IBM and Motorola Diverging

[alex_ant]

Re: Why not using IBM POWER

Originally posted by sunbear191
Is there any reason why Apple could not use IBM's POWER3 or POWER4 cpus? They are compatible (except for AltiVec), aren't they?

I was under the impression that IBM is not keen on letting other companies use the cream of their crop. Also, these chips are very expensive and complex. Because you can't buy these chips off the shelf like you can an Intel or AMD chip, it's difficult to say exactly what they cost, but just to give a bit of perspective, the Power4 is not currently sold in any system costing less than $40,000. The least expensive Power3 system can be had for just under $6,000. There are heat and power consumption issues as well... but hoo boy, would I ever want a Power4 on my desk! F*cking wicked!

Apple has to have something big up their sleeves. They just have to. THEY HAVE TO!!! (*Bursts into tears*) OHHHH GOD, PLEASE TELL ME THEY DOOOO!!!!!!!

Alex

 

[Rower_CPU]

Originally posted by alex_ant

The Mac's only speed advantages in Photoshop are due to AltiVec, which makes it kind of an unfair comparison.

How about SPEC_CPU2000, which would make a good approximation of Photoshop performance if not for AltiVec, which skews everything...

How does AltiVec make the comparison unfair? If a chip has performance enhancing instructions, why is it "unfair" to use them? That's like saying "The P4 has a quad-pumped memory bus, so that skews everything". If a chip has advantages over another they should be taken into account.

SPEC has long since been discounted as a completely unfair benchmark for the PPC platform. Here's a thread that discussed it a while ago:
https://forums.macrumors.com/showthread.php3?threadid=2845

It simply floods the CPU with perfect instructions, not taking into account dependancy issues or the "pipeline" tax". NOT a good indicator of overall system performance.

 

[TechLarry]

Re: Case closed.

Originally posted by gjohns01
I could still build the same machine from scratch for less than $1000. I could have paid an extra $1000 grand to get a server motherboard, cpus; SCSI drives and built a "real" server. I still would have come out cheaper than buying a brand new Apple G4 Tower/Server.

As for the OS. I paid for the MSDN Universal Subscription. Do you know what that is? For $2800 bucks I own every Microsoft App/OS/SDK/Dev Tool they make. Do I receive the same resources from Apple? LOL. Right!

Have I set up an IIS server before absolutely. Is my box secure? Absolutely! Can I run and Oracle database? (and Ellison is on the damn board) DB2? Sybase? Exchange server? Websphere? MQSeries? Use a Java IDE that doesn't suffer from redraw problems because the OS isn't completely finished? NO!

I can do all of that on my PC. I can run all of those applications in the comfort of my home so that I don't have to go to work everyday. Do I give a damn about streaming video? Using Photoshop? Do I care if I can arrange all of my photos nicely and order a book? Care if my machine runs hot? Hell NO!

What's your point? That you don't care about the enterprise because that's not what you do for a living? Well I do. And Apple isn't playing on the same level. Case closed.

Just curious as to who this response was meant for...

TL

 

[TechLarry]

Re: Re: Re: Case closed.

Originally posted by alex_ant

Actually I think he's got a point...

$100 Tyan Tiger dual P3 motherboard
$100 Full-tower ATX case w/ 500 watt power supply
$250 Two 1.2GHz Pentium IIIs ($125 each)
$250 1GB DDR SDRAM
$50 cheapo video & sound cards (this is a server after all)
$50 32X SCSI CD-ROM
$560 Two 50GB Ultra160 SCSI hard drives ($280 each)
$100 Ultra160 SCSI controller
$80 Dual gigabit Ethernet controllers ($40 each)
$20 Cheapo keyboard/mouse, if needed
$0 OS (Linux or one of the BSDs)
--------------------------------------
<$1600 for a machine that will match or beat a $4500 Power Mac server. These are not no-name parts, either (well, except for the keyboard, mouse, video, and sound cards, but none of those are important in a server). Add a monitor and a few other misc. parts and this thing is still less than half the cost of the Mac. Sorry but it's true.

