Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
More 90nm PowerPC Notes
- Thread starter MacRumors
- Start date
- Sort by reaction score
Henriok:
Mac-Xpert:
mjtomlin:
a) The 90nm G5 throws off more heat than a 130nm Opteron, or
b) The 90nm G5 is not currently available above 2.0ghz, or
c) Apple decided to make Xserves less compeditive than they could be.
Why would you think this? The 130nm G5 wasn't maxed out at 2.0ghz; 2.0ghz was pretty much it's launch speed. 2.4ghz seems like a very attainable speed at 130nm, and I don't think asking 25% higher for the 90nm chip is unreasonable. If you ask me, people are so set on the idea of a "980" that they are underestimating the scalability of the 970. I am not expecting any "980" till after the summer/fall tower refresh (although I must say that IBM's choice of 512k L2 seems to leave the high-end wide open for a "980").
Mac-Xpert:
Heh, the G5 has not one but two things higher about its FSB vs the P4's. One is the bandwidth, and the other is latency. (Unfortunately high latency is bad.) Further, you should not compare processor clock speed to FSB speed and draw conclusions, what you want compare is processor performance to memory speed. What matters is how much data the processor can process, not how fast it is clocked. By that measure the G5 does not have a FSB advantage over the P4. Yes, a larger L2 would help.
mjtomlin:
Consider AMD's 130nm 2.2ghz Opteron that works fine 2 per 1U machine. One of three things is possible:
a) The 90nm G5 throws off more heat than a 130nm Opteron, or
b) The 90nm G5 is not currently available above 2.0ghz, or
c) Apple decided to make Xserves less compeditive than they could be.
Re: L2 cache
Opteron's biggest advantage is that it has the fastest memory subsystem out there bar none for a normal CPU. As in, the G5 has TWICE the latency of the Opteron. If you are running an app that doesn't fit into the L2 cache and regularly has to access the main core, the superior latency specs of the Opteron's memory subsystem will dominate the performance. This is due in significant part to the fact that Opteron has an on-chip memory controller. Combine the faster memory access and larger L2 cache, and you have a chip that is much better at keeping the execution units busy in the core.
A lesser advantage Opteron has is that the execution units in the core are more general purpose in scope than the ones in the G5, and therefore can be used more efficiently in some cases. But I personally think that this is largely unimportant compared to the substantial differences in the memory subsystems.
If the G5 had a memory subsystem that had super-low latency like the Opteron, the performance gap in the general case would be closed.
Opteron's biggest advantage is that it has the fastest memory subsystem out there bar none for a normal CPU. As in, the G5 has TWICE the latency of the Opteron. If you are running an app that doesn't fit into the L2 cache and regularly has to access the main core, the superior latency specs of the Opteron's memory subsystem will dominate the performance. This is due in significant part to the fact that Opteron has an on-chip memory controller. Combine the faster memory access and larger L2 cache, and you have a chip that is much better at keeping the execution units busy in the core.
A lesser advantage Opteron has is that the execution units in the core are more general purpose in scope than the ones in the G5, and therefore can be used more efficiently in some cases. But I personally think that this is largely unimportant compared to the substantial differences in the memory subsystems.
If the G5 had a memory subsystem that had super-low latency like the Opteron, the performance gap in the general case would be closed.
Maybe were both thinking of SSOI meaning another thing? I'm referring to "strained silicon." This is Intel's baby. It looks like a very promising technology across the board. IBM would do well to use this in the 970/980 chip.
Here are two articles of reference.
http://www.eetimes.com/semi/news/OEG20020813S0012
and
http://www.eetimes.com/story/OEG20020816S0048
2GHz 90nm 970...
ddtlm:
Well it seems they each sacrificed one thing for something else...
Apple: 3 drive bays
AMD: 2
Apple: 2 full length PCI-X slots
AMD: 1 full, 1 half
Apple: 8 RAM slots
AMD: 16
...etc
Apple's system may use a smaller cooling system and they needed to wait for the smaller chip (less heat dissipation)
The 2.0GHz? Well Apple may have wanted to keep the price down, there may not be a ton of them available yet, or they didn't want to overstep the towers, which is their bread and butter so to speak.
ddtlm:
Well it seems they each sacrificed one thing for something else...
