Slashdot Mirror
Comparative G5/G4 Tests
rocketjam writes "Barefeats.com has posted test results comparing a 2GHz G5 MP, 1.8GHz G5, 1.6GHz G5 and G4 MP's at 1.4GHz, 1.25GHz and 1GHz. They use Photoshop, Cinebench 2003 and a Bryce 5 render for tests. Bottom line is the G5 2GHz MP has the best bang for your buck, but you might think twice about trading in that dual processor G4 for a solo G5...the G4s hold their own quite well. They also say tests in Panther yielded significant increases in the G5 scores."
A lot of programs are yet to be at all tweaked for the G5s. There are a couple of new processor functions (hardware square root for instance) that may bring big gains. Also, there are old functions from the G4 that slow down the G5. Check out Apple's G5 performance primer. I read that the changes are enough to make the guy who made the mac rc5 engine want to re-write it for the G5 (no simple tweaks). He was hoping to get by with just some minor tinkering, but the chip does require a lot more than that to take advantage of it's potential.
Let me take another example; regarding CineBench.
http://www.postforum.com/forums/read.php?f=6&i=87
In that post, a guy called Richard from the Maxon development labs says this:
This compares to a 1.8ghz G5 score that I've seen of 188. Which means they're aiming to get a fair bit out of optimisation for the chip. Just as a means of further comparison...
http://www.imashination.com/bench.
You'll see the top score these people have recorded is for a dual Xeon 2.4ghz - with a score of 502. If the G5s make it up to over 500, that says a lot about the chip.
Finally, you've also got to include a mention of the compilers. Whilst some optimisation has been made to GCC, the GCC guys rejected a whole heap of improvements for the G5 because they were too platform specific. There's a good thread over at Ars Technica that discusses some pretty big gains when using the IBM XLC compiler. Other Mac-specific compilers should yield some pretty awesome gains too.
So, in summary - take these scores with a grain of salt. They're just the beginning.
-- james
g{The G5 is NOT going to excel at Altivec optimised code, the G4 will remain the Altivec champ until IBM puts the kind of Altivec resources onto the 970 that Moto put onto the 745x series}g
That's not strictly true, the problem is that altivec that's optimal for the 745x chips is worst case for the 970, since best case on the 745x makes heavy use of the vec_dst (prefetch) instruction, which is handled serially on the 970 and hence stalls the Altivec hardware.
also the G5 is a lot less likely to run into bandwidth bottlenecks with altivec code on large datasets.
Why on earth should I choose a G5 over a dual G4?
The G5 is faster.
The G5 accommodates four times the RAM (roughly), and the RAM is considerably faster.
The G5 has AGP 8x.
The G5 has USB 2.0.
The G5 has FireWire 800. (The last generation G4 did, but the generation currently being sold by Apple is next-to-last and so does not.)
The G5 has Airport Extreme. (Ditto.)
The G5 has Bluetooth. (Ditto.)
The G5 has internal serial ATA.
The G5 has digital audio I/O.
The G5 is considerably quieter.
If none of that stuff matters to you, get the G4. If it matters, get the G5. This isn't a clear-cut case of "G5 better than G4 in every way." But neither is it even remotely a case of "no reason to buy G5."
The vec-dst issue is one problem, but just review the G4e article at arstechnica and have a look at the G4e's Altivec execution unit design - it's a generation ahead of that in the PPC 970 (which is approximate in execution resources to the Altivec unit of the MPC 7400 - albeit running at 2-4 times the 7410's clock rate!)
That was classic intercourse!
Having listened in to a few programmers discussing that, it seems that it's pretty much only RC5 that actually manages to make full use of the Altivec unit in the 745x series chips anyway, and the benches I've seen that _haven't_ been littered with vec_dst show that the 970 scales in a fairly linear fashion over the G4 as far as altivec code codes (except it also maintains that scaling on larger datasets, since its got a fast memory subsystem, rather than a slow memory subsystem with a fast L3)
". The code is optimized for the mac. My friends P4 2.8ghz is only a little faster than my G4 733. Apple took advantage of this for years and had test results based on photoshop."
