PDA

View Full Version : more on the 970 and G4 vs. Intel


gopher
Nov 1, 2002, 02:36 PM
http://www.igeek.com/browse.php?id=1101

Very good read...what are the advantages and disadvantages of both processing methods and where Apple's going to be able to benefit and the Unix world will too.

Catfish_Man
Nov 1, 2002, 03:18 PM
...the arstechnica thread about thise article. A lot of it is ************ (like the G4+ having a double pumped memory bus), and the rest is as PowerPC biased as he accuses the arstechnica article of being x86 biased. He makes some good points, but nothing that hasn't been made elsewhere, but better written and less biased (such as the article he's so scornful of). Basically, the Pentium 4 will probably be faster at single precision fixed point and double precision vector math (because of its double clocked ALUs and the fact that SSE2 supports double precision better than Altivec does). The 970 will be faster at double precision fixed point, all floating point, and single precision vector (because it's 64 bit, it has 2 FPUs, and Altivec is much better designed than MMX/SSE/SSE2). The 970 will also win at memory limited tasks (6.4GBps vs. 5.3GBps), will have more room to scale (by switching to .09 micron), and will support multiprocessing (and possibly multicore, later).

ddtlm
Nov 1, 2002, 05:23 PM
gopher:

Don't be so eager to accept everything that puts PPC chips in good light... the topic has been discussed to death in places like Arstechnica which have large, intellegent Mac sub-communities.

Catfish_Man:

double precision fixed point
Definately the wrong terminology... all integers have the same precision. Just say "64 bit integer math" and you'll be OK.

6.4GBps vs. 5.3GBps
Eek, errors. First of all, the PPC-970 bus runs at exactly 50% the clock of the core, so that number you provide is only relevant for the 1.8ghz PPC-970. Second, the P4's do not have a 5.3GB/sec bus... 533mhz yes, but do notice that there is not a power of 10 scaling between clock speed and bandwidth. :) Actual bandwidth: 64bits x 533,333,333 hz = 34133333312 bits/second = 3.97GB/second. For 1.8ghz PPC-970: 64bits x 800,000,000 hz = 6.0GB/second. Notice I am using the correct powers of 2 (1024KB = 1MB, etc) and not the powers of 10 (1000KB != 1MB, etc) that IBM seems to use when talking about their bus speed.

SSE2 supports double precision better than Altivec does
Yes, but can SSE2 keep up with the two fused-add-multiply FPU units on a PPC-970? It may not be able to.

will have more room to scale (by switching to .09 micron)
And you think the P4 won't? Of course it will, and probably before IBM even releases the PPC-970 at 130nm. I'm serious too, by mid summer Intel is going to be rolling out 90nm production on some chips.

and will support multiprocessing (and possibly multicore, later).
Multiprocessor P4's exist and are called Xeons, so no win for the PPC-970 there. Intel will also be rolling out "hyperthreading" this year, which may in some cases act as two cores on one chip. Yep, this year.

Chaszmyr
Nov 1, 2002, 05:42 PM
In my opinion, if Apple uses the 970, at the time it is released it will still be behind Intel (who will still be the industry leader at the time). However, I believe the 970 has more potential than Intel's current chips, or even their future hyperthreading chips. If IBM keeps their act together they might pull ahead.

ddtlm
Nov 1, 2002, 06:15 PM
Chaszmyr:

In my opinion
Unfortunately very very few people are qualified to argue the finer points of PPC-970 vs P4. I'm not one of them, I bet you're not, and in that case neither of our opinions mean anything. At this point looking far into the future and declaring the PPC-970 the winner sounds an awful lot like Apple-worshipping.

Computer_Phreak
Nov 1, 2002, 06:55 PM
performance wise, megahertz wise, the P4 is the current leader, and will be in the foreseeable future. Intel has gobbs of money to throw at R&D... which chipmakers like AMD lack (on the Intel scale). IBM isn't a chipmaker, the PPC 970 is kind of a side thing that they came out with.

Telomar
Nov 1, 2002, 09:47 PM
Originally posted by ddtlm
Actual bandwidth: 64bits x 533,333,333 hz = 34133333312 bits/second = 3.97GB/second. For 1.8ghz PPC-970: 64bits x 800,000,000 hz = 6.0GB/second. Notice I am using the correct powers of 2 (1024KB = 1MB, etc) and not the powers of 10 (1000KB != 1MB, etc) that IBM seems to use when talking about their bus speed.Actually you are using the wrong numbers. When discussing bandwidth it has always been the practice to use powers of 10.