Apple: 3 drive bays
AMD: 2
Apple: 2 full length PCI-X slots
AMD: 1 full, 1 half
Apple: 8 RAM slots
AMD: 16
...etc
Apple's system may use a smaller cooling system and they needed to wait for the smaller chip (less heat dissipation)
The 2.0GHz? Well Apple may have wanted to keep the price down, there may not be a ton of them available yet, or they didn't want to overstep the towers, which is their bread and butter so to speak.
Re: Faster PowerMacs soon?
I don't recall anyone stating a 2.6-2.8 GHz top end for the 130nm 970. A search of the archives should show that the early info on the 90nm 970 indicated a top end of 2.8 GHz. Although, as you say, these limits may be subject to a thermal target Apple has set, rather than the maximum frequency the process can produce.
To clarify, SSOI is a combination of the manufacturing technologies Strained silicon and Silicon On Insulator, thus the name SSOI. As it is now Intel uses strained silicon while IBM uses SOI, it's rumored that IBM will begin producing SSOI chips sometime late Q2. It's estimated that IBM will reduce power consumption on there chips by ~30% by using SSOI.
Re: 2GHz 90nm 970...
Or, maybe they're reserving the faster chips for the new Power Macs that will be announced in a couple of weeks?
Or, maybe they're reserving the faster chips for the new Power Macs that will be announced in a couple of weeks?
Alot of people seem to be unaware that while we are reaching 90nm technology you are assuming the chips will run cooler, quite the oposite is true.
Taking a 130nm design and reducing it to 90nm will lower the power dissipation to a degree but all is not so jolly the leakage current are much greater meaning higher static power dissipation the P4 prescot 3Ghz=104W under max load, rougly 80-90W when idle so ive been told, though 104W seems suspicious to me considering the 3.4Ghz eats around 140-150W. Cooling for the prescotts is an Intel custom flower HSF design with fan @ 45db noise level with ideal temps around 70 Degrees C, err thats quite hot and loud dont you think what are the load temps?
The die shrink makes cooling the chip more difficult as there is less contact area and with speed boosts you actually end up with a hotter chip than you started with (albiet at a highter speed and also much more noisy fans etc..). There is nothing wrong with this unfortunetly air cooling for me went obsolite more than a 2 years ago and has currently hit a brickwall in terms of performance. Anyone thinking a 3GHz G5 is gona be a cool, quiet, and less power hungry chip is misstaken.
Taking a 130nm design and reducing it to 90nm will lower the power dissipation to a degree but all is not so jolly the leakage current are much greater meaning higher static power dissipation the P4 prescot 3Ghz=104W under max load, rougly 80-90W when idle so ive been told, though 104W seems suspicious to me considering the 3.4Ghz eats around 140-150W. Cooling for the prescotts is an Intel custom flower HSF design with fan @ 45db noise level with ideal temps around 70 Degrees C, err thats quite hot and loud dont you think what are the load temps?
The die shrink makes cooling the chip more difficult as there is less contact area and with speed boosts you actually end up with a hotter chip than you started with (albiet at a highter speed and also much more noisy fans etc..). There is nothing wrong with this unfortunetly air cooling for me went obsolite more than a 2 years ago and has currently hit a brickwall in terms of performance. Anyone thinking a 3GHz G5 is gona be a cool, quiet, and less power hungry chip is misstaken.
Re: More 90nm PowerPC Notes
I'm suprised the transistor cout is so low. The Athlon XP has around 50million and the Opteron has just ove 100, with this intel cores being above that.
I'm suprised the transistor cout is so low. The Athlon XP has around 50million and the Opteron has just ove 100, with this intel cores being above that.
Yes cache is expensive because large caches mean more silicon area hence a higher failier rate, you think throwing away a 9MB cache Itanium 2 is cheap because of a few dodgy transistors in the caches.
Why are you mentioning the Itanium 2 and saying its expensive, its a workstation ment for EDA tools, the cost of these babies is just a rounding error compared to the software which they run.
