[HEADSUP] Atlas changed libraries

Susi Lehtola jussilehtola at fedoraproject.org
Mon Sep 23 14:38:34 UTC 2013


On Mon, 23 Sep 2013 16:15:16 +0200
Frantisek Kluknavsky <fkluknav at redhat.com> wrote:
> On 09/22/2013 09:32 PM, Susi Lehtola wrote:
> >
> > I might mention that OpenBLAS (successor to GotoBLAS) is in Fedora,
> > which is often 2x faster than ATLAS. But, it's only available on
> > ix86 and x86_64. It does have runtime CPU detection, though for the
> > 20-odd CPUs supported.
> >
> 
> Could you please add more details? I can hardly imagine a situation 
> where properly tuned Atlas is below 50% of maximum possible floating 
> point performance. Packaged Atlas tuned on a completely different 
> machine can of course be slow. The theory goes that if you are into 
> serious high-performance computing, you are willing and able to
> rebuild a few packages. Plus there is a high chance that you are
> using some more exotic architecture than those power-hungry x86_64.
> 
> On the other hand pre-packaged Atlas is probably the second worst 
> alternative for desktop use cases (the worst being canonical Blas).

Yes, this is regarding pre-packaged Atlas. The cluster guys at my
university did some benchmarks of ACML, MKL, OpenBLAS, ATLAS and
reference BLAS. The first three are pretty much equal, then came
prepackaged ATLAS with 50% less performance and then last came
reference BLAS with an order of magnitude or two less speed.

But yes, you are right that ATLAS should be recompiled if you really
want performance. But then again, you can just use libraries that
already give optimal performance.
-- 
Susi Lehtola
Fedora Project Contributor
jussilehtola at fedoraproject.org


More information about the devel mailing list