New barefeats article proves G4s sometimes faster, sometimes slower than AMD/Pentiums

[ddtlm]

alex_ant:

Great to see you fighting the good fight!

others:

As true as it is that the G4 is slower than most of its compeditiors, when it is performing as bad as the numbers that some people have posted here then I can just about assure you that the Mac is at a severe software disadvantage. I mean really, look at the specs of a G4, the worst case performance delta between it and a top-of-the-line PC should be maybe 4x or 5x, not these 10x and higher numbers. There are very few situations when a G4 should do less work per clock than a P4.

So lets try to remain realistic here. It is virtually gaurenteed that the actual performance potential of a 1.25ghz G4 falls between that of a 1.3ghz P4 and the 2.8ghz P4.

EDIT:
Almost forgot to talk about SPEC. Some time ago, the only SPEC results that I know of for Macs were obtained by c't:

http://www.heise.de/ct/english/02/05/182/

In these they showed the G4 was more or less the same speed as a P3 of equal clock (1.0ghz) in the integer tests, when both where done done with GCC. Intel's compiler can give the P3 at 30% edge or something, so we know that the quality of compiler is hurting the G4 here. It is not fair to look at SPEC and declare other chips to be a zillion times faster than the G4, simply because they are all using very good compilers whereas Apple is stuck with GCC. Apple is working to improve GCC however, so things may get better.

(In SPEC FP the G4 get beat worse, I might add. Compilers played a role for sure, but can't explain the whole loss.)

 

[MacCoaster]

Originally posted by ddtlm
I can just about assure you that the Mac is at a severe software disadvantage.

And hardware. No hardware double precision floating point Altivec unit.

 

[Backtothemac]

Originally posted by benixau
maybe, anyway I tell my buddies that a mac works. It is great to have all that speed but here is a thought:

I have a PC that is really 5x as fast as a mac
I spend 5x as long setting it up as i do the mac
I am also 5x less productive on it then a mac as it keeps breaking

I may not be a great mathematician but 5x5 = 25. 25x less usable than a mac. Personal experience proves this.

Long Live King Mac!! Long Live King Mac!!

For the dark side to wonder at how easy I get my life done

There someone in this thread actually gets it. Sure the PC may be faster, SO FRIGGIN WHAT! I will never go back to a PC, and do not know a single Mac user that really would. Sure we bitch and moan, but the fact is that we know that we are on a much better platform!

 

[ddtlm]

MacCoaster:

(Don't be offended if I repeat myself a few times, I want to make sure everyone gets it. Not trying to say anything about you in particular.)

Anyway, you missed my point. I know very well that the G4 is at a hardware disadvantage. I pretty much said that when you see a G4 being beat by margins greater than 4x or 5x, then you can be pretty sure there is ALSO, note ALSO, a software disadvantage. Hopefully everyone will see what I meant that time.

I'm glad to see that many people here agree that the G4 isn't really a faster chip than the x86 competition, but I want to see moderation and understanding of the "benchmarks" that have popped up showing an unbelievably bad situation for the G4.

Remember folks, if the test shows a G4 slower than a P4 per clock cycle then the test probably is handing the software advantage to the P4. Note, for perfect clarity, that I said per clock cycle performance and not overall performance.

 

[MacCoaster]

ddtlm:

Ah, thanks for clarifying.

 

[javajedi]

Originally posted by ddtlm
MacCoaster:

(Don't be offended if I repeat myself a few times, I want to make sure everyone gets it. Not trying to say anything about you in particular.)

Anyway, you missed my point. I know very well that the G4 is at a hardware disadvantage. I pretty much said that when you see a G4 being beat by margins greater than 4x or 5x, then you can be pretty sure there is ALSO, note ALSO, a software disadvantage. Hopefully everyone will see what I meant that time.

I'm glad to see that many people here agree that the G4 isn't really a faster chip than the x86 competition, but I want to see moderation and understanding of the "benchmarks" that have popped up showing an unbelievably bad situation for the G4.

