topic Random numbers depend on version of mil libraries in Intel® oneAPI Math Kernel Library

Random numbers depend on version of mkl libraries

lillian — Tue, 10 Jul 2012 14:07:12 GMT

I have been using the intel mkl library for about a year, but still consider myself new to it. I have found that different versions of the library output different random numbers. Specifically I am using thevdRngGaussian function with inputVSL_RNG_METHOD_GAUSSIAN_ICDF on mac osx. The differences in the random numbers appear to be in about the ninth decimal places for mean zero std 1 normal distribution. Is there a mistake in the libraries?

Random numbers depend on version of mil libraries

Andrey_N_Intel — Tue, 10 Jul 2012 14:35:50 GMT

Hello,

Intel MKL VSL Random Number Generators do not guarantee bitwise results betweenversions of the library. Optimizations we do in thealgorithms of the generatorscan result into those fluctuations, there is no a bug in the library.

Can you please provide additional details why ~10^(-9)difference in the output of Gaussian RNG is important for your applications?

Also, you might want to have a look at Intel MKL 11 Beta, http://software.intel.com/en-us/articles/intel-mkl-11dot0/ which provides a new Conditional Bitwise Reproducibility feature.

Thanks,
Andrey

Random numbers depend on version of mil libraries

lillian — Tue, 10 Jul 2012 15:04:43 GMT

Thank you for the quick reply. I don't necessarily need that level of accuracy. I would just like to be sure that differences in outcome of my code represent changes to the code that I make rather than changes that are made by intel. But now that I know, it's no big deal. Thank you.