G5 Altivec Carbon Fractal Benchmark

Wait a second.... I though Altivec was toned down on the 970s.

EDIT: Nevermind, I guess it actually is.

450 x 3 (to get to roughly 1.42GHz) x 2 (for SMP) G4 = 9 GigaFlops

2000 x 2 G5 = 14 GigaFlops

So, the 970 can only crank out a linear increase despite its enormous FSB.

What, so approximately 9 times faster then?

Alia

14 GigaFlops!?! Cool. Remember, the last G4 had a theoretical max of what... 18? This (if correct) is actual. And pretty good.

Edit: Wait... single or double? Wasn't the dual G4 18 (2x 9)?

Anyone know what the #s are for Intel or AMDs top offerings?

Edit again: Sorry, didn't notice the Altivec part. But still, does anyone know how many Flops the P4s and Athlons can do? (Obviously not on this, but in general)

Originally posted by rice_web
So, the 970 can only crank out a linear increase despite its enormous FSB.

Click to expand...

That seems odd, no? I looked at the link and they said that a 450 G4 (which I have) does 1.5 GFlops. That means that the Dual 2.0 G5 is a little shy of 10x the power of my machine at home. You'd think it be a little better.

Although it might depend on the application as well..

I'd love to do a Lightwave Rendering test....

D

Well I submitted a story link to Macrumors yesterday which showed the same sorta disappointing linear type performance gains (in fact even worse) . The story is on Think Secret and the relevant quote is:


Another screenshot shows results from Skidmarks GT, a benchmarking utility that's part of Apple's CHUD performance tools. Based on the test we ran, this provides a rough comparison of G5 and G4 performance. The Skidmarks scale has "100" equal a Power Mac G4 at 1GHz. The Dual 2GHz received scores of 172 for integer performance, 270 for floating point performance, and 208 for vector performance.



http://www.thinksecret.com/news/wwdc03g5.html

Originally posted by fred
Well I submitted a story link to Macrumors yesterday which showed the same sorta disappointing linear type performance gains (in fact even worse) . The story is on Think Secret and the relevant quote is:


Another screenshot shows results from Skidmarks GT, a benchmarking utility that's part of Apple's CHUD performance tools. Based on the test we ran, this provides a rough comparison of G5 and G4 performance. The Skidmarks scale has "100" equal a Power Mac G4 at 1GHz. The Dual 2GHz received scores of 172 for integer performance, 270 for floating point performance, and 208 for vector performance.



http://www.thinksecret.com/news/wwdc03g5.html

Click to expand...

But was the benchmark optimised for 64bit and the PPC970? Optimising ofr 64bit can make a big difference. Maybe the PPC970 does not run 32bit code as good as Apple wants us to believe. Anyway the G5 is much faster in games like Quake3, so I am ok with the G5 performance.

Originally posted by solvs
Anyone know what the #s are for Intel or AMDs top offerings?

Click to expand...

Errm did you see the name of the benchmark? Altivec Carbon Fractal Benchmark When was altivec and carbon are implemented in and on their cpu's?

Originally posted by fred
Well I submitted a story link to Macrumors yesterday which showed the same sorta disappointing linear type performance gains (in fact even worse) . The story is on Think Secret and the relevant quote is:


Another screenshot shows results from Skidmarks GT, a benchmarking utility that's part of Apple's CHUD performance tools. Based on the test we ran, this provides a rough comparison of G5 and G4 performance. The Skidmarks scale has "100" equal a Power Mac G4 at 1GHz. The Dual 2GHz received scores of 172 for integer performance, 270 for floating point performance, and 208 for vector performance.



http://www.thinksecret.com/news/wwdc03g5.html

Click to expand...

How can you be disappointed about results in a Benchmark you know nothing about? Can anyone shed some light on Skidmarks? The numbers alone mean nothing. Does it accurately test Dual Processor configs?

Optimization

Was the Benchmark recompiled or even optimized for the G5?
If not, note that the Altivec code in the Fractal benchmark was handcoded with respect to the specific instruction scheduling characteristics of the G4 processor.
I wouldn't be suprised if a G5-optimized version lead to significantly better results.

Skidmarks Background

Originally posted by nuckinfutz
How can you be disappointed about results in a Benchmark you know nothing about? Can anyone shed some light on Skidmarks? The numbers alone mean nothing. Does it accurately test Dual Processor configs?

Click to expand...

I'm working on that

Uh, according to some mumblings <A HREF="http://www.xlr8yourmac.com/OSX/os_x_benchmark_tools.html">here</A>, Skidmarks supported only single processor tests, at least as of March 2002. The Apple Architecture and Performance Group claimed it was a processor benchmark with little system-level interaction and had a bug which occasionally posted extremely high scores.

I'm going to try and dig up some more recent informatoin, but in the meantime, I hope tihs is a good start.

