Athlon vs. PIII

Luc Vereecken Luc.Vereecken at
Tue Feb 20 09:44:50 PST 2001

>On Tue, 20 Feb 2001, Ole W. Saastad wrote:
>> I have run a few tests to check the performance of
>> Athlon vs. PIII. 
>> The Atlas (tuned BLAS libs) tests suggest very high 
>> performance for Athlon, however in real tests like
>> Gaussian 98 there is little difference. 
>> Summary of performance Athlon vs. PIII 
>> - with Gaussian98 runs.
>> pgf77 (cache size 512k)
>> Athlon:  22 minutes 26.8 seconds.
>> PIII:    17 minutes 57.3 seconds.
>> pgf77 (cache size 64k)
>> Athlon:  23 minutes  5.8 seconds.
>> PIII:    17 minutes 56.3 seconds.
>> Compiled with pgf77.
>> pgf77  -O2 -tp p6 -Munroll -Mvect=cachesize:524288 -c redund.f
>             ^^^^^^
>> Cache size 512k.
>> Included Blas code compiled. 

>At 10:59 20/02/01 -0500, Troy Baer wrote:
>Compiling with "-tp p6" is arguably the wrong thing to do on an Athlon, if
>that's what you did.  Try either compiling with "-tp athlon" on the Athlon
>or leaving out the -tp flag altogether (the compiler is supposed to figure
>out what processor it's on if it's not specified).
>Also, enabling prefetching (-Mvect=prefetch) may improve performance by
>10-15% on both the Pentium III and the Athlon.

As far as my G98 tests go, running the -tp p6 version on an athlon is about
as fast as the -tp athlon. Same for the the -Mvect=prefetch flag.  I never
managed to get Atlas compiled into G98 ? I use the pgf blas/lapack now due
to compilation problems and lack of time to look into it.

Luc Vereecken

More information about the Beowulf mailing list