I find it hard to believe this, in my opinion as I said before the 90nm version is very likely running hotter than the 130nm version due to the process shrink and also the 1u heatsinks, bear in mind any 1u rack server runs hotter than thier bigger cousins.
It may actually be a bit more like option 'c'.
Consider how Apple got burnt by demand for the original 2gHz machine - I would bet they're playing safe so not to repeat the delivery problems - especially for servers which could hurt their reputation a lot more than for desktop machines.
The move to the 90nm process undoubtedly raised clock speeds. Probably as high as the 2.6 mark that everyone has been speculating about. And with the interest in G5 clusters (and with the release of the server cluster node) I'd guess that Apple is expecting demand for these babies to outstrip supply of the higher clock procs. Or maybe they're just playing it safe.
So possibly they're either stockpiling higher clocked chips, or they're underclocking them. I'd expect a speed bump in a month or so when demand for the servers has been established. And I think then it will be to 2.3gHz or so, leaving another step before 3.0gHz at the end of the year. The last thing Apple need as they're trying to break into the enterprise market is a reputation for late delivery (which they already have to a certain extent).
How is Apple gona enter the enterprise market by using ATA drives?
Do you see servers using ATA drives?
OK SATA has some SCSI features but mechanically it is still built to be a desktop drive.
Who puts a drive that is meant for desktop use (home/games/multimedia/Digital content creation etc....) into a server?
Where is SCSI which is designed to run 24/7, has outstanding reliability, performance especially under heavy/constant/multiple user load. Iam afriad ATA dont cut it for servers period even for lowend ones, people would probably use WD Raptors.
Yes SCSI is expensive but how valuable is your data?
Can you put a price on it?
Common Philosophy of a SCSI drive in my opinion:
Reliability 1st
Performance 2nd
Total Cost of ownership 3rd
I can only see Apple entering the low-end server market nothing more.
Do you see servers using ATA drives?
OK SATA has some SCSI features but mechanically it is still built to be a desktop drive.
Who puts a drive that is meant for desktop use (home/games/multimedia/Digital content creation etc....) into a server?
Where is SCSI which is designed to run 24/7, has outstanding reliability, performance especially under heavy/constant/multiple user load. Iam afriad ATA dont cut it for servers period even for lowend ones, people would probably use WD Raptors.
Yes SCSI is expensive but how valuable is your data?
Can you put a price on it?
Common Philosophy of a SCSI drive in my opinion:
Reliability 1st
Performance 2nd
Total Cost of ownership 3rd
I can only see Apple entering the low-end server market nothing more.
I bet the problem for apple is keeping these chips cool, i think that even 2.6GHz with the 970 would be possible just by moving to liquid cooling, of coarse this would be no easy feat to adapt this kind of cooling technique for home use although it has gone mainstream via complete kits.
That's like saying that the G3 design pushed from 200 MHz to 1.1 GHz. It did but 750 and 750GX are quite different chips produced at completely different fabs. IBM introduced copper, SOI, larger and on die L2 cache and faster/wider memory buses on its road to 750GX.
There are many cores and development stages in a Pentium 4-design just as there are in the G3 or G4-designs.
The Willamette/Northwood/Prescott Pentium4s going from 1.5 to 4 GHz (+266%) on 130 nm and 90 nm is directly comparable to the the development Motorola did from 7450@533 MHz to 7457@1.42 GHz (+266%).
Look! Motorola did just as good as Intel! And.. Intel haven't got to 4 GHz just yet.
Letting the 970 go from 1.4 to 2.8 in just one step is not to shabby at all. Doing it in under a year is fenomenal.
If you look at the individual 750 models, they didn't scale at all. They are introduced at one frequenzy range and then they didn't climb. 970 was supposted to range from 1.4 to 1.8 GHz, Apple pushed it to 2 GHz. I have never seen any official data on the complete 970-family ranging higher than 2.5 GHz, and that'd be the 970FX I presume.
AppleInsider's sources clamied that 970 will only be updated once, to 90 nm, before 980 will take its place. 980 will, according to AI, be updated two times after the first model. In this matter I have no objection to what AI's sources claims.
