The G5 sucks
I don't know if this has already been posted. German c't magazine has run a couple of tests on a dual 2GHz G5 and compared them to several other systems.
SPECint2000 SPECfp2000
Opteron246 1243 1249 (win, intel 7.0)
Opteron246 1040 ---- (linux,gcc3.2.2)
Athlon64FX51(2.2/PC3200) 1399 1406 (win, intel 7.0)
Athlon64FX51(2.2/PC2700) 1302 1354 (win,intel 7.0)
Athlon64FX51(2.2/PC3200) 1185 ---- (linux,gcc3.2.2)
Athlon64 3200+(2.0GHz) 1266 1188 (win,intel 7.0)
PowerG5 2x2.0Ghz,single 767 1001 (IBM 6.0,Fortran8.1)
Cinebench 2003 Rendering (more = better)
2x 2 GHz G5: 504
1x 1 GHz G4: 92
2x Xeon 3.06 GHz: 655
1x Athlon 64FX 2.2 GHz: 310
1x P4 3.2 GHz: 380
Photoshop 7 (less = faster)
2x 2 GHz G5: 278 s
1x 1 GHz G4: 796 s
2x Xeon 3.06 GHz: 287 s
1x Athlon 64FX 2.2 GHz: 337 s
1x P4 3.2 GHz: 362 s
Mathematica 5 (less = faster)
2x 2 GHz G5: 1021 s
1x 1 GHz G4: 2023 s
2x Xeon 3.06 GHz: 725 s
1x Athlon 64FX 2.2 GHz: 553 s
1x P4 3.2 GHz: 678 s
FileMaker 5.5 (less = faster)
2x 2 GHz G5: 82 s
1x 1 GHz G4: 147 s
2x Xeon 3.06 GHz: 70 s
1x Athlon 64FX 2.2 GHz: 46 s
1x P4 3.2 GHz: 62 s
MP3 encoding (less = faster)
2x 2 GHz G5: 98 s
1x 1 GHz G4: 284 s
2x Xeon 3.06 GHz: 68 s
1x Athlon 64FX 2.2 GHz: 89 s
1x P4 3.2 GHz: 91 s
MPEG-4 transcoding (less = faster)
2x 2 GHz G5: 42 s
1x 1 GHz G4: 85 s
2x Xeon 3.06 GHz: 35 s
1x Athlon 64FX 2.2 GHz: 41 s
1x P4 3.2 GHz: 39 s
UT 2003 Asbestos flyby 1024x768
2x 2 GHz G5: 67 fps
1x 1 GHz G4: 33 fps
2x Xeon 3.06 GHz: 197 fps
1x Athlon 64FX 2.2 GHz: 203 fps
1x P4 3.2 GHz: 203 fps
The MPEG4 and MP3 tests are probably not really optimized and should be much faster IMO. But I'm worried about the SPEC scores.
Code:
SPECint2000 SPECfp2000
Opteron246 1243 1249 (win, intel 7.0)
Opteron246 1040 ---- (linux,gcc3.2.2)
Athlon64FX51(2.2/PC3200) 1399 1406 (win, intel 7.0)
Athlon64FX51(2.2/PC2700) 1302 1354 (win,intel 7.0)
Athlon64FX51(2.2/PC3200) 1185 ---- (linux,gcc3.2.2)
Athlon64 3200+(2.0GHz) 1266 1188 (win,intel 7.0)
PowerG5 2x2.0Ghz,single 767 1001 (IBM 6.0,Fortran8.1)
Cinebench 2003 Rendering (more = better)
2x 2 GHz G5: 504
1x 1 GHz G4: 92
2x Xeon 3.06 GHz: 655
1x Athlon 64FX 2.2 GHz: 310
1x P4 3.2 GHz: 380
Photoshop 7 (less = faster)
2x 2 GHz G5: 278 s
1x 1 GHz G4: 796 s
2x Xeon 3.06 GHz: 287 s
1x Athlon 64FX 2.2 GHz: 337 s
1x P4 3.2 GHz: 362 s
Mathematica 5 (less = faster)
2x 2 GHz G5: 1021 s
1x 1 GHz G4: 2023 s
2x Xeon 3.06 GHz: 725 s
1x Athlon 64FX 2.2 GHz: 553 s
1x P4 3.2 GHz: 678 s
FileMaker 5.5 (less = faster)
2x 2 GHz G5: 82 s
1x 1 GHz G4: 147 s
2x Xeon 3.06 GHz: 70 s
1x Athlon 64FX 2.2 GHz: 46 s
1x P4 3.2 GHz: 62 s
MP3 encoding (less = faster)
2x 2 GHz G5: 98 s
1x 1 GHz G4: 284 s
2x Xeon 3.06 GHz: 68 s
1x Athlon 64FX 2.2 GHz: 89 s
1x P4 3.2 GHz: 91 s
MPEG-4 transcoding (less = faster)
2x 2 GHz G5: 42 s
1x 1 GHz G4: 85 s
2x Xeon 3.06 GHz: 35 s
1x Athlon 64FX 2.2 GHz: 41 s
1x P4 3.2 GHz: 39 s
UT 2003 Asbestos flyby 1024x768
2x 2 GHz G5: 67 fps
1x 1 GHz G4: 33 fps
2x Xeon 3.06 GHz: 197 fps
1x Athlon 64FX 2.2 GHz: 203 fps
1x P4 3.2 GHz: 203 fps
The MPEG4 and MP3 tests are probably not really optimized and should be much faster IMO. But I'm worried about the SPEC scores.
