I have been working with CST for the significant part of the past year and half, and it can sometimes be frustrating to wait for the solver to finish (with all that excruciating pain of listening to laptop cooling fans turning into jet turbines :) ). This made me think about using the GPU for computing as well. (to be honest, i wanted to see what is all the fuss about in the CUDA and GPU computing community). It was indeed very exciting to see how Time Domain Solver computation time reduced significantly.
But how I did it? Unfortunately, in the case of CST it is really hard to find proper documentation (Comparing to MATLAB or ANSYS documentations or online communities). This is more accurate when you want to setup a GPU hardware acceleration if you have an unverified GPU unit in your PC (almost all of the NVIDIA GPUs that are not intended for workstations are unverified by CST (Mine is an Geforce RTX 3060 Mobile).
So my point here is, please share your experience with the CST GPU computing here in this discussion for newbies like me. Mine is as following:
I am using a CST Microwave Studio 2021 release on a Core i7 10850H + 64GB Ram with RTX 3060 Mobile.
The first step would be to check your GPU model to see if it has any CUDA cores that you can use for the GPU computing.
Then you should check if it supports at least CUDA 9.2 code. Download CUDA toolkit from NVIDIA website. Install it.
Once you are done with that, you can select acceleration in solver options and activate "Hardware Acceleration" with setting the number of GPUs you have on your PC.
Note! Remember that CST does not support GPU computing on every solver ( :( unfortunately for me, frequency domain and eigenmode solvers dont support Hardware acceleration). But at least time domain solver is supported.
You can check the supported solvers in the GPU computing Guide. I will attach the guide for CST 2021 release here.
Additionally, I must note that if you want to see the GPU performance in task manager during a simulation, you should change the graph in the GPU section to CUDA.
I have quadro p620 with 512 CUDA cores. Did all according CST guide and download toolkit from NVIDIA too. It is convenient to use frequency solver now because it is faster for ma tasks.
Time domain test = scattering plane wave on dielectric sphere
I5 10300H, 16Gb RAM, Nvidia Quadro p620
I checked different accuracy and mesh, GPU gave X3 boost ( it is maximum which i get)
Yeah you are right. Time domain solver has its advantages but you dont have to use it all the time. It requires higher amount of mesh and technically GPU computing will help you if you dont have any other choice.
I, too, use frequency and Eigenmode solvers more frequently, but unfortunately for Voxel data you need to use the transient solver most of the time (if not always).
Please correct me if I am wrong.
By the way 3X is not bad though. :D
GPU can be essetial for large scale problems like transient simulations of very high-Q resontaors or self-consistent PIC problems. For example, there is no way that you can solve beam dynamics in particle accelerator without a GPU. For T-solver, I also observe 3-5x time improvement with Volta.
Sergey V. Kutsaev Now that you mentioned. Is it better to simulate the high resonant structures in Time domain or Frequency Domain? I had the impression that frequency domain FIT is better for resonant structures.
Frequency domain is very good, when you are only interested in stationary processes. When transinet effects are required, you have to go with time-domain.This is particulary improtant for accurate simulations of EM wave and beam interaction.
- Badges
- Science topic
- Similar topics
- Computer Communications (Networks)
- Simulators