[Beowulf] Intel 64bit (emt) Fortran code and AMD Opteron

Mikhail Kuzminsky kus at free.net
Fri Oct 29 08:55:26 PDT 2004

In message from Greg Lindahl <lindahl at pathscale.com> (Thu, 28 Oct 2004 
13:28:35 -0700):
>On Thu, Oct 28, 2004 at 01:39:52PM -0600, Craig Tierney wrote:
>> However, for most applications the vectorization
>> is going to give you the big win.
>People think that, but did you know that SIMD vectorization doesn't
>help any of the codes in SPECfp?
It's interesting ! 
Opteron SPECfp2000 results obtained w/help of PGI 5.1-3 includes
-fastsse copmiler option. SPECfp2000 results (for Opteron) based
on old ifc 7.0 compiler include options like -xW which allow
to create SIMD instructions. Etc. There is 2 possibilities

a) These compilers didn't generate SSE2-containing codes for any
program from SPECfp2000 - what looks strange for me 
b) In the case we'll re-translate the source of SPECfp2000
w/suppression of SSE commands generation, performance results will be
the same. Do I understand you correctly, that you say about case b) ?

BTW, if I remember correctly, ATLAS dgemm codes for Opteron are better 
if they are using SIMD fp operations - but of course, it's "out of
SPECfp2000 codes"

> Remember that the Opteron can use
>both fp pipes with scalar code. This is very different from the

Yes, but 32-bit ifc compilers (which don't know about Opteron 
microarchitecture) gave better results than pgi compilers oriented
to "right" microarchitecture. Of course, I don't say about yours 
PathScale compilers which usually are the best (in the perofrmance
of codes generated) but too expensive :-( . 

Mikhail Kuzminsky
Zelinsky Institute of Organic Chemistry
>I'd say this myth is the #1 myth in the HPC industry right

More information about the Beowulf mailing list