Welcome to the MacNN Forums.

Abbas · Jul 22, 2006, 11:37 AM

This should give a little idea on the performance of upcoming Conroe based Macs:

http://www.tbreak.com/reviews/article.php?id=461

mduell · Jul 22, 2006, 02:27 PM

Which Macs are going to be based on Conroe?

Originally Posted by The Article

Steve Jobs promised a 3.0GHz PowerPC based Mac long before Intel released their 3.0GHz Intel Pentium 4 CPU.

Does the author mean SJobs made the "3Ghz in a year" claim before Intel released a 3Ghz P4, or that SJobs' claim was that a PPC chip would hit 3Ghz before the P4 hit 3Ghz?
Either way, the author is completely wrong.
The 3Ghz G5 claim was made on June 23, 2003.
The P4 hit 3Ghz on May 21, 2003 and even 3.2Ghz on June 23, 2003.

Catfish_Man · Jul 22, 2006, 02:52 PM

That Rosetta benchmark makes me question the accuracy of the rest. 3% difference indicates one of three things: a) their benchmark was of a task that amounts to a single tight loop, b) the Windows version of photoshop is horribly slow, or c) the testing methodology is flawed. There's just no way Rosetta is only a 3% speed hit.

mduell · Jul 22, 2006, 02:55 PM

Originally Posted by Catfish_Man

That Rosetta benchmark makes me question the accuracy of the rest. 3% difference indicates one of three things: a) their benchmark was of a task that amounts to a single tight loop, b) the Windows version of photoshop is horribly slow, or c) the testing methodology is flawed. There's just no way Rosetta is only a 3% speed hit.

Their Photoshop benchmark could be so poorly designed that it is disk-bound instead of CPU bound.

Abbas · Jul 22, 2006, 03:48 PM

Thanks for the comments. I've corrected the sentence of G5/3.0Ghz before the P4 3.0GHz release.

Regarding the Photoshop bench, I've used the image from this archive and ran the Radial Blur filter at 100% using Best settings. The score in Windows seems correct as its slightly lower in seconds over the E6700 and Athlon FX-62 CPUs that I've tested.

Even I was astonished by the performance increase but you're more than welcome to download it and see if its HD bound so I can correct those results. Do keep in mind that I've used 2GB RAM testbenches on all the configurations.

Lateralus · Jul 22, 2006, 04:49 PM

Don't you find citing the lack of delivery on a 3GHz G5 as the reason for the switch to Intel at least a little bit over simplified and uninformed? After all, Intel's new architecture isn't even at 3GHz yet.

I myself have never understood why so many people are fixated on the lack of a 3GHz G5, especially given the other improvements IBM made to the G5 that more than make up for their inability/lack-of-desire to reach 3GHz.

So; Why do so many people think it matters? Certainly not because anybody thinks that the additional 300MHz on top of IBM's achieved 2.7GHz would have had any truly perceivable effect on the G5's performance. It matters because it was one indicator that IBM's commitment to PowerPC on the desktop was and is uncertain beyond the immediate future. It's as simple as that. The driving force behind Apple, AMD and Intel is not MHz.

mduell · Jul 22, 2006, 05:16 PM

Originally Posted by Lateralus

Don't you find citing the lack of delivery on a 3GHz G5 as the reason for the switch to Intel at least a little bit over simplified and uninformed? After all, Intel's new architecture isn't even at 3GHz yet.

Xeon 5160 (an Intel Core Microarchitecture part) launched at 3Ghz nearly a month ago. You can buy one today for about $1000.

Originally Posted by Lateralus

I myself have never understood why so many people are fixated on the lack of a 3GHz G5, especially given the other improvements IBM made to the G5 that more than make up for their inability/lack-of-desire to reach 3GHz.

I agree. Change 3 to 4 and Intel had the same problem with the P4; they promised 4Ghz, only got to ~3.93Ghz, and it was much later than initially stated.

Catfish_Man · Jul 22, 2006, 06:00 PM

All the manufacturers did, really. The 90nm node just turned out to be hellishly difficult. I was fairly astonished at how smoothly 65nm went*... I guess they readjusted their focus.

*not counting Prescott here. That design was a train wreck for fairly clear non-manufacturing reasons.

ScottEllsworth · Jul 24, 2006, 01:51 AM

Originally Posted by Lateralus

Don't you find citing the lack of delivery on a 3GHz G5 as the reason for the switch to Intel at least a little bit over simplified and uninformed? After all, Intel's new architecture isn't even at 3GHz yet.

While the game is not all about GHz, clock rate is a pretty good proxy for a lot of things. Both Intel and IBM have had trouble hitting performance targets, but Intel had something on deck (Core/Core2) that has a fair amount of legs to it. As best as I can tell, IBM kinda tanked with the G5, and never managed to produce a low power version appropriate for a laptop.

The top of the line PPC laptop was a 1.6 GHz G4, remember, which was getting its head handed to it by even Pentium M systems, let alone what the Core Duo/Core 2 designs were likely to do. Intel was not the only choice, but the list of good laptop chip choices really did not include much from IBM.