Remember folks, if the test shows a G4 slower than a P4 per clock cycle then the test probably is handing the software advantage to the P4. Note, for perfect clarity, that I said per clock cycle performance and not overall performance.

If you recall the java program I created ran without modification on a p4/g4, in addition others on this board have ran it on their Athlon systems. The code is unbelievably simple, I did not give the p4 any "software advatage" whatsoever (and as I said, the code remained changed).

The only difference (and this could be a big difference), is the different versions of the jvm on the mac, and on windows. On my p4 pc I was using jvm version 1.4.x, while Mac OS X is limited to 1.3.x. To factor this variable out of the equation I decided to port it directly to Mac OS X and created a cocoa application. Java is now out of the equation.


The cocoa version, as well as it's source is located at http://members.ij.net/javajedi/FPMathTest.dmg.gz

My PowerBook G4 800 now takes *only* 94 seconds running natively. The P4 running the slower java version (slower because it’s interpreted and the byte code translation) finishes it in 5.9 seconds. Please feel free to take a look. I don't see how the P4, or any other of the x86 processors are cheating. I've tried to make it as fair and possible - to the extent of creating a cocoa app.


Thanks for your thoughts!

Kevin

 

[UnixMac]

my 500mhz MP did it in 87 seconds.

I am amazed however that a P4 can do it in 1/20 that time.... It almost says to me that there is a flaw in how software is coded. Can you use the Altivec to improve this?

I seem to recall that FP on the G4 was superior to PIII of double the MHz, so how can the P4 be THAT MUCH faster?

 

[javajedi]

This weekend I'm going to try to vectorize it with the Altivec and make it available for you guys. Frankly I don't know how simple or difficult this will be, but I'm going to look into it. As far as the code itself.. well.. it's very basic. Not only was it not vectorized for the G4, it wasn't vectorized for the P4. If anything the other platforms are at a software disadvatage because it's being ran under Java. The Mac OS X version is native code.

I think what we have learned from all this is that the G4 has a *REAL* problem with integer and double precision floating point. Ofcourse going to Altivec would bypass these registers and help considerably. But that just goes to show you without Altivec code you are far behind everything else.

Once again I'll see what I can do about an Altivec version, it should be very intresting indeed.

EDIT: I should also note that your 500MP didn't benifit from the extra processor, all of the math is being done in the main event loop.

 

[javajedi]

I just ran the cocoa version on a 700mhz iBook..



get this: 73 seconds! Still very slow compared to x86, but considerably faster than my 800mhz G4 w/L3 cache??
If I recall the processor in the iBook is a 750FX IBM chip.

Amazing. All the more reason to go with an IBM chip for the future Pro Macs.


Thought you guys might find this intresting..

Kevin

 

[ddtlm]

javajedi:

Yes, the JVM is the deciding factor here. If the Java takes that damn long on a G4 but goes fast on a P4, can can rest assured that the JVM Apple is distributing sucks compared to whatever one the x86 machines are using.

There is no way in heck that the performance delta can be so large without a large difference in quality of JVM. G4's may be slower, but they are not as slow as those number indicate.

Like I've been saying, when you start to see 5x leads by the PCs you need to start asking questions about the fairness of the benchmark. The G4 is better than 1/5 the speed. There are very few things were a P4 can get better performance per clock than a G4.

BTW:
Your G3 results as bizzarre as well, because of the contrast between them and the G4 results. Do not take it as proof one way or the other of the G3 or other IBM chips being superior to the G4. What we have here are raw numbers that defy a simple explanations. We should ask why these numbers are popping up, rather than running off with them as if they were uttered by a great voice in the sky or somthing.

 

[javajedi]

Originally posted by ddtlm
javajedi:

Yes, the JVM is the deciding factor here. If the Java takes that damn long on a G4 but goes fast on a P4, can can rest assured that the JVM Apple is distributing sucks compared to whatever one the x86 machines are using.