End result is PIV = ~4.2 GB/s and PPC = 7.2 GB/s. Those are theoretical numbers though and the PPC970 bus is considerably more efficient, although the PPC970 bus is actually 3.6 GB/s each way.

Only the OS and RAM really use the powers of 2 and they are meant to use a different prefix for those except damn people and programmers are too lazy or something.

Also at 1.8GHz the PPC970 should have a 900 MHz bus. I'm assuming that was just a typo though because you had it right earlier with the 50% core clock.

ddtlm
Nov 2, 2002, 01:47 PM
Telomar:

What? Your entire position is rediculous. Bandwidth is a measurement of flow of data. Data is beasured in bytes, kilobytes, megabytes and gigabytes, etc etc. These each are different from each other by exactly a factor of 1024, not 1000. 1000 is wrong, and always will be. There can be no discussion of that claim because it is absolutilty totally wrong.

The P4 and PPC-970 have the number I have presented. Marketeers prent inflated, wrong numbers because they can get away with it. But they are still wrong. The P4 bus cannot transfer the data that is claimed, nor can the PPC-970 bus, and therefore the figures you quote are wrong.. Ok? They are wrong. End of discussion.

Only the OS and RAM really use the powers of 2 and they are meant to use a different prefix for those except damn people and programmers are too lazy or something.
Totally wrong. Only marketeers use factors of 1000 because it distorts the truth in their favor, in all other cases all matters of memory are referred to in powers of 2, and factors of 1024 (2^10).

Also at 1.8GHz the PPC970 should have a 900 MHz bus. I'm assuming that was just a typo though because you had it right earlier with the 50% core clock.
Because of overhead only 800mhz of this is available for transferring data. The P4 bus, and no other common bus, suffers from this design drawback. IMHO, IBM is stretching the truth to claim it is 900mhz because it does not transfer useful data at that rate under any circumstances.

atomwork
Nov 2, 2002, 05:23 PM
Hi guys. I'd like to ask this. I read more and more about the posibility that some day Apple runs on a IBM chip. Now, I know that the great advantage for us designers is the motorola processor since its thinks or caculates or works different then the intel chips. Will this not change the way an apple mac works or looks. Just curiouse.

D

Telomar
Nov 2, 2002, 05:58 PM
Originally posted by ddtlm
Telomar:

What? Your entire position is rediculous. Bandwidth is a measurement of flow of data. Data is beasured in bytes, kilobytes, megabytes and gigabytes, etc etc. These each are different from each other by exactly a factor of 1024, not 1000. 1000 is wrong, and always will be. There can be no discussion of that claim because it is absolutilty totally wrong.

The P4 and PPC-970 have the number I have presented. Marketeers prent inflated, wrong numbers because they can get away with it. But they are still wrong. The P4 bus cannot transfer the data that is claimed, nor can the PPC-970 bus, and therefore the figures you quote are wrong.. Ok? They are wrong. End of discussion.


Totally wrong. Only marketeers use factors of 1000 because it distorts the truth in their favor, in all other cases all matters of memory are referred to in powers of 2, and factors of 1024 (2^10).


Because of overhead only 800mhz of this is available for transferring data. The P4 bus, and no other common bus, suffers from this design drawback. IMHO, IBM is stretching the truth to claim it is 900mhz because it does not transfer useful data at that rate under any circumstances. You really don't have the first clue what you are talking about. Next you will try and teach people a kilowatt is really 1024 watts and not 1000 and the power companies are just out to extort money from you :rolleyes:

The following prefixes are all predefined SI prefixes:

Kilo, Mega, Giga, Tera, Peta, etc.

Any engineer, scientist or person who has passed high school is aware that they are defined as official prefixes meaning 10^3, 10^6, 10^9, 10^12 and 10^15. You can check that with any national standards body.

Computer OSs and memory rely on a binary system however and it was easier in the early days to just take the nearest SI prefix (especially since they didn't have prefixes of their own). The answers aren't actually correct though and 1024 bytes isn't actually a kilobyte, that's 1000 bytes. By convention it was accepted due to the lack of other prefixes but it isn't correct nor has it ever been. The error has simply been ignored.

As time went on the error associated with this method of binary calculation and using SI prefixes becomes a larger and larger problem to the extent that in 1998 a new set of prefixes were defined for binary numbers to deal with it. You take the first two letters of the old prefix and add bi. ie:

Kibi, Mebi, Gibi, Tebi, etc = 2^10, 2^20, 2^30, 2^40 and so forth.

