Announcement

Collapse
No announcement yet.

More AMD FX-4100, FX-8150 Bulldozer Linux Details

Collapse
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • phoronix
    started a topic More AMD FX-4100, FX-8150 Bulldozer Linux Details

    More AMD FX-4100, FX-8150 Bulldozer Linux Details

    Phoronix: More AMD FX-4100, FX-8150 Bulldozer Linux Details

    On Friday some benchmarks of the AMD FX-8150 Bulldozer on Linux were shared, thanks to an early adopter running some benchmarks under Gentoo and uploading them to OpenBenchmarking.org. But there's more to come...

    http://www.phoronix.com/vr.php?view=MTAwMTg

  • AgY!
    replied
    Originally posted by sunweb View Post
    According to this http://openbenchmarking.org/result/1...AR-FXCOMPILE20 optimized with Open64 compiller Bulldozer is dream for rendering and multithreaded aps.
    Stil no real programs like Blender with internal/Cycles render, LuxRender but its a start.
    +
    http://openbenchmarking.org/result/1...AR-GCCLLVMCL67
    +
    http://openbenchmarking.org/result/1...AR-FXCOMPILE51

    Just saw those tests. Awesome work Michael.
    However please take a closer look at the compiler flags.

    http://developer.amd.com/assets/x86Open64QuickRef.pdf

    Open64 (amd and the other one) support the following parameters:

    -mp enable the multiprocessing directives (basically like -fopenmp)
    -mso Optimize for multicore scalability
    -lno Loop Nest Optimizations
    -apo Enables automatic parallelization

    The -mp flag is a important one because it seems gcc activates this on -O2/-march or maybe it is somehow detected.
    Also there is -ipa .. which i cant get to work because of various system specific issues...

    Apo and ipa are things amd seems to be proud of while -mso and -lno are actually "cool" optimizations.
    I think its important to include those in the test so nobody can say "but hey you didnt test with .."

    Leave a comment:


  • sunweb
    replied
    According to this http://openbenchmarking.org/result/1...AR-FXCOMPILE20 optimized with Open64 compiller Bulldozer is dream for rendering and multithreaded aps.
    Stil no real programs like Blender with internal/Cycles render, LuxRender but its a start.

    Leave a comment:


  • skies
    replied
    Originally posted by AgY! View Post
    Thats john/stream:

    http://openbenchmarking.org/result/1...AGY-COMPILER21

    It doesnt really reflect what i saw/experienced earlier. ;P Maybe i only saw what i wanted to see.

    Anyway, a compiler cant fix bulldozer.
    Thanks for the benchmark, interesting with the Stream tests.

    I do think it is a compiler issue, I believe if proper optimizations are done the Zambezi is in par with the Sandy Bridge in speed (and SB is like +6 months old by now).

    Leave a comment:


  • PsynoKhi0
    replied
    Originally posted by hermiti View Post
    My motherboard supports up to 2133MHZ overclocked. I will try turning off the cores as discussed in that thread once I finish tweaking my system.
    I stand corrected: http://whatswithjeff.com/amd-black-am3b-socket/

    Leave a comment:


  • AgY!
    replied
    Originally posted by AgY! View Post
    I did some benches (amds open64 vs gcc 4.6.1) and in Multithreaded programs you can see a large improvement. But Singlethreaded programs run slower.
    Thats john/stream:

    http://openbenchmarking.org/result/1...AGY-COMPILER21

    It doesnt really reflect what i saw/experienced earlier. ;P Maybe i only saw what i wanted to see.

    Anyway, a compiler cant fix bulldozer.

    Leave a comment:


  • AgY!
    replied
    Originally posted by skies View Post
    I would love to see a run of the benchmark suite using AMD reference compiler Open64 (a forked version of GCC with features from SGI's MipsPro).

    It's free and open source just like GCC. Available here http://developer.amd.com/tools/open6...s/default.aspx

    As far as I know the stock GCC can not do any optimizions for AMD's new AVX/FMA3/FMA4/XOP instruction sets, you need Open64 for that.
    I did some benches (amds open64 vs gcc 4.6.1) and in Multithreaded programs you can see a large improvement. But Singlethreaded programs run slower.
    And i realized that Open64 is a pita if it comes to regular stuff. Still trying to compile any version of glibc ;P Using it as system compiler is almost impossible.

    Originally posted by deanjo View Post
    Shouldn't make any difference since a 990FX is a rebranded 890FX and identical in every way.
    Oh didnt know that*hides* thx for pointing that out.

    Leave a comment:


  • deanjo
    replied
    Originally posted by AgY! View Post
    but its not a 990FX board ;(
    MT performance seems to suffer .. a bit.
    Shouldn't make any difference since a 990FX is a rebranded 890FX and identical in every way.

    Leave a comment:


  • hermiti
    replied
    Originally posted by PsynoKhi0 View Post
    Welcome and thanks for the submission, I'm actually starting to get really interested in that particular BD chip.


    I think 1866MHz is one of the things the extra pins on AM3+ provide, correct?


    Neat! Can you poke around the cores to forcibly turn them off? There's likely to be extra punch to be had on lightly threaded programs.
    My motherboard supports up to 2133MHZ overclocked. I will try turning off the cores as discussed in that thread once I finish tweaking my system.
    Last edited by hermiti; 10-17-2011, 12:37 PM.

    Leave a comment:


  • curaga
    replied
    Originally posted by skies View Post
    I would love to see a run of the benchmark suite using AMD reference compiler Open64 (a forked version of GCC with features from SGI's MipsPro).

    It's free and open source just like GCC. Available here http://developer.amd.com/tools/open6...s/default.aspx

    As far as I know the stock GCC can not do any optimizions for AMD's new AVX/FMA3/FMA4/XOP instruction sets, you need Open64 for that.
    http://gcc.gnu.org/ml/gcc-patches/2009-09/msg01792.html

    Leave a comment:

Working...
X