Electronics

Electronics

HFSS GPU acceleration for my hardware setup

    • grybka
      Subscriber

      I am looking to speed up my HFSS calculations. Currently, my desktop is Windows 10, Intel i7 3.4 GHz with onboard graphics, and 48 GB of RAM. I am thinking of purchasing the NVIDIA K40c or NVIDIA K80 to have GPU acceleration, but I am very scared of wasting >$1k for not much benefit.

      1. Will my hardware setup benefit from an NVIDIA Tesla card? Is it possible that my CPU won't be able to effectively implement GPU acceleration?
      2. Roughly, what performance increase would I achieve with GPU acceleration from K40c and K80?


      Thank you.

    • slouie
      Ansys Employee

      Hi grybka,


      HFSS will tap the GPU for a few specific scenarios:


      -If you're using the Direct Solver and your mesh count reaches greater than 2 million --- we've seen cases where speedup is ~2x when compared to running on 8cores alone.


      -Some applications that use the SBR+ solver --- very significant speedup!! 


      -If you're running HFSS Transient hybrid solver --- very significant speedup!! 


      If none of those cases apply to you, then GPU is not necessary. 


      Hope that helps your decision making process!


       

    • grybka
      Subscriber

      This is very helpful slouie!!

    • kailash
      Subscriber
      Will it help for CFD calculations as well?n
    • AndyJP
      Subscriber
      This will not be very helpful because the written above is simply not very true (confusing incomplete truth, to be fair).nhttps://forum.ansys.com/discussion/comment/109374#Comment_109374nhttps://forum.ansys.com/discussion/comment/105255#Comment_105255nThe short answer:nGPU acceleration is always activated for isotropic dielectric media in Transient and Driven Modal. Probably SBR is in the list, I don't use it.nI did not find any threshold mesh elements count for that.nIt does not work in other solvers. It does not work when any kind of anisotropy is found. I do not remember, does it work with shell boundaries; it is still a question.nGPU acceleration gives crazy benefits when solving longer than 5 minutes per pass. For short simulations there is no difference because of long and unreliable license negotiation with license server, and long pre/post-processing.nGPU acceleration in unfairly blocked for GTX boards, but work on absolutely equivalent Quadro/Tesla boards.n
    • AndyJP
      Subscriber
      P.S. Kepler architecture (Quadro K) will soon be phased out, because nVidia has limited life of upcompatibility in CUDA sdk. Newer versions of SDK will not provide newer capabilities for these boards, so the software will stop computing with them.n
Viewing 5 reply threads
  • You must be logged in to reply to this topic.