Alex

LOL!

First, it was $500. Then somewhere along the line it changed to "Less than $1000".

Now, it's "<1600".

I just love this stuff

TL

 

[TechLarry]

Re: Re: Re: Re: Case closed.

Originally posted by TechLarry


LOL!

First, it was $500. Then somewhere along the line it changed to "Less than $1000".

Now, it's "<1600".

I just love this stuff

TL

Oh, since you need server software, let's throw in Server Software for, oh, 1000 users.

And to keep it interesting, let's keep it commercial

The Mac server will obviously use MacOS X Server.

Just for fun

TL

 

[TechLarry]

Originally posted by gjohns01


I'm sure I can put together a single processor AMD system that will clean a dual G4's clock on pretty much any test you want to offer up. It's the sad truth buddy. They picked the wrong chip partners and haven't moved their motherboard to new technology.

I have to agree with the last sentence.

I believe I started all this fun nonsense with something similar

TL

 

[Rower_CPU]

Originally posted by gjohns01
... Macs offer a consumer machine, and a pro machine. They offer "servers" but they are not what I consider a server to be. The fact of the matter is that there is more than one kind of Pro. Apple's definition of Pro is centered around the creative types. They failed to address other needs. So by all means use OS X and OS X Server. I'm just letting you know that as a product (hardware and software) Apple is seriously deficient in some areas..

Once again, what do your servers do that OS X doesn't? How are they not "server" enough for you?

I'm fully aware of Apple's "deficiencies", but for me the pros still outweigh the cons. I'm not missing playing Starcraft or any of that crap. If there's something that you absolutely have to run, get VPC. Can your Wintel do that?

I use Photoshop, Premiere and Final Cut all of the time. If I'm on a slow machine, the process is painful. If I'm on a fast one, it's effortless. Therefore, they are a good determinant of system performance for my needs.
I'm sorry you have less creative things to do. What do you do with your machine?


BTW - You have still failed to address the ROI question, and you're now ignoring my (quite fair) challenge...why?

 

[gjohns01]

Originally posted by Rower_CPU


Once again, what do your servers do that OS X doesn't? How are they not "server" enough for you?

I'm fully aware of Apple's "deficiencies", but for me the pros still outweigh the cons. I'm not missing playing Starcraft or any of that crap. If there's something that you absolutely have to run, get VPC. Can your Wintel do that?

I use Photoshop, Premiere and Final Cut all of the time. If I'm on a slow machine, the process is painful. If I'm on a fast one, it's effortless. Therefore, they are a good determinant of system performance for my needs.
I'm sorry you have less creative things to do. What do you do with your machine?


BTW - You have still failed to address the ROI question, and you're now ignoring my (quite fair) challenge...why?

Actually you've completely ignored me. I've stated what I need to do. Read through the posts. But just to speed things up here I'll repeat myself. There are no enterprise databases available on OS X. (i.e. No 8i, No DB2, No Sybase, No SQL Server) Using mySQL, Postgresql, Frontbase, or Openbase won't cut it. I can't develop to those we don't use them. There is no official J2EE release for Mac OS X. And yes I am aware that you can use one from a JBoss distribution or some other source. There is no JBuilder for Enterprise available for OS X. There is no Visual Studio. There is no Visual Age. No Websphere, No IPlanet, No Vignette, or any other application I deal with in my everyday life. Therefore, OS X Server and hardware have ZERO to offer me. I'm sure everyone's favorite board member Larry Ellison feels the same way, because he hasn't ported 8i to OS X. He has no reason to. Apple doesn't have the hardware to run it and doesn't command any respect in the enterprise for Oracle to make the effort. Now go design something. And will you're at it, tell me how Photoshop, Premiere and Final Cut can tell me how many transactions per minute I can achieve on my hardware.

 

[alex_ant]

Originally posted by Rower_CPU
How does AltiVec make the comparison unfair? If a chip has performance enhancing instructions, why is it "unfair" to use them? That's like saying "The P4 has a quad-pumped memory bus, so that skews everything". If a chip has advantages over another they should be taken into account.

