Announcement

Collapse
No announcement yet.

AMD's FX-8150 Bulldozer Benefits From New Compilers, Tuning

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

  • AMD's FX-8150 Bulldozer Benefits From New Compilers, Tuning

    Phoronix: AMD's FX-8150 Bulldozer Benefits From New Compilers, Tuning

    After looking at how Intel's Sandy Bridge processor performance has evolved with the new GCC 4.7 compiler and Apple's forthcoming LLVM 3.1 with Clang, here are benchmark results from the AMD FX-8150 "Bulldozer" Eight-Core processor with GCC 4.7.0 and the latest LLVM/Clang 3.1 development code along with looking at the performance impact of various compiler tuning flags for this latest-generation AMD CPU. Making things even more interesting, AMD's Open64 4.5.1 compiler was also tossed into the testing mix.

    http://www.phoronix.com/vr.php?view=17265

  • Luke_Wolf
    replied
    Originally posted by Kano View Post
    Maybe take a look at the 4 cores of bulldozer, those are combined of 2 integer units and 1 fpu. You have got 8 integer units and 4 fpus then. If you find benchmarks where bulldozer shines you can say that those definitely do not use many fpu commands. Bulldozer is definitely a fake 8 core,
    It's not 2 interger units with 1 FP unit, it's 2 interger units with 2 128 bit FP units that can act together as 1 256 bit FP unit, the problem is that the compiler and schedulers need to be set up properly in order to take advantage of this fact.

    Originally posted by Kano View Post
    you have to select your software very well to get increased speed, the older Thurban (x6) had "real" cores. One big problem with all amd cpus is that they work pretty inefficient compared to intel ones. If you look at povray (1 core bench) or cinebench 1 core results then even very cheap intel cpus can beat amd's top models. And when you know that most apps are only using 1-2 cores then you know what cpus are faster if you don't compile all the day
    Of course Intel beats out AMD at single threaded operations, however AMD beats out Intel at heavily threaded ones, this has been the status quo forever, this doesn't mean that AMD's is any less efficient. Also the point of the Bulldozer design is completely on multithreading and in the future heterogenous compute.

    And simple fact is that if the application is old enough to be so lightly threaded you're not going to see the difference anyway because it doesn't require that much single threaded performance, however where you need it with multithreaded performance AMD will be faster.

    Leave a comment:


  • Qaridarium
    replied
    Originally posted by Kano View Post
    Maybe take a look at the 4 cores of bulldozer, those are combined of 2 integer units and 1 fpu. You have got 8 integer units and 4 fpus then. If you find benchmarks where bulldozer shines you can say that those definitely do not use many fpu commands. Bulldozer is definitely a fake 8 core, you have to select your software very well to get increased speed, the older Thurban (x6) had "real" cores. One big problem with all amd cpus is that they work pretty inefficient compared to intel ones. If you look at povray (1 core bench) or cinebench 1 core results then even very cheap intel cpus can beat amd's top models. And when you know that most apps are only using 1-2 cores then you know what cpus are faster if you don't compile all the day
    that is how the intel monopole game works.
    amd chose to build a server cpu for there opensource customers.
    the desktop cpus and notebook cpus are just unimportant for amd!
    unimportant because they go for APU on notebooks and lowend desktop!
    the GPU in the APU goes bigger and bigger and the cpu dosn't matter.

    on the other side be sure the opteron 6308 will be a nice cpu! (also single-threated just because of higher clock speed because of the new clock engine! and some speed-bug-bugfixes!)

    Leave a comment:


  • ssam
    replied
    Originally posted by crazycheese View Post
    Of course CPU matters! Everything adds up, but CPU and I/O throughtput are most important.


    In Linux, thread=process, so both threading and parallel processing are utilized, so no - you have not lost if your app is "single threaded", you just make sure that app forks good.
    sorry i should have said "single threaded or single processed" or "only uses a single core"

    if you have an app that is only using 1 core, and you buy yourself the processor with the best single core performance, then you might make it 20% faster. congratulations. if you could make that app multithreaded (or multi processed, it makes no difference if you use openmp, mpi, fork, whatever), then you will probably get close to a 2 or 4 times speed up on your existing hardware. win. if you code is already multithreaded, then you will want to add more cores to make it faster.

    now if you have a CPU bound job, that really is non-parallelisable, then you are getting into interesting problems. but i think most people here are interested in things like games, graphics, and video.

    Leave a comment:


  • crazycheese
    replied
    Originally posted by ssam View Post
    most apps are not CPU bound so it does not matter what CPU you have.
    Of course CPU matters! Everything adds up, but CPU and I/O throughtput are most important.

    Originally posted by ssam View Post
    if you have CPU bound apps that are single threaded, then you have already lost.
    In Linux, thread=process, so both threading and parallel processing are utilized, so no - you have not lost if your app is "single threaded", you just make sure that app forks good.

    Leave a comment:


  • ssam
    replied
    Originally posted by Kano View Post
    And when you know that most apps are only using 1-2 cores then you know what cpus are faster if you don't compile all the day
    most apps are not CPU bound so it does not matter what CPU you have. if you have CPU bound apps that are single threaded, then you have already lost.

    Leave a comment:


  • Kano
    replied
    Maybe take a look at the 4 cores of bulldozer, those are combined of 2 integer units and 1 fpu. You have got 8 integer units and 4 fpus then. If you find benchmarks where bulldozer shines you can say that those definitely do not use many fpu commands. Bulldozer is definitely a fake 8 core, you have to select your software very well to get increased speed, the older Thurban (x6) had "real" cores. One big problem with all amd cpus is that they work pretty inefficient compared to intel ones. If you look at povray (1 core bench) or cinebench 1 core results then even very cheap intel cpus can beat amd's top models. And when you know that most apps are only using 1-2 cores then you know what cpus are faster if you don't compile all the day

    Leave a comment:


  • kukreknecmi
    replied
    I wonder if all C-ray tests are under same conditions? If so when comparing

    i7 - 3960x with gcc 4.7 rc1 = 25,7secs and
    FX-8150 with gcc 4.7 march=bdver1 = 26,92

    doesnt it give fx-8150 a uberbig performance jump??

    Leave a comment:


  • fuzz
    replied
    Originally posted by BlueJayofEvil View Post
    That's what Sabayon is trying to achieve with its Entropy system.
    Thanks, didn't know that. Now I have cause to check Sabayon out... .

    Leave a comment:


  • BlueJayofEvil
    replied
    Originally posted by fuzz View Post
    Would be neat to see something like Gentu (Ubuntu + Gentoo lol) with a spawn of portage + aptitude... binaries for all different processors (you could do binaries for all different USE flags too, theoretically, but that would just be a HUGE undertaking with some packages having 100+ USE flags that's at least 100^100 builds). I think it could be done but that just makes the repositories that much more critical etc.
    That's what Sabayon is trying to achieve with its Entropy system.

    Leave a comment:

Working...
X