Floating point computation for games.


I know this might be seen slighly off topic, but please indulge me.

I am very keen flight simulator enthusiast, and as such i’m very keen to build a new system for flight simulator X…Please indulge me a little further before i hear cries of wrong forum.

Now the crux of my question. From what i’ve read, Flight simulator is very CPU intensive and tends to use CPU power more than GPU.

So in a general scenario where a game / application is CPU intensive, would a tesla board provide any computational advantages. I guess i’m wondering if it will almost work like the good old days of adding a maths Co-processor.

I know it should help with dedicated cad systems, but what about standard applications likes games or even the likes of large spreadsheets in excel with complex forumla’s ? Would there be any benefit ?


Plugging in a Tesla board will not magically make existing apps faster, they need to be written specifically with CUDA in mind.

Now with a flight simulator, one could do some intersting things. With the power that the GPU offers for computation, realistic aerodynamics could potentially be simulated in real time.

CUDA is best for problems where you need to operate on large numbers of items in parallel. I mean, really large arrays, like 10k and up. Less than that, and the overhead of communicating with the graphics card often swamps any speed benefit, unless you have a very large amount of arithmetic to do per element.

GPUs are unlikely to be useful for an Excel spreadsheet. Problems with large enough datasets should be tackled with something more specific than Excel.

I could imagine CUDA has some potential in a flight simulator, since that sounds like it might involve fluid dynamics, finite element analysis, or something like that. Those sorts of problems are a better fit for a graphics card.

And, of course, there is AMD’s push (“Fusion”, I think they are calling the chip) to put GPU-like elements onto the same silicon as the CPU. This will lower the communication barrier quite a bit, and make GPGPU applicable to more middle range problems.