[AeroTH]

Greetings,


I'm trying to optimize a machine to run RapidCFD, and would like to know more opinions about what GPU to buy for it. I think I've read through most of the posts in this forum that mention RapidCFD at this point, and there is quite the mixture of Teslas, Titans, and Quadros being used on that code, but I haven't seen any discussions on the best setup for it.


RapidCFD and their GitHub mention the Tesla K20. I'm already basically aware of the differences between "workstation" cards Teslas and Quadros vs. "gaming" cards like Titans and GeForces (other than just the raw TFLOPS). I'm more wondering about other practical considerations. Will a bigger card always provide better results? How cheap can we make the CPU since RapidCFD runs on GPU? Multiple smaller GPUs vs. one bigger GPU? What would be a good "reference machine" (incl. motherboard, ram, etc.) for this?



Just FYI, I am a professional engineer looking to apply CFD to drone design. Maybe there's nothing more to say than the stations people have mentioned in this thread, but if there is more to say, please post it here. Thank you.

[flotus1]

Sifting through the information that is readily available, I have to say that I am not particularly impressed. Or surprised. What would surprise me is if the devs could answer the detailed questions you have.
This piece of software seems to suffer from the exact same issues that are plaguing GPU acceleration on closed source software, with the headaches of open source software slapped on top. Using CUDA is just the icing on the cake.

There is only one reason why GPU acceleration can be viable for closed source CFD codes: the license cost structure has been adjusted to make it more attractive than just using more CPU cores.
With license costs not being a factor to consider, I see no compelling reason to invest in GPU acceleration, instead of higher CPU performance.

[hsnyder]

Originally Posted by flotus1

This piece of software seems to suffer from the exact same issues that are plaguing GPU acceleration on closed source software, with the headaches of open source software slapped on top.

This made me chuckle (though I'm sure you were dead serious). But what do you mean, exactly, regarding the issues plaguing GPU acceleration in closed source software? I've only used open source so maybe this is obvious...

[flotus1]

One of the biggest problems is again made evident by this thread: lack of information.
You can't really know in advance if your specific simulations will benefit from GPU acceleration. A solver being able to run on GPUs does not necessarily mean that it will actually run faster than on CPUs. For this software in particular, they put one extremely ambiguous benchmark for one solver on their page. It is left to the user to figure out the rest.
Then there are general hardware restrictions. Particularly when it comes to GPU memory, limiting the size of simulations that can be run.
And it might just be me, but I always feel like marketing is way ahead of reality on the topic of GPU acceleration. People with limited hardware budgets regularly stop here, either to ask which GPU to buy, or why their "expensive" RTX Quadro card does not deliver the level of performance they expected after reading some marketing slides.