Comments
Wow, look, all these G4 optimized apps don't run so hot on a G5!!!
Keep your panties on.
Originally posted by mooseman
....lol, reminds of when the P4 first came out and all the 'tards were bitching about it being slower than the PIII before the apps were optimized. Then, once they were recompiled with the proper flags, they stomped the PIII.
Wow, look, all these G4 optimized apps don't run so hot on a G5!!!
Keep your panties on.
Cinebench is optimized. They used the IBM compilers for the SPEC scores.
It doesn't exactly blow the doors off the various PC offerings as Apple would have you believe, but it's in the right ballpark.
I also have a gut feeling that Apple will be shipping 3GHz G5s mid-year next year, at which point Intel will be struggling towards 4 GHz, and probably only achieving 3.5GHz.
Without knowing how much RAM is in the machine and what the Power Manager is set to, it's hard to judge those scores. I trust c't to be thorough about these things, so I assume they logged these details somewhere nearby.
Originally posted by 123
Cinebench is optimized.
Since when?
Originally posted by Amorph
Link? I couldn't find them in a quick check of c't's home page.
Without knowing how much RAM is in the machine and what the Power Manager is set to, it's hard to judge those scores. I trust c't to be thorough about these things, so I assume they logged these details somewhere nearby.
They want you to buy the magazine, only the first part of the article is available online: http://www.heise.de/ct/03/20/098/. The numbers were posted somewhere on the net with no additional information given, I saw them at ars. I agree that we need to know more before we can judge, but this is c't... I expect them (more than any other magazine) to have run the tests thoroughly. Their very first G5 report (wwdc) was biased, though.
Originally posted by Zapchud
Since when?
c't is a German magazine. Maxon is a German company.
Originally posted by 123
c't is a German magazine. Maxon is a German company.
Their relative locations don't matter.
The last anyone had heard from Maxon they were awaiting shipment of a G5 so they could try out the code on the machine. They probably have it by now, and I imagine they're hard at work profiling their code. But until we have specifics (such as whether c't was able to coax a preliminary build out of Maxon, which is unlikely) it's hard to say for sure what the relevance of the score is. Except that, as a benchmark, it's relevance is automatically suspect. (As for PSBench, Adobe employees including AltiVec guru Chris Cox have dismissed it as worthless.)
There is at least one potent G5-specific optimizations that can only run on a G5, so simulation will only get you so far. It would also be interesting to see how it runs under the more efficient Panther, compiled with a final version of the IBM compiler (I don't know if GCC can be told about the hardware square root function, for example, and that's a big speedup when it's used - but I don't know how much it would be used in 3D rendering?).
As Intel will tell you, half of a processor's speed comes from the quality of the compiler. gcc sucks. Given what I've seen regarding the architecture of the G5, it should be capable of destroying any x86 architecture on a watt-for-watt basis given a good compiler.
I also wonder where they got these machines, since AMD hasn't released the processor yet. I am skeptical that AMD would just hand over a bunch of new stuff to some German magazine. . .several months ahead of the release date.
