tag:blogger.com,1999:blog-1352688022620133175.post5765803394040409..comments2014-11-12T05:08:08.435ZComments on The NAG Blog: Bitwise Reproducibility with the NAG LibrariesKatie O'Harehttp://www.blogger.com/profile/09366741271809330805noreply@blogger.comBlogger5125tag:blogger.com,1999:blog-1352688022620133175.post-59111647019182143172012-12-12T09:30:10.572Z2012-12-12T09:30:10.572ZSome versions of the NAG Toolbox use MKL and some ...Some versions of the NAG Toolbox use MKL and some ACML. None of the current versions of the Toolbox (they are all at Mark 23) were compiled with the flags that would guarantee bitwise reproducibility, so we'll need to think carefully about what we do at Mark 24.Mick Ponthttp://www.blogger.com/profile/15149585002054415615noreply@blogger.comtag:blogger.com,1999:blog-1352688022620133175.post-48865081176708199132012-12-08T18:32:06.047Z2012-12-08T18:32:06.047ZBTW, is NAG Toolbox for Matlab MKL based or not? I...BTW, is NAG Toolbox for Matlab MKL based or not? In other words, are results from Matlab using NAG reproducible? <br /><br />With respect to RobMNAG's question: I am pretty sure ALL major banks will use non-MKL version. Banks are required to report consistent numbers.<br /><br />Thanks.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-1352688022620133175.post-20124274973245125852012-11-30T13:43:39.822Z2012-11-30T13:43:39.822ZThanks for this Mick. As usual, you've made us...Thanks for this Mick. As usual, you've made us all better informed. It'll be interesting to see how many users choose the more consistent version at Mark 24.RobMNAGhttp://www.blogger.com/profile/08987706670505944198noreply@blogger.comtag:blogger.com,1999:blog-1352688022620133175.post-28603803672738329152012-11-30T09:25:56.727Z2012-11-30T09:25:56.727ZYes, of course SMP does also throw a spanner in th...Yes, of course SMP does also throw a spanner in the works, and can lead to similar problems of reproducibility (which can also sometimes have the effect of masking real errors).<br /><br />I don't think any panel notes are available yet, but I've asked Intel.<br />Mick Ponthttp://www.blogger.com/profile/15149585002054415615noreply@blogger.comtag:blogger.com,1999:blog-1352688022620133175.post-3830444975889807272012-11-29T21:48:49.667Z2012-11-29T21:48:49.667ZInteresting. But what about SMP libraries, where w...Interesting. But what about SMP libraries, where we tend to assume the associative law hols when adding up the results from different processors? Any notes of this panel - it would be nice to have something to cite?James Davenporthttp://staff.bath.ac.uk/masjhdnoreply@blogger.com