Now computer operating systems and memory manufacturers had good reason to use binary numbers and since they didn't have their own prefixes could previously be excused but bandwidth and storage never had such a reason so have used the prefixes as properly defined.

As for bus overhead every bus has overhead. The MPX bus, which is among the most efficient around runs at roughly 15% overhead. And it only even achieves those numbers in very specialised cases. Generally the number is closer to 25%. PC buses are generally worse still operating at around 65% of theoretical numbers or so. No small part of this is just inherent in operational costs of a bi-directional bus.

No other bus suffers that same area of overhead but every bus suffers from overhead. Also uni-directional buses allow for some reduction in overhead associated with a bi-directional bus.

I'd be expecting much better throughputs out of IBM's pair of uni-directional buses than from a PIV bus (except in the case where data streams in almost solely one direction). All buses have overheads so if you plan to factor in the PPC970's then factor in the PIV's.

Telomar
Nov 2, 2002, 07:46 PM
Oops wrong button.

Borborygmus
Nov 2, 2002, 08:14 PM
Originally posted by atomwork
Hi guys. I'd like to ask this. I read more and more about the posibility that some day Apple runs on a IBM chip. Now, I know that the great advantage for us designers is the motorola processor since its thinks or caculates or works different then the intel chips. Will this not change the way an apple mac works or looks. Just curiouse.

D

First of all, ignore the hot air being tossed around here. Most of these guys are just spouting meaningless techno-babble to impress each other. Most of the arguments about the "slow" 970 are utterly foolish to anyone who knows what is really going on. Clock speed is only one of five factors that affect actual performance and is actually one of the least of these factors in predicting actual data flow, but it is the easiest for marketers to sucker the buyer with. It's a simple and easy to understand number... but it really means almost nothing. It's like predicting how fast your car will go by bragging about its idle speed.

To answer your original question... the IBM chip is still a PPC chip, albeit it an incredibly powerful one. It is also a full 64 bit processor but can also process older 32 bit PPC data and code. That means it will run "Cocoa" and "Carbon" apps under Mac OS X just fine.

In short... it's still a Mac, it will run Mac OS X apps just fine but a lot FASTER!

This means that a Mac running a PPC 970 will be a full 64 bit RISC Unix workstation... not a 32 bit glorified game player's toy, like Pentium 4 machines.

Also... as a comparison... Intel is touting the Itanium-2 at 900 MHz ( only 1/3 the "speed" of a Pentium 4) as the most powerful thing they make. If clock speed was really the measure of performance... then why is Intel telling the average PC Weenie that speed is everything (so as to sucker them into buying P4 boxes) but telling their corporate buyers not to believe that hype and believe that the "slow" Itanium-2 is really the faster chip?

Which lie is the truth?:confused:

Chaszmyr
Nov 2, 2002, 08:48 PM
Originally posted by ddtlm
At this point looking far into the future and declaring the PPC-970 the winner sounds an awful lot like Apple-worshipping. [/B]

Yes and no. If in fact Apple is going to use the IBM 970 you have to assume they had a reason for choosing that chip.

ddtlm
Nov 3, 2002, 12:52 PM
Chaszmyr:

You mean such as it being PPC? Going to x86 would be no small undertaking because every app everyone has would suddenly not work (unless the maker could be bothered to recompile and distribute it)... it would be a bad thing. Apple wants to stick will PPC. The PPC-970 is PPC. The P4 is not.

Again, it sounds to me like you are viewing Apple in an unrealistically positive light.

Catfish_Man
Nov 3, 2002, 03:15 PM
Originally posted by ddtlm
gopher:

Don't be so eager to accept everything that puts PPC chips in good light... the topic has been discussed to death in places like Arstechnica which have large, intellegent Mac sub-communities.

Catfish_Man:


Definately the wrong terminology... all integers have the same precision. Just say "64 bit integer math" and you'll be OK.


Eek, errors. First of all, the PPC-970 bus runs at exactly 50% the clock of the core, so that number you provide is only relevant for the 1.8ghz PPC-970. Second, the P4's do not have a 5.3GB/sec bus... 533mhz yes, but do notice that there is not a power of 10 scaling between clock speed and bandwidth. :) Actual bandwidth: 64bits x 533,333,333 hz = 34133333312 bits/second = 3.97GB/second. For 1.8ghz PPC-970: 64bits x 800,000,000 hz = 6.0GB/second. Notice I am using the correct powers of 2 (1024KB = 1MB, etc) and not the powers of 10 (1000KB != 1MB, etc) that IBM seems to use when talking about their bus speed.


