CFD Online Discussion Forums

CFD Online Discussion Forums (http://www.cfd-online.com/Forums/)
-   Main CFD Forum (http://www.cfd-online.com/Forums/main/)
-   -   what should we do for the application of CUDA in CFD? (http://www.cfd-online.com/Forums/main/113705-what-should-we-do-application-cuda-cfd.html)

aerosjc February 25, 2013 11:13

what should we do for the application of CUDA in CFD?
 
Hi, guys! I am a newbie of CFD and an undergraduate who try to dig something in CFD. So, please forgive my ignorance in the following question. Thanks in advance! :)
I have read some papers about the application of CUDA in CFD and done some google on it. Also, I've written some codes to solve a simple heat conduction problem. I realized that CUDA will offer CFD scientists a great tool to accelerate their computation.

However, I'm not able to figure out what should we do next for the application of CUDA in CFD.

After reviewing google scholar search with keywords(cuda cfd) and time limitation of 2013(I just want to find the newest work on this subject. The link: http://scholar.google.com/scholar?as...-CN&as_sdt=0,5), it seems that scientists have already done a lot of work on how CUDA can speed up some specific problems in CFD. In my point of view, their works are just the same. They just test a specific GPU vs a specific CPU, if more work, several specific GPUs vs several specific CPUs. I see only a little reference value for other scientists.

Is there an possibility to give a universal reference for scientists to choose GPU card when they need run CFD codes on CUDA? If no for this question, why did so many scientists pay lots of attention to the speedup between GPU and CPU, and why is this trend still alive?

In addition, since Chinese scientists have not done so many works on the application of CUDA in CFD, they all think that it's imperative to do some tests on a specific problem by their own hands, although a lot of similar works have been done by foreign scientists. What's the value of these works, especially for the advancement of CFD?

FMDenaro February 25, 2013 11:18

At present, I know that many CFD problems are solved by using a mix of CPU and GPU's devices ...however, I am not an expert about GPU systems.
See:
http://www.nvidia.com/object/computa..._dynamics.html

aerosjc February 25, 2013 11:31

FMDenaro,
I have read this offical page. Like this picture:
http://otho.douban.com/view/photo/ph...1882315395.jpg
It points out that some speedup by CUDA. OK, that's a great acceleration. However, after hundreds of similar papers published, what should we do next to obtain progress for CUDA application in CFD? Should we still select a specific problem in fluid dynamics and apply CUDA into this problem?

FMDenaro February 25, 2013 11:43

my opinion is that CFD is much more than using GPU to accelerate a run ...
there are still so many open questions ...

andy_ February 25, 2013 11:57

Adapting CFD codes to run on particular types of processors has been going on for decades. It is more development than research. As time progresses different processors become the most cost effective platforms for CFD and so codes get ported to them. The effort involved in porting is normally pretty small.

aerosjc February 25, 2013 11:58

yes, I also know that in CFD there are so many problems to be tackled, not just only GPU acceleration. However, I just want to figure out what should do after so many tests done, which are for what kind of extent the speedup of CUDA can reach. Could you provide some advices, please?
I want to do some work on the capture of shock wave by using a high accuracy scheme and executing on CUDA. Should I just test it?

aerosjc February 25, 2013 12:01

thanks, andy! I understand, it's not a research, it just is a kind of development.


All times are GMT -4. The time now is 18:44.