I tried this, and found that on opteron libmkl_p4 dgemm is much slower (2.2 GFLOP) than with Atlas (3.3 GFLOP). I agree that on P4/Xeon mkl61 is the best choice, but Atlas Linux_P4SSE2 is only slightly slower. I have never seen a factor of 2. Joachim Werner On Sa, 20 M?r 2004, Sigismondo Boschi wrote:
>Hi, > >have you tryed it with the MKL from intel? On P4/xeon it performs about >twice the atlas on many benchmarks. Still to try on Itanium. > >Regards, > Sigismondo Boschi > > > >H. -J. Werner wrote: >>Performance numbers for opteron 2.0 GHZ ifc/i4 serial and parallel (2 >>processors >>only) can be found under www.molpro.net (click benchmarks). On the >>average, the Operon 2.0 GHZ ifc/i4 is the fastest of all currently >>tested machines for single-processor, 2-processor mpp, and 2 processor >>throughput (2 identical jobs at the same time). The 2.0 GHZ version >>used atlas blas built for this machine. But using the prebuilt >>Linux_ATHLON Atlas blas is only slightly slower. >>ACML performes much less well than Atlas since it is slow for small >>matrices. For the i8 version, compiled with pgf90, the >>prebuilt Linux_HAMMER64SSE2_2 performes well, but the overall >>performance is worse than for ifc/i4 due to less good optimization of >>pgf90. The Portland compiler people told us that this will >>be improved in the near future. >>Joachim Werner >> >>P.S. There is one unresolved problem with the atlas blas/lapack (i4 >>version): >>the dsyev diagonalization routine crashes on opteron (only for large >>matrices). This problem can be avoided by adding "olddiag2" to the >>FTCFLAGS in CONFIG: >>FTCFLAGS="blas1 blas2 blas3 lapack olddiag2" >>After doing this, delete src/util.o and remake the program. >> >> >>----- Forwarded message from Alex Brown <[EMAIL PROTECTED]> ----- >> >>From: Alex Brown <[EMAIL PROTECTED]> >>To: Matt Hodges <[EMAIL PROTECTED]> >>Cc: [EMAIL PROTECTED] >>Subject: Re: Molpro on opteron >>Date: 10 Sep 2003 08:32:52 -0600 >>Message-Id: <[EMAIL PROTECTED]> >>X-Mailer: Ximian Evolution 1.2.2 (1.2.2-5) >>Organization: >> >>Dear All (and Matt in particular), >> >> The news/update on the port to the Opteron is very useful. Thanks. >> >>You mention that the ATLAS libraries perform significantly better than >>the ACML ones, but can you give an indication of the overall performance >>of Molpro on the Opteron. Of course, one would have the advantage of >>both 32-bit and 64-bit computing, but it would be nice to have an idea >>of how the Opteron is benchmarking against other systems (even at this >>early stage before the port is complete). >> >>I have been thinking about the Opteron and news that performance (even >>now) is good, may encourage me to leap in anticipation of even better >>performance in the future. Any info would be useful. >> >>Cheers, >>Alex >> >>On Wed, 2003-09-10 at 03:37, Matt Hodges wrote: >> >>>>>>>>M D Wheeler writes: >>> >>>> Does anyone know if Molpro has been compiled on an opteron system?? >>>> If not, does anyone know if it is possible?? >>> >>>The Opteron port is work in progress. There is a patch for 2002.6 that >>>contains changes to the configure process to support such systems, and >>>licence objects for i4 and i8. At the moment, the i4 version works >>>well (all testjobs succeed), and we have found that performance with >>>the ATLAS libraries is significantly better than with the ACML ones >>>(<URL:http://developer.amd.com>). >>> >>>With all patches applied, it should be possible to configure the i4 >>>version of Molpro 2002.6 by first making sure that the 32-bit compiler >>>is in your path (i.e., 'which pgf90' should return something like >>>/usr/pgi/linux86/5.0/bin/pgf90), then using './configure -pgf90'. >>> >>>Note that there are some unresolved problems with the i8 version, and >>>all the above is for serial execution; we don't yet have a version >>>working with Global Arrays on this platform. > > > >-- >Sigismondo Boschi, Ph.D. tel: +39 051 6171559 >CINECA (High Performance Systems) fax: +39 051 6137273 - 6132198 >via Magnanelli, 6/3 http://instm.cineca.it >40033 Casalecchio di Reno (BO)-ITALY http://www.cineca.it -- Prof. Hans-Joachim Werner Institute for Theoretical Chemistry University of Stuttgart Pfaffenwaldring 55 D-70569 Stuttgart, Germany Tel.: (0049) 711 / 685 4400 Fax.: (0049) 711 / 685 4442 e-mail: [EMAIL PROTECTED] ----- End forwarded message ----- -- Prof. Hans-Joachim Werner Institute for Theoretical Chemistry University of Stuttgart Pfaffenwaldring 55 D-70569 Stuttgart, Germany Tel.: (0049) 711 / 685 4400 Fax.: (0049) 711 / 685 4442 e-mail: [EMAIL PROTECTED]