/somewhat/ different than those for prior PowerPC processors and the optimizations that took place for the G4 (such as using vec_dst in the AltiVec optimized portions of the code) actually work against the G5.
:-)
Your working with outdated information--by about 7 years.
The code is optimized for both platforms very heavily. A large portion of Photoshop's userbase is on Wintel machines and they were also paid (by Intel) awhile back to optimize their filters for Intel processors. Presumably they care about performance and keep both versions highly optimized. Saying that it is "more highly" optimized for the Mac is a fallacy--it hasn't even had more time to become optimized for the Mac, due to processor changes. The optimizations for the G3 and G4 are
I also heard a rumor, though I am not sure that it is true, that Photoshop would offload its interface to the graphics card for display on its own and that, with the introduction of QE, Photoshop was effectively rendering its display elements twice. This may be completely unsubstantiated though
While in general your point is good--that apps which are highly optimized for one platform are not necessarily highly optimized for another--Photoshop is a bad example.
Integrate Keynote and LaTeX
Yes, if we want to compare a Pentium4 Xeon or a Opteron, and even in comparisons such as this, we have to take the results with a (largish) grain of salt and realize that the PowerMac G5 *will* perform better in these tests given time for further optimizations--using xlc/xlf (which /preliminary/ reports show for the Jet3D benchmark will raise the score for a single 2 GHz G5 to around three times that of 2.66 GHz P4), recompiling in 64-bit, a system which is better optimized to take advantage of the processor, etc.
/now/ from a dual-processor G4 to either a single or a dual processor G5 the current benchmarks are more relevant. Otherwise, you are absolutely right, when it comes to benchmarks we still are premature.
:-)
This goes for everything.
OTOH, if you are thinking of upgrading
That being said, it is fun to look for the results
Integrate Keynote and LaTeX
"The G5 is NOT going to excel at Altivec optimised code, the G4 will remain the Altivec champ until IBM puts the kind of Altivec resources onto the 970 that Moto put onto the 745x series. "
/bandwidth limited/. The architecture that the G5 is based on is radically faster when it comes to things like loading from main memory or making optimal use of both processors.
Another poster has already mentioned the execution sterilization that occurs thanks to optimizations that worked quite well in the G4 (vec_dst et al), but there is another point on this issue that needs to be mentioned.
Even if the G5 performs slightly slower than the G4e with AltiVec enhanced code on an individual, it has the potential to be much much faster overall simply due to the architecture. One of the main problems with AltiVec on the G4e currently is that many of the apps that take advantage of it are
Integrate Keynote and LaTeX
The G5 supports PCI-X.
Integrate Keynote and LaTeX
I'm getting a G5 for my home, but that's because I do heavy video editing and effects, which should benefit a LOT from the faster system.
And you're not concerned with the 50% reduction in internal drive bays as compared to the G4? I already have 650 GB shoehorned into my Blue & White G3 for video (upgraded to a G4 processor), and that's not counting the boot drive. The G5 only has two hard drive bays and one optical drive bay compared to the four and two in the G4.
Oh, say does that Star-Spangled Banner entwine / The myrtle of Venus with Bacchus's vine?
If you want a prebuilt unit, I can verify that the Maxtor Personal Storage 5000XT works well. I have used it on several different Macs running OS X. No problems whatsoever.
Using large drives in Firewire enclosures may have been a problem at some point but I can assure you this is no longer the case (no pun intended).
Apple has said it will continue to make G4 Powermacs for as long as people will buy them.
If you read Hannibal's appendium to his 970 articles he points out that the Altivec looks better than he initially thought. This isn't to say there isn't room for improvement. But to say that the G4 is the Altivec champ seems somewhat incorrect.
Hannibal's Followup
It is true that with G4 altivec code the G5 won't be as efficient as it could. Perhaps for some code the G4 might do better. But this is a simple matter to fix with a grep through your code. But initially altivec code won't show off the G5 until programmers do a simple recompile.
Actually there was that trojan that was on some web pages. I seem to recall reading about that on the Battlefront over at ARS last week. Of course almost no one *got* it. I don't know if the vulnerability has been plugged by Apple or not. It was a problem in Safari 1.0. (And I don't believe there have been any updates to Safari outside of the Panther betas)