BTW, if the single-processor bit is true, that means that the G5 is somewhere around twice as fast as the G4, at least in the Skidmarks test, if the numbers posted above were correct

Translation from single to dual results

Are we overlooking the fact that a dual machine isn't going to have double the benchmark score of a single machine with the same speed processor? There is going to be some loss in processing power over 2 chips right?

If a single 1 Ghz G4 gets a benchmark of 100, a dual 1 Ghz G4 machine might only get 170 not necessarily 200. Am I on the right track? This is the case in the x86 world.

So we can't really judge the dual 2.0 G5 unless we have a dual G4 benchmark to go by. Anyone have a newish dual G4, 1 Ghz or above to run the test on?

My 1 Ghz TiBook gets 3000-3500 Mflops when I run the test on it...

-Brian

Re: Translation from single to dual results

Originally posted by Ulmer
Are we overlooking the fact that a dual machine isn't going to have double the benchmark score of a single machine with the same speed processor? There is going to be some loss in processing power over 2 chips right?

If a single 1 Ghz G4 gets a benchmark of 100, a dual 1 Ghz G4 machine might only get 170 not necessarily 200. Am I on the right track? This is the case in the x86 world.

So we can't really judge the dual 2.0 G5 unless we have a dual G4 benchmark to go by. Anyone have a newish dual G4, 1 Ghz or above to run the test on?

My 1 Ghz TiBook gets 3000-3500 Mflops when I run the test on it...

-Brian

Click to expand...

Yes, you're right about the loss of processing power, although the G5 is supposed to be much more efficient in dual-processor configurations than the G4. Jobs said this at WWDC; I'm also going on information from IBM (can't remember particular links from them, but it might be mentioned in the PowerPC 970 reports over at http://www.arstechnica.com).

Re: Translation from single to dual results

Originally posted by Ulmer
Are we overlooking the fact that a dual machine isn't going to have double the benchmark score of a single machine with the same speed processor? There is going to be some loss in processing power over 2 chips right?

If a single 1 Ghz G4 gets a benchmark of 100, a dual 1 Ghz G4 machine might only get 170 not necessarily 200. Am I on the right track? This is the case in the x86 world.

So we can't really judge the dual 2.0 G5 unless we have a dual G4 benchmark to go by. Anyone have a newish dual G4, 1 Ghz or above to run the test on?

My 1 Ghz TiBook gets 3000-3500 Mflops when I run the test on it...

-Brian

Click to expand...

Yes you are correct in your assumptions.....but what I find disconcerting is that according to the pre-970 hype these chips were supposed to be orders of magnitude superior to the G4 and that now seems to be a pipe dream...

we wont know for sure until the real machines are in the hands of the people and then we can run some benches, i guess this means sept2 until we can do this. i wish apple would have shown us how UT2003 runs on a 1.6, 1.8 and duallies.

Skidmarks Background

Originally posted by nuckinfutz
How can you be disappointed about results in a Benchmark you know nothing about? Can anyone shed some light on Skidmarks? The numbers alone mean nothing. Does it accurately test Dual Processor configs?

Click to expand...

I'm working on that

Uh, according to some mumblings <A HREF="http://www.xlr8yourmac.com/OSX/os_x_benchmark_tools.html">here</A>, Skidmarks supported only single processor tests, at least as of March 2002. The Apple Architecture and Performance Group claimed it was a processor benchmark with little system-level interaction and had a bug which occasionally posted extremely high scores.

I'm going to try and dig up some more recent informatoin, but in the meantime, I hope tihs is a good start.

BTW, if the single-processor bit is true, that means that the G5 is somewhere around twice as fast as the G4, at least in the Skidmarks test, if the numbers posted above were correct

Dual 1 gig - I score about 5500

i'm getting about 5500 on a dual 1 gig with 512 ram - also have lots of apps open - dont know if that matters

Re: Re: Translation from single to dual results

Originally posted by jcdenton
Yes, you're right about the loss of processing power, although the G5 is supposed to be much more efficient in dual-processor configurations than the G4.

Click to expand...

The G5 benefits from not having to share a single FSB between two chips. Otherwise, it's still subject to the rules of SMP scaling which usually state roughly 80% as the typical performance increase when adding a second proc.

Which makes a dual 2Ghz G5 roughly equivilent to a 3.6 Ghz G5 when things are SMP aware. (How the dual having 2Ghz of bandwidth between two procs compared to the 1.8 Ghz a 3.6 would have affects overal performance between the two setups, I'd have to leave to someone else to sort out.)

