Slashdot Mirror
Apple's G5 Speeds Challenged
An anonymous reader was the first of a seemingly infinite stream of people to submit a URL to an argument that makes the case that the G5 isn't quite what Apple wants you to think of it. The evidence? Apple's own press material. Worth a read.
--- Often in error; never in doubt!
I can vouch for those unfortunate enough to have worked in the grocery industry (and have an idea of what that kind of mechandising entail) but this is hardly insightful. It happens on nearly everything that you buy.
When asking the pricing managers (which work for the chain, not an individual store) they replied that there was a study once done, indicating that there is a psychological tendancy to shy away from certain "maker" numbers as being too big. For example, the masses statistically believed that twenty dollars was too much to pay for item x, but for some reason, nineteen ninety-nine was not too much to pay for the same item. Funny thing is that with the same item, eighteen dollars would again be too much, but seventeen ninety-five wouldn't.
Even if the study is flawed or bogus, it is still being taught in the "front-line" marketing schools, (ie. grocery, drug-store, clothing, etc.) , and so I expect we will see nineteen ninety-five for many many years to come.
The author of the article makes the point that most programs use a single processor unless specifically written for using two, so we should downplay the dual processor results. A good point on the surface but examine it more deeply and it has two flaws:
1) This is Apple's Pro machine and many of the users are in the Graphic Arts, Audio and Film industry. The most siginificant programs in these fields do get optimized for the Mac platform.
2) I don't know about you, but it is normal for me to be doing several things at once on my computer. Listening to music, downloading email, munging video, plus about a hundred background tasks. The OS itself balances these separate tasks between the processors, so there is a very real and significant advantage to the dual processor even if the individual programs don't take advantage.
-I have no Sig yet I must scream...
32 bit memory addressing is 4GB not 2GB.
hmm.. that image seems to suggest they've tested a dual g5 against a single proc p4...?
Go read Jon "Hannibal" Stokes article about the world of benchmarking., over on his site, Ars Technica.
h tm l
http://arstechnica.com/cpu/2q99/benchmarking-1.
This will give you at least a basis for understanding why benchmarking is used, and what makes or breaks any given set of results. Also, feel free to argue about anything and everything that is said about these benchmarks, since, apparently, everyone of you is in the benchmarking labs day in and day out, testing systems and looking at the results on a scientific level.
I also think benchmark scores are, quite frankly, marketing bullshit. A processor designer can tweak a program and a compiler any number of ways to increase thier scores. The true test would be to use the SPEC benchmark suite with no flags set on the compiles for either platform. That way you are testing just the base processor, with no SIMD instructions, no disabling of the software prefetch algorhythms, no "cheats" as it were. Then test those same systems with every trick in the book thrown in. Then look at the difference. This will probably give you a better picture of the performance you will see in real world activities.
If you have a machine that absulotely sucks donkey when using no "cheats" and then you see this amazing boost in performance when the "cheats" are enabled, you probably are dealing with a highly optimized and specialized instruction set, which can be either very good for specific applications, but absolutely horrible for programmers who don't have access to, or don't bother to research, the abilities of that processor.
These are the benchmarks I'm interested in most. And it'll be at least late September before we see any of that.
Also, while all this is interesting, in an intelllectual sort of way, what about the actual perfomance gains over the current crop of G4's? Why not take a look at the difference between the SPEC scores of the dual 1.42GHz G4 towers, vs. the dual 2GHz G5's? That alone will tell you more about the increase in speed and power that has been delivered. If Apple had been smart, instead of trying to impress and piss off the x86 sparkheads they should have posted those scores as well, to give a real side by side comparrison between the speed and power of the G5 vs the bottlenecked, processor starving, gimp that is the G4. But that would make too much sense, wouldn't it? And you know marketing is all about confusing your consumer into beleiving that the latest and greatest is really what they want, not some old machine from 3 months ago...
Don't Ask Questions. I don't know the answers and even if I did I wouldn't tell you.
With regard to price, if you're after a high-end system, he represents that the high end of the Dell line comes in at $3680, yet rapidly returns to promoting the idea that a $2000 Dell is equivalent. In an effort to configure up an Intel system comparable to the new high-end Apple PowerMac G5, I ran the Dell configurator. It clocks in at $3939:and that's with a lesser video card and a smaller, slower IDE hard drive (add $840 for SCSI, a better comparison with Serial-ATA). I don't think I was being unfair in my selection of components. (OK, add $30 for a USB floppy on the Mac if necessary)
This guy certainly has a point about the non-optimized Intel benchmarks, but he reveals his prejudice by not offering a fair price comparison.
My God, it's Full of Source!
OUTSIDE_IP=$(dig +short my.ip @outsideip.net)
4 GB is for the OS and application TOGETHER. The stock behaviour on Linux and Windows is to give 2 GB to the OS and 2 GB to the application. You can go as high as 3GB to the application Linux, but there are some serious warnings against going even that high.
-sirket
Obviously you have never run CAE/CAM programs. (That being Computer Aided Engineering / Computer Aided Manufacturing.. I-DEAS, Pro Engineer, SolidWorks, Catia, etc.) This is the hard core stuff that Boeing, Ford and Toyota use that cost hundreds of thousands of dollars for one license. At my university, the undergrads get to use it on P4 machines and the grad students get to use it on 64 bit HP-UX workstations.
Even if you use a fairly simple FEA (finite element analysis) on something, for example finding the levels of stress in some objects you have modelled when it is bent in different ways, or modelling the flow of water or air through some pipe bends, this amount of RAM is very desireable. Basically the program builds and solves a bunch of 2000x2000 matrices for you. Even a simple one like the pipe bend took something like 3 hours on a P4/512MB and there was a multi-GB swap file needed. I was in the lab very late that night. That is where super-large amounts of RAM are necessary.