Unfair was the wrong word, I suppose. The reason I said unfair was because, at the moment, AltiVec 1) requires a fair amount of hand-tuning of code in order to be taken advantage of, and 2) is taken advantage of by only a few apps. This is atypical for a CPU. Any good CPU is not supposed to require tedious optimization; you're supposed to be able to compile your code for it, get great performance off the bat, and be done. If/when there ever exists a decent compiler for AltiVec-enabled chips, then I will stand corrected.

SPEC has long since been discounted as a completely unfair benchmark for the PPC platform. Here's a thread that discussed it a while ago:
https://forums.macrumors.com/showthread.php3?threadid=2845

It simply floods the CPU with perfect instructions, not taking into account dependancy issues or the "pipeline" tax". NOT a good indicator of overall system performance.

I've read that thread. All I can say is, SPEC is the most widely accepted CPU benchmarking test there is. Its so-called pipeline tax doesn't seem to affect any other CPUs. You can whine all you want about unfairness, but that's to be expected from anyone getting butt-raped like Motorola's PPCs are.

If only software that has been hand-optimized for AltiVec can squeeze a decent amount of speed out of the G4, then that makes the G4 a speciality chip. Nobody wants to optimize for AltiVec because doing so results in completely non-portable code.

If it looks slow... and smells slow... it probably is.

Alex

 

[alex_ant]

Re: Re: Re: Re: Case closed.

Originally posted by TechLarry


LOL!

First, it was $500. Then somewhere along the line it changed to "Less than $1000".

Now, it's "<1600".

I never claimed I could build a server for $500 or $1000. What I proved was that it's possible to build a server faster than the fastest Mac server for less than half the price. Not taking into account service, support, reliability, etc... but it can be done.

Alex

 

[Rower_CPU]

Originally posted by alex_ant

Unfair was the wrong word, I suppose. The reason I said unfair was because, at the moment, AltiVec 1) requires a fair amount of hand-tuning of code in order to be taken advantage of, and 2) is taken advantage of by only a few apps. This is atypical for a CPU. Any good CPU is not supposed to require tedious optimization; you're supposed to be able to compile your code for it, get great performance off the bat, and be done. If/when there ever exists a decent compiler for AltiVec-enabled chips, then I will stand corrected.

I've read that thread. All I can say is, SPEC is the most widely accepted CPU benchmarking test there is. Its so-called pipeline tax doesn't seem to affect any other CPUs. You can whine all you want about unfairness, but that's to be expected from anyone getting butt-raped like Motorola's PPCs are.

If only software that has been hand-optimized for AltiVec can squeeze a decent amount of speed out of the G4, then that makes the G4 a speciality chip. Nobody wants to optimize for AltiVec because doing so results in completely non-portable code.

If it looks slow... and smells slow... it probably is.

Alex

By the same token we should eliminate MMX and SSE instructions on the x86. Give me a friggin' break.

In the perfect world, a CPU never has branching instructions, and a perfect flow of data to it. Guess what? We don't live in that world.
By not taking into account REAL WORLD performance issues you create a completely worthless benchmark. The pipeline tax applies to P4s, not PPCs.
Their longer pipeline causes a slowdown when they have to dump the whole thing when they get a bad instruction.

You seem to be quite knowledgeable about sodomy...you wouldn't happen to be a Catholic priest, would you?

 

[alex_ant]

Originally posted by Rower_CPU
By the same token we should eliminate MMX and SSE instructions on the x86. Give me a friggin' break.

Intel's compiler optimizes for both MMX and SSE automatically. There is no such compiler that optimizes for AltiVec. I'm not saying AltiVec should not be a part of the picture, I'm just saying that it is not an automatic performance boost, and is also not a very desirable performance boost for software developers because it requires them to eliminate the portability of their code. Nor does it support double-precision FP, crucial to quite a few scientific apps. AltiVec is a hack designed to save the PPC from utter FP disgrace (which it does, more or less). So again, until someone delivers a compiler that takes advantage of AltiVec on non-specialized code by default, any benchmark performance gains due to AltiVec should have an asterisk next to them.

