I made have the very same experience on GNU/Linux on my atom N330 the
OpenBLAS shipped by my distro outperforms even the ATLAS I specifically
compiled for that machine by ~ 30%, I switched to use OpenBLAS on all my
Help-octave mailing list
[hidden email] https://mailman.cae.wisc.edu/listinfo/help-octave
OpenBLAS is the continuation of the old GotoBLAS library. No new code
has added in order to speed the computations has added, so I think
that the results should be the same as for the GotoBLAS.
El día 3 de abril de 2012 13:07, Martin Helm <[hidden email]> escribió:
> Am 03.04.2012 11:45, schrieb Michael Goffioul:
>> On Tue, Apr 3, 2012 at 10:40 AM, nitnit <[hidden email]> wrote:
>>> OpenBLAS (ver 0.1.0 alpha 2.5) is provided as part of recent mingw
>>> octave-3.6.1 build on sourceforge. mingw compiled OpenBLAS stable ver. 0.1.0
>>> can be downloaded from
>>> https://skydrive.live.com/redir.aspx?cid=03f7acf7a4a62765&resid=3F7ACF7A4A62765!132&parid=3F7ACF7A4A62765!122&authkey=!ANxnqi-7QHzQg74 >>>
>>> I have tried it on several machines and it outperforms the ATLAS 3.8.4
>>> version. There are some minor diffs between the ATLAS and OpenBLAS
>>> computation results.
>> I confirm this: OpenBLAS outperforms ATLAS 3.8.4. The difference is
>> small on an older P4 (~10%), but large on an Atom N570 (2x).
> I made have the very same experience on GNU/Linux on my atom N330 the
> OpenBLAS shipped by my distro outperforms even the ATLAS I specifically
> compiled for that machine by ~ 30%, I switched to use OpenBLAS on all my
> Help-octave mailing list
> [hidden email] > https://mailman.cae.wisc.edu/listinfo/help-octave
If you use Intel Sandy Bridge (AVX), you should use ATLAS 3.9.xx
version which supports AVX instructions.
We are going to optimize OpenBLAS on Sandy Bridge in this year. In
our preliminary experiment, we obtained the MKL comparable DGEMM
(double precision matrix mulitipication) performance on Intel Core