Yes, but can SSE2 keep up with the two fused-add-multiply FPU units on a PPC-970? It may not be able to.


And you think the P4 won't? Of course it will, and probably before IBM even releases the PPC-970 at 130nm. I'm serious too, by mid summer Intel is going to be rolling out 90nm production on some chips.


Multiprocessor P4's exist and are called Xeons, so no win for the PPC-970 there. Intel will also be rolling out "hyperthreading" this year, which may in some cases act as two cores on one chip. Yep, this year.

Integer and fixed point are referring to the same thing in this case. IIRC, they're both accurate (and both used).

I'm comparing the top of the line 970 to the Pentium 4s that will be out when it is out (Prescott based), so they're already .09 micron with hyperthreading (which is already out, btw) and a 667MHz bus. I don't remember if 5.3GBps is exactly right for 667MHz, but it's fairly close. Obviously the dual fmadd on the 970 will be extremely fast, but that's the scalar FPU isn't it? I was talking about vector. Hyperthreading actually slows things down in many cases as the two threads kick each other out of the cache. Also, I'm comparing desktop CPUs here (which is why I didn't include Opteron). If we get into higher end stuff like Opterons and Xeons, then the 970 will have a bit more trouble. So what I stated about the P4 not supporting SMP is correct. What I was saying about scaling is that I expect this comparison to hold for a .13 micron 970 vs. a .09 micron P4. When the 970 goes to .09 micron, I would expect it to win handily.

Catfish_Man
Nov 3, 2002, 03:23 PM
Originally posted by Computer_Phreak
performance wise, megahertz wise, the P4 is the current leader, and will be in the foreseeable future. Intel has gobbs of money to throw at R&D... which chipmakers like AMD lack (on the Intel scale). IBM isn't a chipmaker, the PPC 970 is kind of a side thing that they came out with.

...IBM not a chip maker? What about the POWER4 (the fastest server chip in the world until quite recently), or the 604, or the 750cx, 750cxe, and 750fx? How about the chip in the gamecube? Or the one they're working on for the PlayStation 3? IBM has what is currently the most advanced manufacturing process in the world (.13 micron SOI) and some awesome designs (POWER4). The 970 is not a side thing.

ddtlm
Nov 3, 2002, 04:07 PM
Catfish_Man:

Integer and fixed point are referring to the same thing in this case. IIRC, they're both accurate (and both used).
I'm not concerned with them both being used, saying "fixed point" is demonstrating ignorance.

Obviously the dual fmadd on the 970 will be extremely fast, but that's the scalar FPU isn't it? I was talking about vector.
Who cares if it's scalar or vector? Why is that important? There is nothing sacred about vector math, and if it can be beat by scalar than what is there to worry about?

Hyperthreading actually slows things down in many cases as the two threads kick each other out of the cache.
Cache conflict is an interesting theory which I had not heard of before, but then again seeing as how the 90nm P4 will sport 1024k of L2 than it can afford a little more usage than the PPC-970, can't it?

If we get into higher end stuff like Opterons and Xeons, then the 970 will have a bit more trouble.
Normal Xeons are almost exactly the same as normal P4's and they don't cost much more. Excluding them as being high end is rediculous in light of Apple's pricing scheme.

What I was saying about scaling is that I expect this comparison to hold for a .13 micron 970 vs. a .09 micron P4. When the 970 goes to .09 micron, I would expect it to win handily.
Based on what reasoning and/or evidence? Why would Intel not being going to a smaller proecess than 90nm when IBM gets to 90nm? This sounds like Apple worship to me.

nixd2001
Nov 3, 2002, 05:15 PM
Originally posted by Telomar
.... You take the first two letters of the old prefix and add bi. ie:

Kibi, Mebi, Gibi, Tebi, etc = 2^10, 2^20, 2^30, 2^40 and so forth.


Well I never. Live and learn. "Kibi" has a nice sound to it.

DakotaGuy
Nov 3, 2002, 05:17 PM
Originally posted by atomwork
Hi guys. I'd like to ask this. I read more and more about the posibility that some day Apple runs on a IBM chip. Now, I know that the great advantage for us designers is the motorola processor since its thinks or caculates or works different then the intel chips. Will this not change the way an apple mac works or looks. Just curiouse.

D

Correct me if I am wrong, but isn't the G3 chip in the iBook an IBM chip? Does it change the way the Mac works? Not that I can tell...

pilotgi
Nov 3, 2002, 05:34 PM
Yes, the G3 is a powerpc chip from IBM. And we haven't seen the most powerful version of it in an Apple computer yet.