Lastly, the SPEC bench numbers are for a single processor, and are lower than the ones Apple presented (Am I wrong to think that all SPEC should be the same?). The G5 has two processors.
Originally posted by Amorph
Thanks for the update on the location of the benchmarks. That means I did manage to find the article, or at least the part they posted online.
[...]
But until we have specifics (such as whether c't was able to coax a preliminary build out of Maxon, which is unlikely) it's hard to say for sure what the relevance of the score is.
You were able to find the article but unable to read it?
Near the bottom:
"Eine löbliche Ausnahme ist der 3D-Renderer Cinema 4D von Maxon, zu dem es mit dem Cinebench 2003 einen Benchmark in Mac- und PC-Version gibt, der nicht nur bereits das Hyper-Threading aktueller Intel-Prozessoren unterstützt, sondern mit einer uns zur Verfügung stehenden Vorabversion auch bereits für den G5 optimiert ist."
Which basically translates to: [There's hardly any benchmarking software available that runs on all platforms], one exception being Maxon's Cinebench 2003 which is available for the PC with hyper-threading support and for the Mac as a G5 optimized prebuild which is available to us.
Originally posted by Amorph
Except that, as a benchmark, it's relevance is automatically suspect. (As for PSBench, Adobe employees including AltiVec guru Chris Cox have dismissed it as worthless.)
I think the normalized PS7Bench score is a good CPU/System mix (integer & SIMD & bandwidth) benchmark. It is NOT a good application benchmark, though, which was what Chris was referring to IIRC. If c't's PS test was indeed PS7Bench, then this score is useless.
Originally posted by Splinemodel
I also wonder where they got these machines, since AMD hasn't released the processor yet. I am skeptical that AMD would just hand over a bunch of new stuff to some German magazine. . .several months ahead of the release date.
I don't think your scepticism is justified. c't is on of the most respected non-specialized computer magazines. If they say they tested something, they have tested it. Unlike some PC Magazine that hasn't even run the Xeon tests themselves (maybe I'm mistaken) and includes highly dubious stufff like "controls" that took minutes to load. They were able to get their hands on a lot of pre-release products in the past.
Lastly, the SPEC bench numbers are for a single processor, and are lower than the ones Apple presented (Am I wrong to think that all SPEC should be the same?). The G5 has two processors.
Apple's scores are: 800/840. c't used the new fortran compiler from IBM to improve the FP score (the question is, why do they get a lower SPECInt score?). I'm sure, rate tests are included in the print version, but I haven't been able to find them online.
Originally posted by 123
You were able to find the article but unable to read it?
Near the bottom:
"Eine löbliche Ausnahme ist der 3D-Renderer Cinema 4D von Maxon, zu dem es mit dem Cinebench 2003 einen Benchmark in Mac- und PC-Version gibt, der nicht nur bereits das Hyper-Threading aktueller Intel-Prozessoren unterstützt, sondern mit einer uns zur Verfügung stehenden Vorabversion auch bereits für den G5 optimiert ist."
Which basically translates to: [There's hardly any benchmarking software available that runs on all platforms], one exception being Maxon's Cinebench 2003 which is available for the PC with hyper-threading support and for the Mac as a G5 optimized prebuild which is available to us.
I can't read German, no.
At any rate, we'll see what happens when it comes out. But since Cinebench has always lagged on the Mac, regardless of architecture or CPU, I would not consider it a reliable indicator of overall performance. If it was, nobody serious about performance would get a P4, because the Athlon stomps all over it in Cinebench.
This is still early days.
No Panther. No decent compiler. Limited or no optimisation in apps.
Are any of those apps optimised to run on the G5? Mathematica is ongoing? PS 7 seems limited in its optimisations. But still a raw 2 gig G5 sees off a top Xeon system...and trails a little in Lightwave. An app' which hasn't been opt' for G5 yet. Wait until PS8 and Panther more likely. Unreal? Something clearly wrong there.
And we aren't that far from Rev B in the scheme of things (my hunch.)
Given all that. We're still not far along.
ie G5 unoptimised, giving twice the performance of a G4 in the above benches. Which is what IBM promised. Sure, the G4 is crap compared to PC scores...but the G5 is certainly competitive in it's raw, unoptimised form.
The AMD 64 looks no pushover. But you get to run it on Windows. With sobig worm virus etc. I can live without that. My Athlon is still down with that.
