c't releases G4 SPEC scores :-))!!

Discussion in 'General Mac Discussion' started by ruhe, Feb 23, 2002.

  1. ruhe macrumors newbie

    Joined:
    Feb 23, 2002
    #1
    G4 1GHz
    SPECint_base2000: 306
    SPECfp_base2000 : 187

    compiled with OS X gcc 2.95.2, Absoft Pro Fortran 7.0

    Pretty weak for a "7.5 GFLOP" CPU, isn't it?!! :)))))
     
  2. ruhe thread starter macrumors newbie

    Joined:
    Feb 23, 2002
    #3
    What "what"? The German c't computer magazine has released G4 Spec benchmark scores (http://www.spec.org). This is the first time somebody is able to run Spec CPU2000 on a Mac and the extremely weak results show why Apple did not do it yet. A Mac system is just a loser in the field of scientific computing....
     
  3. LordRPI macrumors newbie

    Joined:
    Jan 25, 2002
    Location:
    San Diego
    #4
    If I'm not mistaken, not all versions of gcc 2.95 includes AltiVec support. AltiVec is the major source of number crunching on the G4.

    Were the specs done through the AltiVec unit?
     
  4. ruhe thread starter macrumors newbie

    Joined:
    Feb 23, 2002
    #5
    Alivec is no source of number crunching because is does not support double precision floating point. It can help in SPECint, but has no use in SPECfp which is using double precision like in most scientific applications.
    I don't know if gcc 2.95.3 or Abosoft Pro Fortran 7.0 support alivec. I am going to try to find out if they do so...

    Julian
     
  5. Taft macrumors 65816

    Taft

    Joined:
    Jan 31, 2002
    Location:
    Chicago
    #6
    OS X support just came out.

    I can't find any results on the spec web site. Doesn't look like the results from the company have been submitted or approved yet. Do you have a link to the company's web site or where I can find their results?

    And in response to your "Mac sucks for scientific computing" comment, I've got one word: BLAST. The results from the BLAST tests have been VERY positive.

    Way to Troll.

    Matthew
     
  6. ruhe thread starter macrumors newbie

    Joined:
    Feb 23, 2002
    #7
    The results will not be published on the SPEC website. The famous German computer magazine c't did the testing and they usually do submitt the scores to SPEC.
    A SPECfp_base of 187 really sucks and disqualifies the Mac for most of the scientific double precision floating point computations.

    Julian
    ps. We got the equal weak results for the G4 in the ATLAS (Automatically Tuned Linear Algebra Software, http://www.netlib.org/atlas) project, although using handcoded assembly. OK,
    the G4 runs well in single precision due to Alivec, but hardly anybody uses single precision in sci. computing.
     
  7. Gelfin macrumors 68020

    Gelfin

    Joined:
    Sep 18, 2001
    Location:
    Denver, CO
    #8
    Trust me, a computer magazine is not remotely qualified to perform the kind of system tuning that goes into producing a SPEC result.

    For a real comparison, among other things, you really want Apple and Motorola to chain some compiler experts to their desks for a few weeks with a reference compiler for the G4 (instead of gcc).

    The reason this result just came out is that the most recent SPECcpu2000 release supports Mac OS X, allowing curious German computer mags to run it for themselves. They probably don't realize they're doing a disservice to the Mac community by taking it upon themselves to publish results on a system that isn't insanely tweaked. That's the way every official SPEC result is done.
     
  8. Taft macrumors 65816

    Taft

    Joined:
    Jan 31, 2002
    Location:
    Chicago
    #9
    My point...

    ...is that the results have really not been publically compared or analyzed. They haven't been summitted yet to the source of the benchmarking tool--where eventually more than a single benchmark is submitted for a processor, system, etc. and compiled into an overall result. As I pointed out in my previous post, the SPEC system has just been released for OS X. My first question is wether they even used the OS X version and my second question is how many iterations of the software will it take before the software is properly tuned for OS X.

    There is a *lot* more information needed before this benchmark can even be considered anything more than a rumor.

    I will not dispute that *if* the number is a valid and fair benchmark--*if* it is-- then it is not a very good number.

    I have serious doubts about the quality of this number however. My biggest doubts stem from the fact that this is a magazine. Would I trust benchmarking information from MacWorld or PC Reader or something? Not as accurate benchmarks, no.

    Come at us with hard facts and *information* to back up your statement that the Mac floating point benchmark is 187. That single number is practically meaningless.

    A simple web link to more information would even help.

    Matthew
     
  9. ruhe thread starter macrumors newbie

    Joined:
    Feb 23, 2002
    #10
    The table of contents of the latest c't magzine can be found here (serach for "G4"):
    http://www.heise.de/ct/inhalt.shtml

    The article is *not* online available. The c't is qualified to run the SPEC because they always do it for various CPUs since years. The c't is known for its in dept analysis and its technical knowledge.
    German readers of the c't here in this forum will probably confirm my statements.

