New benchmarks available

in General Discussion edited January 2014
Head over to<a href=""; target="_blank">GMP Bench</a> to see a direct comparison of G4s to P4s , Althons, Itaniums, Alphas, Power3s, and Power4s.

These benchmarks are not the same as SPEC, but part of me thinks that they are more honest as it is completely open source. For the base functions, i.e. multiply, divide, add, subtract, and a few others, developers (or manufacturers for that matter) can supply assembly versions of those functions to eke out better performance. Granted, at this time there are only a small number of benchmarks available, but that will be expanded over time.

BTW, the Power4 1.1 GHz is only surpassed by an Alpha 21264 at 1 GHz and an Althon 2200+. The P4 at 2.4 GHz falls behind it. The G4 667 MHz is 1/4 the speed of the listed Power4.



  • Reply 1 of 11
    majormattmajormatt Posts: 1,077member
    Yay! We have the worst proccessor on the list!
  • Reply 2 of 11
    davegeedavegee Posts: 2,765member
    [quote]Originally posted by MajorMatt:

    <strong>Yay! We have the worst proccessor on the list! </strong><hr></blockquote>

    Ummm they were benching a 667Mhz G4... Does Apple even use those any more?
  • Reply 3 of 11
    os10geekos10geek Posts: 413member
    No...they use processors that are over 2X that speed...
  • Reply 4 of 11
    A couple others things of note about these benchmarks. First, these only encompass integer instructions and not floating point. Second, there is no AltiVec, SSE, or other vector instructions affecting the benchmarks.

    The reason I posted this link was to give everyone a better idea of how the 970 will compare to other CPUs as it is based on the Power4.
  • Reply 5 of 11
    hmurchisonhmurchison Posts: 12,341member
    This should definitely be in Current Hardware. When I say the title I thought we'd be treated to some benchmarks of the PPC 970. :cool:
  • Reply 6 of 11
    os10geekos10geek Posts: 413member
    Yeah...Iwanted to see the 970 kick a Pentiums little a$$..more than anything, I want to see how the 970 stacks up to the Xeon / Itanium.
  • Reply 7 of 11
    fran441fran441 Posts: 3,715member
    Moving to General Discussion.
  • Reply 8 of 11
    splinemodelsplinemodel Posts: 7,311member
    what is a benchmark anyway? I'm busy writing version of bzip2 for ppc that will piss all over anything x86. (To hell with compilers ) So, it uses Altivec. Big deal. PPC has Altivec. It's part of the chip. x86 does not. Integer means nothing if you don't consider vector performance as well.

    BTW: I've used platforms that utilize several of these chips. Most aren't impressive. SPARC 1GHz is crap, for one thing. This test is BS. I'll tell you as a firsthand witness. . . . And make sure to download my damn-fast compressor when it's ready this summer. :cool:

    [ 03-14-2003: Message edited by: Splinemodel ]</p>
  • Reply 9 of 11
    Altivec is nice, but only for functions that can be vectorized. It is horribly inefficient with multiprecision numbers as evidenced by Apple's own multiprecision package. Apple's mp package has an implementation of ECM (Elliptic Curve Method, used for factoring large numbers) which is slower than GMP-ECM, a version of ECM based upon GMP. So for the distributed projects I participate in, Altivec doesn't provide me much. Now if someone wants to write a quadratic sieve (either multipolynomial or self-initializing fors) or a number field sieve for Altivec (there are also used in factoring large numbers), I would be very curious to see how they perform compared to PPSIQS or mpqs4linux, which only run on x86.

    One last thing about Altivec, it doesn't support doubles, so GIMPS (Great Internet Mersenne Prime Search) can't use it.

    BTW Splinemodel, are you the same person behind the Altivec version of irred.c? I ask because I know that the person who wrote that is writing an Altivec enhanced version of bzip.
  • Reply 10 of 11
    ast3r3xast3r3x Posts: 5,012member

    in the words of jon stewart


    :confused: <img src="confused.gif" border="0">
  • Reply 11 of 11
    splinemodelsplinemodel Posts: 7,311member
    [quote]Originally posted by rogue master:


    BTW Splinemodel, are you the same person behind the Altivec version of irred.c? I ask because I know that the person who wrote that is writing an Altivec enhanced version of bzip.</strong><hr></blockquote>

    Nah, but I've read about that. I'm actually more concerned about writing a Burrows-Wheeler based assymetric lossless compression. Writing BZIP isn't too much different, but it uses Huffman, which isn't great for the altivec in it's normal incantation. My compression is somewhat different, and is more "vectorizable." Even my version of BZIP arrives at the output quite differently than the GNU version. . . in theory. It's a side project, so I'm still working on it at a slow-ish pace.

    But I guess the bottom line here is that there's more than one way to skin a cat, or in our case, to write an algorithm that takes X and makes it Y. Most of the time there's a clever way to do it in vectors.

    [ 03-16-2003: Message edited by: Splinemodel ]</p>
Sign In or Register to comment.