PowerMac G5 1.6GHz PSBench 7 Results

[Analog Kid]

Originally posted by Cubeboy
Macrumors12345:
Looking back, I see your point, so I'm going to do another comparison which compares the CPUs in terms of how they perform in each filter.

Wow, thanks for the analysis! That was a welcome sight!

I'd still really like to see a comparison of those results with the results of the G5 w/o plugin to see which of those are optimized.

Anyone care to speculate what causes each unit to perform so much better on certain tests?

Rotate 90 is largely a memory op so maybe the caching on the P4 helped it there? It can't fit the whole image in cache, but if the filter is designed well it could exploit it pretty well.

Rotate .9 is largely a floating point op, so the G5 caught up? We know the P4 FPU is weak.

I'd expect these to have been optimized.

What's up with the divergence in the gausian blur numbers? Does the larger radius blow the G5 L2 cache?

RGB-CMYK sounds floating point with a small memory footprint (almost a register op)-- advantage G5?

Lens flair also sounds like a memory user-- advantage P4?

I know so very little about how most of the operations work, that all I can hope to do is spark someone elses interest and get real information.

I am curious though whether the lack of an L3 cache hurts the G5, even with the fat pipe to main memory.

The duals might help a bit here since they can snoop each other's L2 caches.

For the case of upoptimized ops, it sounds like there are common G4 Altivec ops that cause the G5 to stick (unfortunate artifact of switching chip vendors and a potential problem for compatibility)-- so that might explain why some operations seem to give such a huge advantage to the P4.

I think any op that spun up the disk is invalid, but we don't know which those were...

Is there anyone out there who has designed filters like these?

 

[eatme8888]

Maybe it will, maybe it won't.

This benchmark shows the dual 1.8 Opteron not doing too well:
http://www6.tomshardware.com/cpu/20030422/opteron-23.html#3drendering

We need to know what kind of performance increase you get for each Mhz in the Opteron before we can speculate how much faster a dual 2 Ghz Opteron would be. For example, the Xeon is at a point where it gets very, very, very little actual performance increase with every step in Mhz. The dual 3.2 Ghz Xeon will be slightly faster than the dual 3.06. The difference is so small, people who know better are just going to save their money and by the less expensive processors.

Right now, there is a 4 gig limit on the BOXX dual Opteron, because it comes with Windows XP Pro. If/when MS comes out with a special Windows for the 64 bit aspect of those AMD processors, then it can have more memory. The only 64 bit Windows available today starts out at $2,999, just for the OS.

And when you start comparing the prices of these systems, the G5 is the least expensive.

I say wait on pronouncing a speed king until the G5s are shipping with Panther and the Opteron is out with an optimized Windows. The Xeon is dead. It might make it to 5.1 Ghz at the end of the Prescott's run, yet it will have to be 5.1 Ghz this time next year to keep up with the 3 Ghz IBM 980 that is being prototyped today, and proclaimed to be shipping next year by Steve Jobs and IBM last June. The G5 has legs, folks, and in a year, there won't be a need to nitpick .2 seconds for gaussian blur and .3 seconds for rotate...

The IBM roadmap has the G5 going to 10 Ghz by 2006. Damn, Longhorn is going to be out in 2005/6. We'll have two more updates of OS X by then, and a G5 that's between 7-10 Ghz.

The Opteron is very nice, for an X86 processor, but it's still X86 and weighted down with all that baggage that comes with X86.

Resistance is futile.

 

[jamesbsd]

Originally posted by adamfilip


...
anyways..

remember when the first p4 came out.. at 1.4ghz.. people were cryling and complainging that the p3 at 1ghz was faster.. and in many respects it was.. the p4 was superscaler.. it can reach higher clock speeds. and cant do as much per clock speed.. but by being able to reach very high clock speeds it still ending up faster then the p3.. same thing is happening with the g5

if you were able to take a single g4 1.4 and a g5 1.4 the g4 would prob win.

Maybe if you crippled the 970 (G5) with a 167 MHz bus and were running a benchmark using primarily integer or vector code. In that case a G4 with 2 MB of L3 should outperform the G5. But the 970 isn't designed to run on a configuration like that, so the point is a moot one.

Apple's tech note 2087 should clear a few things up:

http://developer.apple.com/technotes/tn/tn2087.html

 

[Cubeboy]

Regarding varying results:

The tester ran each filter three times and got similar results each time, the numbers you see on the chaosmint site are actually the averages of the three times.

 

[acj]

Do the test right, read the results right

For anyone who's read this deep (not me, so sorry if this has been posted) here are a couple trhings worth mentioning:

For one, the biggest factor in this test is how many history states you have. It should be set to 1. Ignore all these crazy results being submitted about how much faster it is than a DP 1.42 MDD, etc.

Second, this test is actually better than Apples because even though it only runs 15 filters, you can actually read the results for each one. One filter can offset the entire results, so total time is meaningless for most professionals who only use a handfull of filters.

Third, sorry but the dual 2.0 is not Apple's fastest machine. It's not even available. Make comparisons there in 3 weeks.

 

[acj]

Still meaningless

Despite my previous post, this benchmark too is fairly meaningless.

I want a computer that will do things fast like paint with a 2500 pixel soft brush in an adjustmentlayer mask, run a batch command on 80GB of images before I start work again in the morning. These are some of the things I do.