HI,
If I overclock my Graphics card, I know that this will enhance 3D gaming performance. What if I do overclocking and apply CUDA to solve any parallel algorithm- will I get performance improvement because of overclocking?
Thanks
Sure, but you need to be aware of the tradeoff between performance and computational accuracy. Calculation errors are not as severe in 3D applications as in most CUDA applications. You might not even see an error in a 3D application which would be fatal for most algorithms.
Sure, you will get a performance improvement. You should keep in mind that if you are compute bound, the shader clock rate sets the improvement, and if you are memory bandwidth bound, the memory clock rate is the important one. This is the same idea as Mr. Nuke described in another post where he adjusted the clock rates down to figure out whether a kernel is compute or bandwidth limited.
But yes, you should consider the stability issues before tinkering with the clocks. Most GPU overclocking (at least without replacing the cooling system) seems pretty pointless. An improvement of 10% is hardly worth the effort and possible failed calculations. Also, you should be very concerned if your GPU temperature exceeds 90C. Memory errors become more probable when the temperature is very high.
Thank you so much Seibert. Would you be kind enough to send me the link where Mr. Nuke described this.
Yes. On CPU-side, I’ve thermally destabilized my memory by not even overclocking but just choosing high settings for the FSB and RAM
(on an older mobo). Something to watch esp. on multi-GPU systems which really heat everything around them.
And yes, it’s totally not worth it, improve your algorithm to gain 50%, I’d say.