There is no way in heck that the performance delta can be so large without a large difference in quality of JVM. G4's may be slower, but they are not as slow as those number indicate.

Like I've been saying, when you start to see 5x leads by the PCs you need to start asking questions about the fairness of the benchmark. The G4 is better than 1/5 the speed. There are very few things were a P4 can get better performance per clock than a G4.

BTW:
Your G3 results as bizzarre as well, because of the contrast between them and the G4 results. Do not take it as proof one way or the other of the G3 or other IBM chips being superior to the G4. What we have here are raw numbers that defy a simple explanations. We should ask why these numbers are popping up, rather than running off with them as if they were uttered by a great voice in the sky or somthing.

I should note that the 90 second and 72 second results I just recently posted are from my cocoa implementation, not java.. so the jvm is out of the picture now on the mac.

Do not take it as proof one way or the other of the G3 or other IBM chips being superior to the G4.

Don't worry, I don't make assumptions like that. And no, I don’t think this does defy simple explanations. I will say that, what we are starting to see here is evidence that the scalar units (integer and fpu) in the IBM 750FX (G3) are more efficient than those in the Motorola G4.

If this is true, then my program hit it right on the nail. Also if this is true, it means there exist theoretical situations when using non altivec code that it would be faster on one of these newer G3 chips.


Also what alex said about how tedious it was to make altivec code, I would agree there is some truth to this. When you vectorize code (either for the P4 or G4), if you don't watch your p's and q's you can actually slow *down* your code. Just because you use the nice and special vector registers on these G4 and Pentium 4 processors does not mean you gain 5 times the speed. You literally have to take your methods back to the drawing board. You will only get peak performance out of pipelined, fully vectorized code.

None the less, scalar operations on both G3/G4 are miserable compared to x86. The JVM is no longer the deciding factor in the performance delta. It's out of the equation on the Mac since the benchmark is now a 100% native cocoa application with c code and no longer java. Mean while on the x86, the benchmark remains java.

70-ish seconds navtive on a G3
90-ish seconds on a native on a G4
5.9-6-ish seconds running under JVM 1.4.1 on a P4

 

[ddtlm]

javajedi:

70-ish seconds navtive on a G3
90-ish seconds on a native on a G4
5.9-6-ish seconds running under JVM 1.4.1 on a P4

Admittedly I am getting lost in what all the numbers people have mentioned are for, but looking at these numbers you have here and assuming that they are doing the same task, you can rest assured that the G3/G4 are running far inferior software. AltiVec and SSE2 or not, there is just nothing that can explain this difference other than an unfair playing field. There is no task that a P4 can do 11x or 12x the speed of a G4 (comparing top-end models here). The P4 posseses nothing that runs at 11x or 12x the speed. Not the clock, not the units, the bandwidth to memory and caches are not 11x or 12x as good, it is not 11x better at branch prediction. I absolutely refuse to accept these results without very substantial backing because they contradict reality as I know it. I know a lot about the P4 and the G4, and I know a lot about programming in a fair number of different languages, even some assembly. I insist that these results do not reflect the actual performance of the processors, until irrefutable proof is presented to show how they do.

I guess the 70 and 90 don't surprise me a lot for the G3/G4, depending on clock speed difference. But all this trendy wandwagon-esque G4-bashing is not correct just cause every one else is doing it. There are things about the G3 that are very nice, but the G4 is no slouch compared to it, and given the higher clock that it's pipeline allows, the G3 really can't keep up. The G4 not only sports a better standard FPU, but it also sports better integer units.

 

[ryme4reson]

ddtlm

If you know about programming languages, and still refuse to accept the scores on the test, check this out:

Here is the code snippet in question: (C not java, for the sqrt function)
double x1,x2,x3 ;
for (x1=1; x1<=20000; x1++) {
for(x2=1; x2<=20000; x2++) {
x3 = sqrt(x1*x2);
}
}
Very simple, 3 variables of double precision. 2 for loops that and 1 to each number, then multiply the 2 and square it.

