function minimization on small arrays - is it worth it ? fitting data to a function for small array

It seems that all this CUDA thing makes sense only for really large arrays of data.

Any idea of a approximate minimum number of elements for which CUDA would start to make sense ?

Say, a fit to a circle for 20 points ?

Or a Gaussian fit for 5 points ?

(for example, Core 2 CPU vs 9600GT GPU)

Sure it’s worth it, if you have to fit 2000 circles, or 10000 sets of Gaussans at once ;)


The best rule of thumb I’m aware of is the thread count. If you can launch at least several thousand threads on the device, it is usually worth it.

Thank you,
you confirmed my disappointment :(