[SciPy-user] Advantages of Intel's MKL

Albert Strasheim fullung at gmail.com
Fri Dec 8 18:04:26 CST 2006


Hello all

On Fri, 08 Dec 2006, Nils Wagner wrote:

> Hi all,
> 
> What are the advantages of Intel's MKL library in connection with
> numpy/scipy ?

I'll take a stab at this question. Feel free to disagree. Mostly I'd 
compare it with ATLAS. Both MKL and ATLAS are orders of magnitude faster 
than NumPy's "LAPACK lite", so you really want to use one of these 
libraries.

Pro's of MKL (mostly vs ATLAS):

Don't have to build the library.

Runtime detection of CPU. Useful if you have a single Python+NumPy
installation shared over NFS to a bunch of heterogenous machines.
This is quite likely in cluster-type environments where you would use 
something like IPython1 -- I have a configuration exactly like this to 
deal with 15 machines in our lab.

I think MKL supports multiple cores through OpenMP by just setting 
OMP_NUM_THREADS. ATLAS also has support for pthreads though.

Probably about as fast as ATLAS.

Con's of MKL:

Costs money if the non-commercial Linux license doesn't apply to 
you and/or doesn't suit your needs (e.g. you need to run on Windows).

Those are the points that jump out for me. I'm sure there are others.

Comments?

Cheers,

Albert


More information about the SciPy-user mailing list