This should not take 73 seconds on my 933 and 6 seconds on a PV 2.4 Gig.

Also, if the MHTZ myth was so true, then why would Apple be investing time and money into merklar. Imagine Head of Hardware: "today we release OSX on x86. Sure it was slower in the past to the the MHTZ myth, but now its not"

I think apple knows the G4 is a pile of crap, but they have no options, so they are hopefully finding a new supplier while doing there best to make the code as optimized as possible.

Not the clock, not the units, the bandwidth to memory and caches are not 11x or 12x as good

If you are a believer of the Myth then you should not find it hard to believe that this CAN be true. The PV is just faster, and better optimized hardware wise to accept standard code.

Apple has pushed that the G4, and the PowerMacs are faster for sometime. They say sure our harddrives are slower, but we are still faster. Our system bus is slower, but the computer is faster. Our Ram is slower, but the computer is faster. Sure our Processors are 50% slower, but still guys are computers are faster, and they look good. If you believe this BS, then surely looking at the sinple code, and listening to the results of people who are not selling or pushing anything cant be 2 hard to believe.

And yes I am a mac user, and if Apple is still in business in another year, my next computer will be a mac.

 

[javajedi]

Originally posted by ddtlm
javajedi:


Admittedly I am getting lost in what all the numbers people have mentioned are for, but looking at these numbers you have here and assuming that they are doing the same task, you can rest assured that the G3/G4 are running far inferior software. AltiVec and SSE2 or not, there is just nothing that can explain this difference other than an unfair playing field. There is no task that a P4 can do 11x or 12x the speed of a G4 (comparing top-end models here). The P4 posseses nothing that runs at 11x or 12x the speed. Not the clock, not the units, the bandwidth to memory and caches are not 11x or 12x as good, it is not 11x better at branch prediction. I absolutely refuse to accept these results without very substantial backing because they contradict reality as I know it. I know a lot about the P4 and the G4, and I know a lot about programming in a fair number of different languages, even some assembly. I insist that these results do not reflect the actual performance of the processors, until irrefutable proof is presented to show how they do.

I guess the 70 and 90 don't surprise me a lot for the G3/G4, depending on clock speed difference. But all this trendy wandwagon-esque G4-bashing is not correct just cause every one else is doing it. There are things about the G3 that are very nice, but the G4 is no slouch compared to it, and given the higher clock that it's pipeline allows, the G3 really can't keep up. The G4 not only sports a better standard FPU, but it also sports better integer units.

Keep in mind this test does not reflect balanced system performance. The point of this exercise has been to determine how the G4's FPU compares to an assortment of different processors and operating systems.

I'd like to know you you qualify "inferior software" on the x86. If the P4 is some how cheating, then all of the other processors are cheating as well. Again, we ran the exact same code. We even made it into C code on the mac for maximum speed. In fact I'd like for you to check the code out for yourself, so you can see there is no misdirection here. Keep in mind, other people here have ran it on Athlons in Linux and still get sub 10 second times. I've also had a friend of mine (who i can trust) run it under Yellow Dog on a G4, he got 100+ seconds. And I did not tell him the scores we've been getting on the Mac, I had him run the test first and tell me how long it took before I even said anything. The JRE and now Mac OS X have been factored out of this equation.


When you look at operations like these, for example scalar integer ops, that's all register. The fsb, bsb, or anything else doesn't matter. This is a direct comparison between the two units on the G4 vs everything else. Also, my question to you is, in what way are the integer and fpu units "better" in the G4? I did not build the chip so I can't say weather they are better or not better than those in the 750FX, but I can say I've ran a fair benchmark comparing the FPU on the G4 from everything to a P4, Athlon, C3, G3, different operating systems, on x86 Windows and Linux, and on the Mac, Mac OS X and Yellow Dog. The results are consistent across the board. What more "proof" do you want?

 

[UnixMac]

How does it run on an UltraSparc III 900?

How does it run on an Alpha?