There is no saying which is right, and I don't think this guy was really trying to. If you read his writeup, he says that Apple claims a certain Dell model benchmarks at value $X, while Dell claims that the same model can do $Y.
He doesn't actually say that one or the other is correct -- he says that the most charitable thing you can do is split the difference and go with the average -- and the kicker is that even that midway point is higher than what Apple claims for the G5.
You've got a good point, but I think this guy is aware of it as much as you are. He's not saying that each vendor's analysis is authoritative, but that the truth probably lies somewhere in the middle, and that middle ground might or might not look to be in Apple's favor (in fact, it doesn't seem to be in Apple's favor).
DO NOT LEAVE IT IS NOT REAL
oops, not HAD enabled it, but HADN'T enabled it. Stoopid me
At work, we just bechmarked the Dell systems a month ago and got very similar results to Apple for the "base" rate. The article seems to be quoting the "peak" rate for the Dells. It's not valid to compare peak rates yet because gcc 3.3 and os 10.3 aren't really fully optimized yet.
The article also complains that using the NAGWare compilers is not a valid test since they're too slow. But I think the NAGWare compiler is a more vallid comparison than intel's compiler because most real-world computing is done with NAGWare because it fully implements the F95 spec and is more portable. In addition NAGWare is well tested for accuracy and it also very much cheaper.
The Dell benchmark numbers are pure fantasy. They never occur in real-world use.
There are 10 types of people in this world, those who can count in binary and those who can't.
I just read the veritest document myself. This guy did not show the complete picture at all. First, in favour of Apple, these tests were run on Mac OS X 10.2.7. I don't think this is the 64-bit kernel that can really take advantage of the G5. I'm not a mac head, but I'm guessing this is Smeagol?
So right from the start the G5 is seriously crippled in these tests. Especially if they don't even take advantage of 64-bit as seems to be the case. Now, on the the other points that the spl dude makes.
Sure the special malloc library seems a bit unfair. But then again, do these tests really focus on memory allocation? I'd think they are limited by CPU power more than memory allocation. In any case, it'd be nice if we saw results without this library.
Now for the comfusing part. The 8300 only has a single cpu. For the base tests, they use hyperthreading and an SMP kernel. They do the exact same test for the 650's base test, hyperthreading with a single processor and an SMP kernel. The G5 system is run with a single processor as well.
So it seems this test is well balanced and fair. The confusing part is the rate tests.
For the 8300 they have no hyperthreading and a uniprocessor kernel. For the 650 they have no hyperthreading and an SMP kernel with two processors. The G5 system is run with two processors. It's unclear why they chose not to use hyperthreading on the rate test. It could be that hyperthreading actually reduced the scores of these tests. I'm no expert on the SPEC tests and hyperthreading, but what I do know is that hyperthreading is an intelligent technology. It can't always increase speed, it depends on what kind of code it's running. In the rate test it's possible that hyperthreading is unable to yeild any improvements, in which case the overhead of enabling hyperthreading may make the scores worse than without hyperthreading.
At anyrate, the tests were a LOT more fair than the dpl guy makes them out. And considering that the G5 could be seriously crippled by not running 64-bit and who knows what other optimisatoins, I'd say that the numbers are still impressive.
My guess is the did use Altivec optimizations. See this section
For the Mac
â Installed theTachyon development environment version 6K452. This provides the appropriate development tools for generating the SPEC binaries and installs Appleâ(TM)s version of the GCC compiler ( version 3.3 build 1379 ) on the test system
For the Dell
â Downloaded GCC version 3.3 ( gcc-3.3.tar.gz ) from http://gcc.gnu.org.
â Followed the documented steps to build and installed GCC v 3.3 on the system.
And here from the appendix
-fast
This flag is used with C and C++ and specifically targeted to the G5 and enables G5 specific instruction usage, tuning and 64 bit arithmetic. In addition to enabling the -O3 optimization level, it also enables the use of C99 aliasing rules and relaxed IEEE math operations.
G5 Specific instruction usage sounds suspicious. I really like the relaxed IEEE math operations.
I also like this part
â Installed a high performance, single threaded malloc library. This library implementation is geared for speed rather than memory efficiency and is single-threaded which makes it unsuitable for many uses. Special provisions are made for very small allocations (less than 4 bytes). This library is accessed through use of the â"lstmalloc flag during program linking.
Doesn't say anywhere that they did the same for the Dell.
I don't think Apple was looking for to even of a field for this test.
Here are some SPEC results I googled for, commisioned from SUN on their Xeon based Fire V65x, running a single 3.06 GHz Xeon. You'll notice that they, too, disabled Hyperthreading. Obviously, Sun would have wanted these benchies to be as fast as possible. So, probably, the single thread used for SPEC scores is best suited by TURNING HYPERTHREADING OFF.
s 20 03q2/cpu2000-20030520-02193.pdf
Meaning, if Apple's results are reliable (which I think they are...levelling both machines by optimizing them for neutral operations and having them run neutral code), they tuned the Dell FOR SPEC. They didn't decrease its performance -- they probably increased it a bit.
http://www.specbench.org/osg/cpu2000/results/re
Just because you put the words "Fast" or "Hyper" in front of a chip's feature doesn't automatically make it faster, as any BIOS hacker knows.
Hey freaks: now you're ju
In reality, icc produces really good code for Athlon/Opteron. So good in fact that the SpecInt scores of Opteron with icc in 32-bit-mode are better than the scores with gcc in 64-bit-mode. -- Despite the extra 8 registers in 64-bit-mode.
Finally! A year of moderation! Ready for 2019?