will be able to provide a table of generic cases run on different GPU's, but I have not been able to find anything like that of value, so I suppose it does not exist yet.

    Kind regards
  • Got it. Can you suggest a GPU in your suggested price around 1k.

    Thank you
  • I would prefer to not give specific recommendations since I am not sure how CUDA is developed on different GPUs, and it would suck if I recommended a card which is not fully compatible with DualSPHysics yet. Hopefully some day a proper database and test simulations will be available to compare different GPU execution times. For now I think you should follow the more general advice given in past answers, else my personal best guess would be to buy the best possible from the 1000-series with the idea that since they have been out for some time they are probably compatible, while I know nothing about the 2000-series. Again, only based on assumption.

    Again, this is how I would look it and is only a recommendation, I am not accountable for any displeasement or anything negative my suggestion might make you experience - I hope you find something suitable.

    Kind regards
  • @TPouzol , Hi, i have the same question here. I tried to buy the below GPU card:
    https://www.gigabyte.com/Graphics-Card/GV-N2080GAMING-OC-8GC
    This card seems mainly for gaming, but i assume it should also be applicable for SPH calculations, is that correct?
  • I have only tested 1000 series cards sorry! (1060 and 1080ti)
    In theory the 2000 series should be able to run DSPH.
    I'd love to have an answer to your question from an official DSPH team member, or a user that has already tested the 2000 series.
  • As a rule of thumb: If the graphics card has CUDA cores, it can be used for SPH. All (Almost?) NVIDIA graphics cards have CUDA cores.

    A warning when looking at graphics cards: make sure the graphics card has only 1 "GPU". Some of the high end graphics cards contain 2 GPU's. Currently DualSPHysics can only use a single GPU for its calculation, so if you either have multiple graphics cards or have a graphics card with 2 GPU's then it will be wasted.

    If you are looking for the best graphics card (with almost no price limitation), look at the GFlops a card has (Processing speed of GPU * number of CUDA cores). The higher the GFlops the better it is. But once again, ensure that there is only a single "GPU" in the graphics card, otherwise the real performance will be half of what you calculated.
  • @bladomas1 How about double precisions? It seems the 20series don't have FP64 units. Do you think this may be a issue for DSPH?
  • @James I cannot answer that question. I'm using a Quatro series GPU with double precision and so far I know there are no problems.
  • @bladomas1 For DSPH, is double precision a must? I went through the guide, It says when dp is much smaller than calculation domain, double precision should be used, Do you know what kind of criteria is "much smaller"?

    I searched online, some adviced that Geforce is much cheaper than Quadro for similar performance level. It seem 1080 is out of stock recently so I am trying to find out whether 2080 can do the work for DSPH.
  • Are the 20series really without FP64 units ?
    I'm not sure...
    You can read here
    "Each SM does contain a pair of FP64-capable CUDA cores as well, yielding a double-precision rate that’s 1/32 of TU102’s FP32 performance."

    However, I wonder if the 20series cards are detected and managed by DSPH ?

    Right now my machines have 1080ti (and I'm happy with it), but for the next purchase it's going to be difficult to procure...
  • Sorry, my fault, it seems 20 series also have FP64 units.

    since the 20 series utilize similar drivers with 10 series, i assume DSPH should be able to support it, is there any DSPH members can confirm this?
  • I can confirm that DualSPHysics works fine with the 2080 ti. So it should also work with the other GPUs of the 20 series.
  • @Admin Thanks ! great news !
  • Thanks @Admin ! I noted that DSPH is complied with CUDA 9.2 while RTX in theory should be working with CUDA10.0. Do we need to recompile DSPH under CUDA 10.0 or the current version under CUDA 9.2 will be working well?
  • @James I tried compiling DualSPHysics with Cuda 10 and it would compile finely, I just had to change some line of code - I did not test the compilation though. It should work fine without compiling Cuda 10 though, and use standard distribution of DualSPHysics.

    Kind regards
  • @James The DualSPHysics compiled with CUDA 9.2 works completely fine with RTX. However, as @Asalih3d pointed out, you can compile it with CUDA 10 as well.
  • @Admin Many thanks for the kind advice! This solved my question very well.
  • Hi everybody. I'm using two NVIDIA Tesla P100... running in native double precision. It works very well !
  • BTW... NVIDIA Tesla P100 is quite expensive : 6600€ per GPU.
  • Geforce GTX family includes a list of many GPU cards that you can check with much lower prices
Sign In or Register to comment.