Intel MKL results differ depending on the underlying hardware architecture

[anonuser123]

I think this is a well known problem that when using MKL_CBWR=AUTO, the results may differ across different hardware architecture due to the floating point arithmetic that occurs.

My question is in such cases, when the results are system-dependent, how do you evaluate which one of the results is the closer to the true theoretical result?

Which MKL_CBWR settings is most appropriate for accuracy? 

[Mahan]

Hi,

Please let me know if the following link helps. It shows how one can obtain numerical reproducibility.

https://www.intel.com/content/www/us/en/docs/onemkl/developer-guide-windows/2024-1/get-started-with-conditional-num-reproducibility.html

[Mahan]

Hi,

Please let me know if there is any update.

[Mahan · Accepted solution]

As there was no response, I assume that the above link might have been able to resolve the issue.

I am closing this thread for now.

Please open a separate thread if you face any further issues.