Lets get an assortment of score, there could be a code bug for the G4, I am not an expert, but 10-20 times slower sounds like science fiction.

 

[jefhatfield]

comparing like products on market

when i got my ibook, which was manufactured in summer-1999 and listed for $1599 us, i got a 300 mhz G3 processor, 32 MB of 66 mhz sdram, 3 GB hard drive, 4 MB agp graphics, and os 9.0

the next day i bought a compaq presario 1272 laptop, manufactured in spring-1999, $1599 us, and i got a 366 mhz amd k6-2 processor, 32 MB of 66 mhz sdram, 4.3 GB hard drive, 2 MB pci graphics, and windows 98

i would clearly say that these two machines were marketed for students and home users who were then looking for a bargain computer under sixteen hundred dollars

while the higher clock speed compaq presario had a larger hard drive, more output ports, more software bundled, pcmcia, and floppy against the single usb ibook;

i found the ibook to be much faster in everyday use for e-mail, internet, and word processing

it would be fun to get an $1199 ibook and get an $1199 dell laptop and use these machines every day for three years and see what kind of performance i get from them

...of course, at $1199, the pc laptop would give me a dvd optical drive vs. the cd-rom in the ibook, and a 14" inch screen vs. the ibook's 12" inch screen, and the pc would include much more software

 

[MacCoaster]

Originally posted by WanaPBnow
How does it run on an UltraSparc III 900?

I don't know. I'll run it on an UltraSPARC II sometime when I can. My step-dad's box isn't loaded up yet.

Lets get an assortment of score, there could be a code bug for the G4, I am not an expert, but 10-20 times slower sounds like science fiction.

Really? Code bug? How? It's a simple C/C#/Java/obj-C program. The G4 shouldn't be so slow with a task oh so simple. It's also no bug that Altivec doesn't include hardware double precision floating point. But then again, we weren't testing them with hardware support--just testing the pure CPU power. In fact, if you don't believe us--please, we beg you, look at the source code. Nothing Altivec/SSE/SSE2/3DNow/any of that crap there. 10-20 times slower isn't science fiction when it comes to double precision floating point on the G4. It simply blows.

 

[UnixMac]

Oh well, I'm out of my league on this.... I'll defer to others..

 

[javajedi]

I think it was Back2TheMac who posted earlier in this thread "x86 plain sucks". The reason why he belives the x86 ISA and CISC are inferior is because Apple put out a bunch of marketing in the early days of the PowerPC touting RISC as superior new technology. In today's world, RISC processos really aren't RISC, and CISC processors really are CISC.


I recommend anyone who still believes in this spin to read this:
http://www.arstechnica.com/cpu/4q99/risc-cisc/rvc-1.html


It's most informative.
Enjoy

 

[jefhatfield]

Originally posted by javajedi
I think it was Back2TheMac who posted earlier in this thread "x86 plain sucks". The reason why he belives the x86 ISA and CISC are inferior is because Apple put out a bunch of marketing in the early days of the PowerPC touting RISC as superior new technology. In today's world, RISC processos really aren't RISC, and CISC processors really are CISC.


I recommend anyone who still believes in this spin to read this:
http://www.arstechnica.com/cpu/4q99/risc-cisc/rvc-1.html


It's most informative.
Enjoy

it's really most fascinating...thank you

some of us hardware side IT people often make fun of the software IT people and it is often because of the introverted way most of them act or their lack of knowledge of the hardware side of things

but what's interesting is that the hardware side techies like network engineers and desktop techs would not have anything to implement and maintain if it wasn't for those coders who make it all possible

i always hear a lot about the hardware side of apple's products and the praise they get when things are done right, but i rarely hear about the heroes in the background, the developers who make it all run smoothly

of all the products apple has ever made, the mac operating systems is what really makes a mac a mac

 

[snoopy]

Hate to drop in late like this, but the G3 had the same FPU as the 603, not the better one in the 604. When Motorola built the G4, they did not upgrade the FPU, but added AltiVec. This is what I understand. So, yes, double precision floating point does run poorly, with that old 603 FPU.