In the perfect world, a CPU never has branching instructions, and a perfect flow of data to it. Guess what? We don't live in that world.
By not taking into account REAL WORLD performance issues you create a completely worthless benchmark. The pipeline tax applies to P4s, not PPCs.
Their longer pipeline causes a slowdown when they have to dump the whole thing when they get a bad instruction.

This explains the P4's good benchmark results. What about the P3, all of the Athlons, the Power3, the Power4, the RS64, the Alpha 21164 and 21264, the MIPS R12K and R14K, the UltraSPARC, the Itanium, and the PA-RISC? Is the PowerPC so unique that any benchmarking done on it is wildly inaccurate and irrelevant?

Again. If it looks slow and smells slow... it's probably slow (at everything not heavily hand-optimized for AltiVec).

Alex

 

[TechLarry]

Re: Re: Re: Re: Re: Case closed.

Originally posted by alex_ant

I never claimed I could build a server for $500 or $1000. What I proved was that it's possible to build a server faster than the fastest Mac server for less than half the price. Not taking into account service, support, reliability, etc... but it can be done.

Alex

>>Not taking into account service, support, reliability, etc... <<

And, of course, no IT manager worth their salt would EVER take those little, niggling details into account...

This has become so funny

TL

 

[Rower_CPU]

Originally posted by gjohns01


Actually you've completely ignored me. I've stated what I need to do. Read through the posts. But just to speed things up here I'll repeat myself. There are no enterprise databases available on OS X. (i.e. No 8i, No DB2, No Sybase, No SQL Server) Using mySQL, Postgresql, Frontbase, or Openbase won't cut it. I can't develop to those we don't use them. There is no official J2EE release for Mac OS X. And yes I am aware that you can use one from a JBoss distribution or some other source. There is no JBuilder for Enterprise available for OS X. There is no Visual Studio. There is no Visual Age. No Websphere, No IPlanet, No Vignette, or any other application I deal with in my everyday life. Therefore, OS X Server and hardware have ZERO to offer me. I'm sure everyone's favorite board member Larry Ellison feels the same way, because he hasn't ported 8i to OS X. He has no reason to. Apple doesn't have the hardware to run it and doesn't command any respect in the enterprise for Oracle to make the effort. Now go design something. And will you're at it, tell me how Photoshop, Premiere and Final Cut can tell me how many transactions per minute I can achieve on my hardware.

Once again you've completely side-stepped my two very simple and straightforward questions...ROI and a side-by-side real world challenge...

 

[Rower_CPU]

Originally posted by alex_ant
This explains the P4's good benchmark results. What about the P3, all of the Athlons, the Power3, the Power4, the RS64, the Alpha 21164 and 21264, the MIPS R12K and R14K, the UltraSPARC, the Itanium, and the PA-RISC? Is the PowerPC so unique that any benchmarking done on it is wildly inaccurate and irrelevant?

Again. If it looks slow and smells slow... it's probably slow (at everything not heavily hand-optimized for AltiVec).

Alex

I'll have to check the specs on all of those CPUs and get back to you.
BTW - Here's the actual article I was referring to: http://www.heise.de/ct/english/02/05/182/ (graphs at the bottom)

Funny...it says the G4 performs poorly, but it seems match or outperform the Pentium at each task. Where are you getting your results from?

 

[alex_ant]

Re: Re: Re: Re: Re: Re: Case closed.

Originally posted by TechLarry
>>Not taking into account service, support, reliability, etc... <<

And, of course, no IT manager worth their salt would EVER take those little, niggling details into account...

This has become so funny

Who said they wouldn't? Certainly not me. You have a good point. Have you been following the thread? The issue of support and reliability has already been brought up.

