2.3 vs 3.0 different results....

Accelerated Computing CUDA CUDA Programming and Performance
1

Hey all,

Has any one else noticed a difference in results when switching to the newly released cuda 3.0?

I just recompiled my code in version 3.0 after having been using 2.3. The answers, in a monte carlo code simulating upwards of millions of trials, are different (both are run with the same seed, so the RNG should be the same, theoretically), but are within the variance of the answer.

Does any one have any idea why this occured?

Thanks for the help,
Adam

2

I realized that I compiled under version 2.3 and then installed version 3.0 to run the two of them. I.e., compiled under 2.3 but run with 3.0’s libcudart, if that matters (which i’m still not sure that it should).

3

And it doesn’t, found old outputs that I could compare with (where code was compiled with cuda 2.3 and run with libcudart 2.3 as well). Both the fresh version 2.3 compile and the old 2.3 show the same results.

Does any one have any idea whats going on?

Thanks,

Adam