 

[Inhale420]

Originally posted by j763




[ANTI-WINDOWS]
BUT... i'd like to raise this important point. wtf are the win32 users using their CPU power for? Typing up word documents really fast? browsing the web with Internet Exporer v6.000.21312.185726351;SP1? or perhaps having to wait only 10 seconds for windows media player to launch? win32 is simply a craptacular operating system to the extent where it shouldn't be recognized (and i certainly don't recognize it) as a real operating system. mac and *nix (excl. linux-on-the-desktop) is where it's at. get over it.
[/ANTI-WINDOWS]

you gotta be ****ing kidding me. it's so amusing to witness the brainwashed and ignorant roam the earth. yes, i use the latest version of ie and browse these forums 10x faster than whatever mac browser you're using. i only have the default ie on my mac, because there's no point in installing other browsers when you have a pc.

i also have a hell of an easier time developing for the web using the tabbeb-based version of dreamweaver and coldfusion studio. i export 3ds artwork to flash, and the performance of my 2 year old 1ghz athlon is amazing. and when i'm done with work, I USE MY PC AS A GAME MACHINE. the only reason i have a mac, is because i really want to use them for 2d graphics, but apple really ****ing do something brilliant if they expect me to upgrade.

so can you explain what you mean by 'not recognizing' windows? that statement made absolutely NO sense. don't be such a bigot.

 

[javajedi]

What you are saying makes a lot of sense. Now that I think about, I too recall reading this somewhere.

Now that we know the real truth about the "better standard FPU", I thought it was time to shed some light on non vectorized G4 integer processing.

It still does 200,000,000 calculations, but this time I'm multiplying ints.

Motorola 7455 G4@800Mhz: 9 seconds (Native)
IBM 750FX G3@700Mhz: 7 seconds (Native)
Intel P4@2600Mhz 2 seconds (Java)

PowerPC 7455 integer processing is consierabley better than floating point (obviously less work doing ints), but still less per cycle than the Pentium 4.

Very intresting the G4 looses both floating point and integer to the IBM chip, at a 100MHz clock disadvantage.

I'm still waiting to see that "better standard FPU" in the G4. It seems the G4 is absolutely useless unless you are fortunate to have vectorized (AltiVec) code.

 

[PCUser]

This isn't going to further the discussion any, but... javahedi, perhaps you could post a link to the C code? I'd very much like to compile it with GCC under Linux and Windows. Just a curious benchmark, IMO. Thanks!

(Doesn't the benchmark do 1,600,000,000 calculations? 20,000 x 20,000 x 4 = 1,600,000,000... am I missing something? It does two adds, one multiply, and one sqrt per cycle. The loop cycles 400,000,000 times... ?)

 

[Backtothemac]

Originally posted by Inhale420


you gotta be ****ing kidding me. it's so amusing to witness the brainwashed and ignorant roam the earth. yes, i use the latest version of ie and browse these forums 10x faster than whatever mac browser you're using. i only have the default ie on my mac, because there's no point in installing other browsers when you have a pc.

i also have a hell of an easier time developing for the web using the tabbeb-based version of dreamweaver and coldfusion studio. i export 3ds artwork to flash, and the performance of my 2 year old 1ghz athlon is amazing. and when i'm done with work, I USE MY PC AS A GAME MACHINE. the only reason i have a mac, is because i really want to use them for 2d graphics, but apple really ****ing do something brilliant if they expect me to upgrade.

so can you explain what you mean by 'not recognizing' windows? that statement made absolutely NO sense. don't be such a bigot.

And I care why? It doesn't matter how fast you can surf on your PC. I can get around fast enough on my Mac. People who say Mac's are too slow are the same people that never take the time to watch a sunset or spend a day with their kid.

They are fast enough. They do what they are supposed to do the way they are supposed to do it.

The don't crash, don't get viruses, and don't look like something from the 1980s!