On the desktop side, the same could be said. A G5/2.7 competes pretty well with a 3Ghz pentium, but the roadmap were just not there. IBMs failure to deliver a 3Ghz part would not have been a catastrophe had everyone tanked, but Intel had new architectures on the drawing board that could make up for the 30% ht going from altivec to SSE3. Not a good thing, and thus at least no loss would take place switching architectures.

Finally, there are more people tweaking the Intel architectures for speed. The Java server VM runs an order of magnitude (yes, 10x) faster on my MBP than on the powerbook that replaced it, and it runs four times faster than anything my g5 could do. Not because of any special cookies of the Core Duo, but because Sun put a lot of time into optimizing the compiler.

So, given that laptops were an ever increasing fraction of the market for Apple, and that IBM had nothing good in the pipeline that could compete, the move makes sense on that front alone. Add in that the desktop market did not look joyous for IBM, and I think the motives for the change seem clear.

Sure, saying 'it was because IBM did not deliver 3Ghz' is a simplification, but it is a reasonable one.

Scott

Simon · Jul 24, 2006, 02:47 AM

I think the bitching about the G5 not reaching 3GHz wasn't inappropriate.

It's not actually 3GHz that matters, it's about over promising and under delivering. When you get up in front of your customers and promise 3GHz within a year and then never deliver you deserve to take some flak. Promises are promises - it's not like business promises don't count. Breaking promises kills trust - it's plain and simple.

If you're not sure about being able to deliver, don't make premature promises - especially not when you're being such a pompous ass about it (like Steve likes to be). If you promise something and then fail to deliver, you'll look like an idiot and that's perfectly justified.

That all said, Apple didn't switch to Intel because IBM didn't reach 3 GHz. That's simplistic and naive.

harrisjamieh · Jul 24, 2006, 04:18 AM

I swear the first post in this thread had NOTHING to do with the whole 3ghz crap... it was just about Conroe benchmarks...

Mamochanx · 08:59 AM

Can we have verification of this test and get back on topic?

Abbas could you bench this test in photoshop?

http://www.retouchartists.com/pages/speedtest.html???

Abbas · Jul 24, 2006, 01:49 PM

Sure thing- I'll post the results of this tomorrow. I can also take a video in case somebody doubts these results

-a

Originally Posted by Mamochanx

Can we have verification of theis test and get back on topic?

Abbas could you bench this test in photoshop?

http://www.retouchartists.com/pages/speedtest.html???

Abbas · Jul 25, 2006, 06:29 AM

OK- the above test took 2:24 (2 minutes and 24 seconds) in total to complete. Not sure how good or bad that is as I'm using this test for the first time but let me search for it on these forums to see how good of a score it is.

-a

Abbas · Jul 25, 2006, 06:49 AM

That same test in Windows using indentical hardware too 3:37 (3 minutes 37 seconds) to complete which really makes me wonder whats going on. Cinebench under Windows is producing almost identical scores to Mac on this setup so it cant be a matter of Windows behaving badly.

-a

Mamochanx · Jul 25, 2006, 08:42 AM

Abbas if this is accurate it is amazing. It took a 20inch Intel Imac with 1.5Gb of ram 4:09 to complete that test.

I dont know what to say really. Unless OS X for intel is using the extra instructions that Core Duo 2 tech has over older chipsets I can not think of a reason why Photoshop under Rosetta would have such high scores.

I am really shocked and trust that the benches are acurate. Could you post a CPU log under both Windows and OSX while testing?

Abbas · 09:04 AM

Originally Posted by Mamochanx

Could you post a CPU log under both Windows and OSX while testing?

Please tell me how to do this and I will have it done. As I mentioned earlier, I can take a video of this thing in action as well.

-a

Mamochanx · 09:19 AM

If you want to post a vid that would be, can you post it with activty monitor open so that we can see how much the cpu is in idle?

Thanks again for the test Abbas.

I know Rosetta takes advanate of SSE instructions, but those results are insane. Here is some info on the new SSE instruction performance on the new chips.

http://www.extremetech.com/article2/...1935731,00.asp

That is the only reason I can see how those results can happened.

Abbas · Jul 25, 2006, 11:44 AM

Originally Posted by Mamochanx

That is the only reason I can see how those results can happened.

Also keep in mind that these Conroes have 4BM cache which could be a factor is adding to Rosetta's performance.

-a

Catfish_Man · Jul 25, 2006, 12:35 PM

Originally Posted by Abbas

Also keep in mind that these Conroes have 4BM cache which could be a factor is adding to Rosetta's performance.

-a

That's a very interesting point. We may be seeing a situation much like the 601->603 transition in reverse, where the half-size L1 cache on the 603 destroyed 68k emulation performance, because the working set of the emulator was bigger than 8k but smaller than 16k.

In theory it is possible for a dynamic recompiling emulator to have higher performance than native code... but the only case I've heard of one doing so is HP's dynamo "emulator".