I've been on my wife's iBook weeks now. I can't wait to get my hands on a G5.
I'm confident an optimised Rev B G5 on Panther with PS 8 will more than hang with AMD's '64' bit cpu.
Until the G5 is optimised from Panther through to Apps then these 'tests' don't tell us very much except that if you're a G4 owner, you can probably get a nice boost by buying a dual 2 gigger G5 and it will get faster in the coming half year.
Sniff. Early days. Beige box AMD 64 vs gorgeous Panther on a gorgeous G5 Box?
No competition.
Lemon Bon Bon
MP3: iTunes 4.01 / Magix MP3 maker 2003 diamond
MPEG-4: QT 6.3 / Windows media Encoder 9
P4 3.2 GHz = Dell Dimension 8300 Toronto
2x Xeon 3.06 GHz (with 1M L3) = Dell Precision 450
all other x86 self built by c't.
Additional Photoshop test with 2GB RAM:
2x 2GHz G5 : 261s
2x 3GHz Xeon : 260s
3.2GHz P4 : 316s
MacOS X was 6S80 for 2GHz and 6S74 for 1.8GHz.
Cinebench was manually optimized by Maxon as they didn't had a G5 optimized compiler.
Memory transfer for the 2x 2GHz: 3.03GB/s for 1 CPU, 5.49GB/s for 2 CPUs.
Originally posted by Splinemodel
There's very little information conveyed here, but the G5 did win in Photoshop, which isn't exactly the most optimized of programs.
As Intel will tell you, half of a processor's speed comes from the quality of the compiler. gcc sucks. Given what I've seen regarding the architecture of the G5, it should be capable of destroying any x86 architecture on a watt-for-watt basis given a good compiler.
I also wonder where they got these machines, since AMD hasn't released the processor yet. I am skeptical that AMD would just hand over a bunch of new stuff to some German magazine. . .several months ahead of the release date.
Lastly, the SPEC bench numbers are for a single processor, and are lower than the ones Apple presented (Am I wrong to think that all SPEC should be the same?). The G5 has two processors.
This is actually a very good point.
The new AMD "Hammer" Athlon64 processors are already in the hands of reviewers, so that they have time to produce their reviews in time for the official release date of the products.
That said, anyone with these processors should be bound by an NDA (non disclosure agreement), so I find it a bit odd that this magazine is allowed to be posting benchmarks prior to the launch date.
Another reply above had a great point, in that nothing has been optimized and released at this point. The forthcoming Mac OS X "Panther" sounds very promising, not to mention third parties reworking their code for the G5s.
Originally posted by smalM
Benches were run with 512 MB on all systems, ATI radeon 9600 PRO, 64MB on Macs, 128MB on x86.
MP3: iTunes 4.01 / Magix MP3 maker 2003 diamond
MPEG-4: QT 6.3 / Windows media Encoder 9
P4 3.2 GHz = Dell Dimension 8300 Toronto
2x Xeon 3.06 GHz (with 1M L3) = Dell Precision 450
all other x86 self built by c't.
Additional Photoshop test with 2GB RAM:
2x 2GHz G5 : 261s
2x 3GHz Xeon : 260s
3.2GHz P4 : 316s
MacOS X was 6S80 for 2GHz and 6S74 for 1.8GHz.
Cinebench was manually optimized by Maxon as they didn't had a G5 optimized compiler.
Memory transfer for the 2x 2GHz: 3.03GB/s for 1 CPU, 5.49GB/s for 2 CPUs.
Ok, so MP3 and MPEG4 benches are useless (except to show us that QT doesn't make use of the second processor?). Different encoders, and not optimized for the G5.
I don't know why they didn't use a 64 MB Radeon for the PCs, seems a bit odd to me. However, I don't know if it really makes a difference in their tests.
Also, Mathematica strikes me as odd -- perhaps they had MP turned off? To enable MP you have to launch a slave kernel on the local machine with the command
LaunchSlave["localhost", "math -mathlink"];
-- Mark
Originally posted by 123
Ok, so MP3 and MPEG4 benches are useless (except to show us that QT doesn't make use of the second processor?).
The MP3 and MPEG4 benches are very useful because most Mac users will have to use quicktime to do those things. Until Apple releases a MP optimized quicktime, Macs will lag in those tasks. This is one thing that we can blame Apple instead of Motorola.