XBENCH - I get a 115 score

hmmm not quite as smoking as i hoped

On the other hand... I just read this on Accelerate Your Mac which augurs well:

"Hey everyone, I've been at Apple's developer conference and had a chance to install and try out After Effects on a new G5.
I ran the Night Flight file that has come to be the standard for AE benchmarking. Since I didn't want to sit there and watch it render for hours, I ran just the first 10 interlaced frames from the project's pre-set render queue...
http://www.aefreemart.com/tutorials/3DinAE/nightflight/nightflight.html
Here are my results for this test on the three computers I have available to me:

1 x 1.0 GHz G4 PowerBook 17" - ~30 minutes (3 min/frame)
2 x 2.66 GHz Pentium Xeon from Boxx - 11 min, 39 sec (1.2 min/frame)
2 x 2.0 GHz PowerMac G5 - 6 min, 1 sec (0.6 min/frame)

I ran the Xeon test on a couple different identical machines to make sure mine wasn't just running slowly, but got identical results.
Of course my Mac bias is well-documented, but I'm sure many people here can vouch for me as an honest person. If the results had gone the other way, I'd just keep my mouth shut and let someone else break the bad news.
Other observations about this test that may ultimately work in the Mac's favor:

1) The machine was not running 64-bit Panther, but only a tweaked version of 32-bit Jaguar. Likewise, AE is obviously not yet compiled to take advantage of the G5 chip in any way. Both or these situations will automatically be rectified in the future.

2) Night Flight is very CPU-intensive, but not very disk I/O intensive. I think the 1 GHz system bus and other details on the G5 will provide greater gains for typical projects that rely more heavily on I/O."

What did you expect?

Originally posted by hvfsl
Maybe the PPC970 does not run 32bit code as good as Apple wants us to believe.

Click to expand...

Of course not. Nothing about the G5 will be as good as Apple wants us to believe.

Steve Jobs is the king of hyperbole. He tells a great fish-story and is wonderful at getting the troops amped, but you'd have to have a serious long-term memory problem to take his statements at face value.

That's not to say it isn't a welcome, but long-coming upgrade. Of course it is, but it's not going to cook you waffles, mow your lawn and save your marriage like the RDF marketers would have you believe.

Re: Re: Translation from single to dual results

Originally posted by fred
.....but what I find disconcerting is that according to the pre-970 hype these chips were supposed to be orders of magnitude superior to the G4...

Click to expand...

As I recall, from the bench marks released by IBM the 970 was expected to be 2 to 3 times faster than a G4 for some tasks (FP?), not orders of magnitudes.

Also, performance generally doesn't scale linearly with clock speed. Look at these Xbench scores.

Dual 1GHz G4 about 100
Dual 533MHz G4 about 75

We may need a new version of xbench to get accurate scores for the G5.

Re: Re: Re: Translation from single to dual results

Originally posted by Dave K
The G5 benefits from not having to share a single FSB between two chips... Which makes a dual 2Ghz G5 roughly equivilent to a 3.6 Ghz G5 when things are SMP aware. (How the dual having 2Ghz of bandwidth between two procs compared to the 1.8 Ghz a 3.6 would have affects overal performance between the two setups, I'd have to leave to someone else to sort out.)

Click to expand...

Thanks for the clarification, Dave.

Originally posted by fred
Well I submitted a story link to Macrumors yesterday which showed the same sorta disappointing linear type performance gains (in fact even worse) . The story is on Think Secret and the relevant quote is:


Another screenshot shows results from Skidmarks GT, a benchmarking utility that's part of Apple's CHUD performance tools. Based on the test we ran, this provides a rough comparison of G5 and G4 performance. The Skidmarks scale has "100" equal a Power Mac G4 at 1GHz. The Dual 2GHz received scores of 172 for integer performance, 270 for floating point performance, and 208 for vector performance.



http://www.thinksecret.com/news/wwdc03g5.html

Click to expand...

Would I be correct, or incorrect, to say that these tests do not utilize the hard drives or memory bandwidth. In that sense, they are pretty poor comparisons. The bottom line is that the architecture of these new machines should prvide a more promosing future for rapid GHz increase, and an overall speed improvement.

I'm not shocked by the scores at all.

Not bad

Back to original topic, the G5 has been scaling more than linearly.

(Assuming dual processors to be 1.8 times the performance of single processors.)

2000MHz/450MHz= 4.444

1500MFlops x 4.444=6666.666 Theoretical 2GHz rate if the CPU scaled linearly

6666.666 x 1.8 = 12000 Theoretical Dual 2GHz linear rate

If the Dual 2GHz G5 scored 14055.2, that's not so shabby.

And that's comparing to the G4/450 which didn't have a system bus bottleneck. Later G4 models had some serious bandwidth problems. Someone wrote that their Dual Gig G4 scored 5500. The Dual 2GHz G5 almost tripled that number despite only being double the clock rate.

BTW, I remember reading somewhere that After Effects is not dual processor aware and not altivec enhanced.