Alex

 

[alex_ant]

Originally posted by Rower_CPU
Funny...it says the G4 performs poorly, but it seems match or outperform the Pentium at each task.

Yes... the 1GHz Pentium III. Which is pretty obsolete today. And don't forget the G4's raw FP speed, which is anemic. If you want to put the 1GHz G4 up against a processor of equal cost, the Athlon XP 2100 actually costs less and will piss all over the G4, AltiVec or no AltiVec. (I heard a figure that Motorola charges $250 for the 1GHz Apollo in quantities of 10,000... it sounds about right.)

Please be aware that I'm not slagging off the Mac. As I said, I think it's a great platform, and I wouldn't want to use anything else as my main computer. But the topic right now seems to be servers and raw CPU performance, and this is one area where the Mac just does not cut it, no matter which way you look at it.

Where are you getting your results from?

http://www.mtl.t.u-tokyo.ac.jp/~nminoru/memo/spec_cpu2000.html

It's in Japanese, but the graphs are still readable.

Alex

 

[alex_ant]

Originally posted by Rower_CPU
Once again you've completely side-stepped my two very simple and straightforward questions...ROI and a side-by-side real world challenge...

I'm just curious about that challenge you've offered. The Mac is slower and generally less capable than just about everything else available as a server at the same price point, as we've already established. What do you need proven that the very telling statistics already presented to you don't? Do you think that the Power Mac has some sort of secret speed-booster chip on the motherboard that causes it to defy all benchmarks but shoot ahead in the real world? Gjohns has already brought up software compatibility. I've brought up CPU performance. What else do you need? Memory bandwidth statistics (on which the Mac will get demolished)? I/O performance statistics? Transactions per second, hits per second, network streams per second, what?

It just sounds like you're unwilling to accept the fact that the Mac is a relatively poor-performing server with a limited selection of proprietary software. Which is not to say it's necessarily a bad choice... I mean, I just don't understand - it seems like the only reason you could possibly be requesting a real-world challenge is because 1) you're hoping your challenge-ee will decline, therefore giving you victory by default, or 2) you are simply too blindly devoted to Apple to believe that the Mac might not possibly be the best choice for every task for which it is marketed, and you are in denial about the fact that it isn't.

Alex

 

[eric_n_dfw]

Originally posted by alex_ant

Yes... the 1GHz Pentium III. Which is pretty obsolete today. And don't forget the G4's raw FP speed, which is anemic. If you want to put the 1GHz G4 up against a processor of equal cost, the Athlon XP 2100 actually costs less and will piss all over the G4, AltiVec or no AltiVec. (I heard a figure that Motorola charges $250 for the 1GHz Apollo in quantities of 10,000... it sounds about right.)
[/B]

But you really can't say "AltiVec or no AltiVec". Look at the RC5 stats or QuickTime rendering times. Code that takes advantage of AltiVec smokes that which uses MMX/SSE/etc... on the x86 chips at even twice the clock speed.

 

[alex_ant]

Originally posted by eric_n_dfw
But you really can't say "AltiVec or no AltiVec". Look at the RC5 stats or QuickTime rendering times. Code that takes advantage of AltiVec smokes that which uses MMX/SSE/etc... on the x86 chips at even twice the clock speed.

Fair enough. As I said before, the G4 with AltiVec will do better at some specialized tasks, like those like RC5 and QuickTime And FCP and Photoshop filters which are able to do well with mad single-precision FP ability.

Alex

 

[Rower_CPU]

Originally posted by alex_ant
Please be aware that I'm not slagging off the Mac. As I said, I think it's a great platform, and I wouldn't want to use anything else as my main computer. But the topic right now seems to be servers and raw CPU performance, and this is one area where the Mac just does not cut it, no matter which way you look at it.

I understand. I never meant to imply that the G4 would beat the P4 in EVERY task out there. I just don't like it when people look at one aspect of a chip's performance and dismiss it entirely.

I am painfully aware of the shortcomings of the PPCs. My 1.33 GHz Athlon rig is much faster in certain applications. But for what I do, the G4 is a far better chip. And I don't have to deal with the horrible Windows OS.

