[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: ATLAS
From: |
David Bateman |
Subject: |
Re: ATLAS |
Date: |
Mon, 17 Sep 2007 08:56:33 +0200 |
User-agent: |
Thunderbird 1.5.0.7 (X11/20060921) |
Tatsuro MATSUOKA wrote:
> Micheal wrote;
> Theoretically, it's suboptimal. In practice, I ran a benchmark script in
> octave
> with 2 versions of ATLAS: one compiled for P4 (512KB L2 cache) and one
> compiled for Centrino (2MB L2 cache). The test was run on the Centrino
> CPU. Both ATLAS versions brought big speed-up compared to the generic
> Fortran libraries, but the gain of using a Centrino-compiled ATLAS instead
> of a P4-compiled ATLAS was not that high (IIRC it was < 10%). I didn't test
> the other possibility: running the Centrino-based ATLAS on the P4 machine.
> The binary installer only provides the P4-compiled ATLAS, which probably
> provides good results on most CPU out there (P4, AMD, Centrino...)
> ***************
> I have a question just for my reference.
> What version ATLAS did you select for ?
> stable 3.6.0
> develeopment(unstable) 3.7.xx
>
> The 3.6.0 is surely stable but I feel it is too old now.
> For 3.7.x, the term unstable make me anxious.
>
> Tatsuro
I had some very bad experiences with the earlier 3.7 releases, with
mathematically incorrect results, particularly for SVDs.. However, 3.7
is heading to a stable release so that should be fixed now. The big
issue with 3.7 is the packaging is entirely different, and so it might
take some time for it to migrate into the distributions..
D.
- ATLAS, Tatsuro MATSUOKA, 2007/09/16
- Re: ATLAS,
David Bateman <=
- Re: ATLAS, Michael Goffioul, 2007/09/17