Abbas · Jul 26, 2006, 08:34 AM

Alrighty- video is uploaded and can be downloaded from the following URL:

http://www.tbreak.com/files/conroe.zip

Its a little over 50MB in size. Enjoy!

-a

Mamochanx · Jul 26, 2006, 12:44 PM

Thanks a lot Abbas.

Abbas · Jul 26, 2006, 03:10 PM

No worries- Just wanna get the expert's opinion on what to make out this incredible Photoshop performance under rosetta.

-a

shawmanus · Jul 26, 2006, 10:01 PM

Originally Posted by mduell

Which Macs are going to be based on Conroe?

Does the author mean SJobs made the "3Ghz in a year" claim before Intel released a 3Ghz P4, or that SJobs' claim was that a PPC chip would hit 3Ghz before the P4 hit 3Ghz?
Either way, the author is completely wrong.
The 3Ghz G5 claim was made on June 23, 2003.
The P4 hit 3Ghz on May 21, 2003 and even 3.2Ghz on June 23, 2003.

Intel released Northwood B 3.06ghz with 533mhz fsb on Nov 14 2002.

http://techreport.com/reviews/2002q4...6/index.x?pg=1

michaelb · Jul 27, 2006, 03:42 AM

Here's a wild theory:

Rosetta emulates Altivec instructions, and Photoshop is intensely optimized for Altivec.

Now, the new Core 2 architecture is meant to have many SSE enhancements, I think up to 16 new opcodes have been added to SSE3 (or is it SSE4 now?).

It is just possible that Rosetta maps these to Altivec instructions in an almost 1:1 fashion.

Which means, under Rosetta, Altivec instructions may be simply passed through to their equivalent Intel SSE instructions, resulting in phenomenal performance when running Altivec optimized PowerPC software on a Core 2 chip.

Just an idea!

Catfish_Man · Jul 27, 2006, 05:06 AM

The biggest improvement for SIMD performance in Conroe is not the new opcodes, but the massively increased resources for the existing opcodes. Instead of one 2x32 bit op every cycle (so two cycles for each 4x32 vector), it can do multiple 4x32 ones, like the G4 and G5 can.

Eug Wanker · Jul 27, 2006, 10:45 AM

For comparison's sake, this is what I got with the www.retouchartists.com Photoshop Actions test comparing the G4, G5, and Yonah Core Duo.

ajprice · Jul 27, 2006, 12:26 PM

Just to straighten things out with that graph, a MacBook running Photoshop is going to be running it through Rosetta, so thats why it's slower than a G5. Once the new Photoshop comes, that graph will look different.

Simon · Jul 27, 2006, 01:00 PM

Originally Posted by ajprice

Once the new Photoshop comes, that graph will look different.

Actually it could be even earlier. Preliminary tests with Conroe show that it's improved instruction set and larger L2 cache boost Rosetta performance by quite a bit. Maybe the wait for CS3 will become less painful soon.

Eug Wanker · Jul 27, 2006, 01:45 PM

Originally Posted by ajprice

Just to straighten things out with that graph, a MacBook running Photoshop is going to be running it through Rosetta, so thats why it's slower than a G5. Once the new Photoshop comes, that graph will look different.

Yeah. I was just offering that up as a baseline, since everyone was talking about Rosetta.

P.S. Photoshop CS3 ain't out until well into 2007.

mduell · Jul 27, 2006, 07:37 PM

Originally Posted by Catfish_Man

The biggest improvement for SIMD performance in Conroe is not the new opcodes, but the massively increased resources for the existing opcodes. Instead of one 2x32 bit op every cycle (so two cycles for each 4x32 vector), it can do multiple 4x32 ones, like the G4 and G5 can.

Doesn't SSE also accept 2x64 instead of 4x32, unlike Altivec on the G4/G5 (4x32 only)?

/slightly offtopic

24klogos · Jul 28, 2006, 08:58 PM

mr duell, you're on fire. Whats you actual oppinion on the performance of upcoming Conroe based Macs? if there was software out there ready.

Catfish_Man · Jul 28, 2006, 10:49 PM

Originally Posted by mduell

Doesn't SSE also accept 2x64 instead of 4x32, unlike Altivec on the G4/G5 (4x32 only)?

/slightly offtopic

Correct. It'll do that in two cycles* though on pre Core 2 chips.

*rather, throughput will be one every two cycles. Latency is higher than that.

Simon · Jul 29, 2006, 01:53 AM

Originally Posted by 24klogos

Whats you actual oppinion on the performance of upcoming Conroe based Macs? if there was software out there ready.

For native CPU bound software I'd expect 20-30% at equal clock rate. For Rosetta emulated apps there's a good chance you'll gain more due to the improved instruction set and larger L2 cache.

mduell · Jul 29, 2006, 11:22 PM

Originally Posted by 24klogos

mr duell, you're on fire. Whats you actual oppinion on the performance of upcoming Conroe based Macs? if there was software out there ready.

Just like the Conroe based Windows boxes: smoking everything else out there. 25% CPU performance (and maybe memory performance if they upgrade to DDR2-800/PC2-6400) boost for an iMac at the same clockrate.