    Don't force me to translate the whole article. I will give you some facts from the article:

    (1) They test a 1GHz G4 under OS X, in single user mode, no gui and second CPU disabled.
    (2) They used gcc 2.95.3 delivered with Apple Developer Tool in Jan. 2002 for the C routines in the SPEC.
    (3) They used gcc because it performed *better* than Code Warrior in SPEC
    (4) They did not tweak the system in any way, they just ran SPEC CPU2000_base like described in the SPEC rules (the same compiler switches in all tests.
    (5) Thy ran SPECint_base on a PIII 1GHz with gcc 2.95.3 for x86 for comparison and they got 309 SPECint_base. Intel C would deliver much better results (see SPEC webpage)
    (6) SPECfp_base was done with Absoft Pro Fortran 7.0 because there is no fast alternative Fortran compiler for the G4

    Julian
     
  10. Taft macrumors 65816

    Taft

    Joined:
    Jan 31, 2002
    Location:
    Chicago
    #11
    Double precision FP on G4.

    I've been doing some reading on the web about the G4e processor. It does appear to have limitations when it comes to double precision floating point calculations. I still don't know about that number however. AND I think that to say the G4 is useless for scientific computing is quite an overstatement. It IS useful for all kinds of scientific computing. Just because it might be slower in DP doesn't mean it is out of the running in this category and it doesn't negate all of the other G4 features that ARE great.

    And don't tell me this thread wasn't meant to anger the Mac faithful. It certainly had a bite to it. I call it flamebait.

    Matthew
     
  11. Taft macrumors 65816

    Taft

    Joined:
    Jan 31, 2002
    Location:
    Chicago
    #12
    Oh...

    And I haven't yet seen any technical information on the Apollo G4s (the processors in the Dual 1GHz). So Gods know about what they are capable of. I'm sure we'll be seeing more technical information and benchmarks about this processor soon. This is a debate that can't be won without more and more detailed information.

    Matthew
     
  12. AmbitiousLemon Moderator emeritus

    AmbitiousLemon

    Joined:
    Nov 28, 2001
    Location:
    down in Fraggle Rock
    #13
    this is blatant "flamebait." your first post just shows this. you just throw out some numbers with no explanation and then throw in a mean spirited comment.

    as far as the later comment about being useless for scientific work. what a joke. nearly every genetics lab in the world uses macs. the only thing better than a mac for data analysis is a supercomputer, which we have to use at time with large data sets.

    the scientific community has embraced macintosh since day one and has never let go. g4s have always been weak with floating points, which has been a point of critizism since day one, but even with this the g4 does remarkably well it fair tests. it is rumored that the g5 will improve this flaw.
     
  13. Gelfin macrumors 68020

    Gelfin

    Joined:
    Sep 18, 2001
    Location:
    Denver, CO
    #14
    I don't care how long they've been running amateur SPEC results, they're still amateurs. First, I wouldn't put much faith in SPEC numbers run by ANYBODY who didn't design the system under test. Second, in almost all cases, you don't want your SPEC results to come from gcc, the fact that it performed better than CW notwithstanding.

    Uh. See, this is a misunderstanding of the SPEC rules. In other words, they ran an unoptimized build (except for maybe /O2).

    Thank you. That's EXACTLY why you don't want your SPEC numbers coming out of gcc. It's not that Intel's C compiler is so much better (though Intel does make an excellent reference compiler). It's that the compiler is designed by people who know EXACTLY what optimizations are possible with the chip and chipset, and they write compilers which expose those features. And notably, I seriously doubt that "c't magazine" could reproduce those high Intel scores, even WITH the reference compiler. It's a lot more complicated than booting single-user and running the test. Running gcc on both systems doesn't tell you much at all, because it could just be that gcc optimizes better on Intel than on Motorola chips.

    You've done an excellent job of proving my point for me. This magazine may represent itself to you as some kind of authority, but the details you gave are not indicative of a serious effort. It certainly isn't possible to compare those numbers meaningfully to published results.
     
  14. Catfish_Man macrumors 68030

    Catfish_Man

    Joined:
    Sep 13, 2001
    Location:
    Portland, OR
    #15
    This is aside from the fact...

    ...that any artificial benchmark (SPEC, GFlops, etc...) is not particularly useful. If I remember correctly, arstechnica has an excellent article on benchmarking that agrees with this. Basically, I care about real world performance in apps that I use (if you tell me it can render a heavily refractive glass mountain in Bryce 5 in 30 seconds I'll be [very] impressed, if you tell me it got a SPECwhatever score of 800 or something, I might care enough to go run some tests of my own on it). I still think someone needs to make a benchmark that compares times/framerate/etc... in all the common categories against a base machine, then takes the average of that. It would still be possible to skew (Optimize Quake III to get 800 fps and it would throw the averages out of whack), but it would give you a general indication of how the machine was going to perform.
     
  15. oldMac macrumors 6502a

    Joined:
    Oct 25, 2001
    #16
    troll/bait/etc...

    Anyway... real scientists use integer math. :)
     

Share This Page