On Thursday 19 September 2002 16:04, you wrote:
> Xiaogang Wang wrote:
> >What is the best blas for an intel pentium III 1.26G
> >cluster (each node has two cpu's)?
> >Is 'Intel Mathematical Kernel Library" the answer?
According to Intel's documentation, the BLAS level 3 and the FFT
routines can use threading to make use of the two CPUs, even if the
application is not multithreaded.
> There is also ATLAS, which will automatically tune itself to the
> processor: http://math-atlas.sourceforge.net/
> It will take longer to install, but I believe it is the recommended
> course (it also has C bindings and is portable on almost any system).
ATLAS will mostly tune its blocking factors to adapt itself to the
machine's cache sizes. If the underlying compiler cannot produce good
floating-point code or prefetching hints, then an assembler kernel will
still be faster.
The Intel MKL also has C bindings.
-erik
--
Erik Schnetter <[log in to unmask]>
Web: http://www.tat.physik.uni-tuebingen.de/~schnette/
My email is as private as my paper mail. I therefore support encrypting
and signing email messages. Get my PGP key from www.keyserver.net.
|