I'm really looking forward to advances in Moto's technology (DDR, RapidIO, etc.). Apple definitely needs some new technology to remain competitive.

 

[eirik]

Originally posted by alex_ant

Intel's compiler optimizes for both MMX and SSE automatically. There is no such compiler that optimizes for AltiVec.

Again. If it looks slow and smells slow... it's probably slow (at everything not heavily hand-optimized for AltiVec).

Alex [/B]

Alex,

What you said about SIMD implementations, quoted above, confuses the hell out of me. From what I've read, it is the other way around.

AltiVec is a generalized vector processor or co-processor. Exploiting AltiVec is generally accomplished by writing your algorithms in terms of vector/matrix operations. That's it, the compiler does the rest.

MMX, SSE, and AMD's flavor (?), are actually much more specialized for multimedia applications. While more specialized, I don't believe there are such convenient functions that one can simply call and get an instant photoshop filter, per se.

In fact, from what I've read, the x86 world is lacking in SIMD optimized applications because implementation is actually harder on the x86 platforms. Has this changed in the last three years? Has Intel finally made it simple?

Well, this is what I've read in every head to head comparison of the SIMD architectures. I haven't actually done any programming in this myself. So, there's a grain of salt to go with this post.

Has anybody on this thread actually programmed in both G4 and x86 SIMD?

Eirik

 

[alex_ant]

Originally posted by eirik


Alex,

What you said about SIMD implementations, quoted above, confuses the hell out of me. From what I've read, it is the other way around.

AltiVec is a generalized vector processor or co-processor. Exploiting AltiVec is generally accomplished by writing your algorithms in terms of vector/matrix operations. That's it, the compiler does the rest.

MMX, SSE, and AMD's flavor (?), are actually much more specialized for multimedia applications. While more specialized, I don't believe there are such convenient functions that one can simply call and get an instant photoshop filter, per se.

In fact, from what I've read, the x86 world is lacking in SIMD optimized applications because implementation is actually harder on the x86 platforms. Has this changed in the last three years? Has Intel finally made it simple?

Well, this is what I've read in every head to head comparison of the SIMD architectures. I haven't actually done any programming in this myself. So, there's a grain of salt to go with this post.

Has anybody on this thread actually programmed in both G4 and x86 SIMD?

Eirik

There's an equally-sized grain of salt to go with my post as well. I was under the impression that although there was no suitable parallel to AltiVec in the x86 world, the SIMD implementations on x86, although somewhat ugly hacks in comparison to AltiVec, are actually taken into account by Intel's compilers without any extra effort on the part of the programmer. I was also under the impression that, while AltiVec does not need to be written for specifically, it does require code to be written in terms of vector/matrix operations... which will render said code sub-optimal on a chip without an onboard vector unit. The Apple C compiler (GCC 2.95.2 is it?) will not vectorize non-vector code for you - you have to do it yourself.

Although I could be entirely full of it...

Alex

 

[eirik]

SIMD

Alex,

I think we're totally in agreement.

I especially like your point about vector/matrix operations. It hadn't ocurred to me that if a platform handles vector/matrix operations poorly, very general operations BTW, then portis to that platform would suck.

On the other hand, with the x86 SIMD implementations, if you port to another platform, from what I understand, the software won't work at all because the functions are too specialized.

I hope somebody whose worked with both can help us better understand the x86 SIMD shortcomings.

Eirik

 

[TechLarry]

Re: Re: Re: Re: Re: Re: Re: Case closed.

Originally posted by alex_ant

Who said they wouldn't? Certainly not me. You have a good point. Have you been following the thread? The issue of support and reliability has already been brought up.

Alex

Who could _possibly_ follow this thread? The 'conditions' keep changing!

Which is basically the point I'm trying to make

To indicate how out of hand it's gotten, please remember that the INITIAL, ridiculous post was the guy claiming he could build a dual-processor computer system for under $500.

And here we are

TL