I guess that 970 and 970FX will scale past its initial range when the fab mature and they eventually get enough individual cores that can be clocked past 3 GHz to start selling them. But I'm certain that, in volume, 980 will get there first and that it's that chip that Steve was referring to.
And.. I'm pretty certain that Intel have introduced their processors at an intentionally low frequency just so they can "speed bump" them a few months later. Intel's processors have almost always have an insane potential for overcloking. This is not the case with IBM and Motorola who have been under quite a lot of pressure just to keep up.
The "970" that was supposted to go to 2.8 GHz was the 90 nm variant, ie 970FX. The fastest 130 nm 970 I've seen even suggested previously is 2.4 GHz! That would be one hot processor. Not likely!
A die shrink in the past have always resulted in speedbump in the ammount of 30-50% if all other factors stay the same. That would be 970 @ 1.8-2.0 GHz shrunk to 970FX @ 2.7-3.0 GHz while beeing an optimist. 970FX @ 2.6 wouldnt surprise anyone, but pushing up to 3 GHz anytime soon would be VERY surprising. At least to me.
IBM are planning for a SSOI deplyment on their 65 nm-process. I guess they might quicken their pace since Intel is getting there first. Intel on the other hand has mountains of problems with power consumptions, IBM has not.
With all due respect, we know almost nothing about the scalability of 970. AFAIK, it has already surpassed its original top speed at 1.8 GHz by 11% to 2 GHz. Going to 3 GHz from 1.8 GHz is a 66% increase and that's too much to ask I think. Why do you think 2.4 GHz is a very attainable speed for the 970 design? Care to elaborate?
If 970 goes to 2.4 GHz (33% above original spec) and 970FX gets a 50% increase due to die shrink to 90 nm, and it then too get a 33% overclock speed bump just for the presumed exceptional scalability, we are pushing for 4.7 GHz.. Not likely!
Even increasing it a mere 25% due to die shrink and then boost it 33% due to the supposed greatness would put 970FX at 4 GHz. Not likely either.
I have seen no evidence at all for supporting any claims for exceptional scalability of the 970 design or IBM's fab. Heck.. even going up to 2 GHz seems to be pushing it with heat sinks the size of bricks and the necessity of 9 fans.
Intel would **** their pants though
If tye arn't already.This seems reasonable. Question: Might using dual core chips or insanely large caches be way to enlarge the die area for cooling purposes? Not the only purpose of course.
BEcause it IS expensive. And Intel initially intended it to take over the world and then some.. Guess they have had some time to rethink that strategy.
They intended on taking over the world what with $5000 machine i doubt it.
The Itanium is expensive for the performance you get compared to say a SUN machine but if the software was actually thier it would sell in the intended market however iam talking 8-10 years ago, roughly around 1992 EDA companies started compiling all thier EDA tools to 64-bit SPARQ machines and hence SUN still remains number 1 in terms of software tools and support of which not even HP or IBM can match. EDA companies cannot just sudenly rewrite, recompile and recertify thier tools, it takes money lots of money and its the reason why so few tools are available on the Itanium.
And Hspice does not count in my opinion as Intel is sysnopsis's biggest customer they were litrally forced to port Hspice from SPARQ to Itanium. Itanium is a good product but it needs far too much effort for the performance which you get in the end.
Itanium pricing starting from $5000
where does it end Ill tell you id say around $34000 for a fully loaded system with 16GB of RAM and TOP spec DUAL 1.5GHz CPUS.
Expensive for a machine that only has 1 major EDA application what a joke. You can get 2 TOP SPEC SUN BLADE 2500's and run them using the distributed networking.
The Itanium is expensive for the performance you get compared to say a SUN machine but if the software was actually thier it would sell in the intended market however iam talking 8-10 years ago, roughly around 1992 EDA companies started compiling all thier EDA tools to 64-bit SPARQ machines and hence SUN still remains number 1 in terms of software tools and support of which not even HP or IBM can match. EDA companies cannot just sudenly rewrite, recompile and recertify thier tools, it takes money lots of money and its the reason why so few tools are available on the Itanium.
And Hspice does not count in my opinion as Intel is sysnopsis's biggest customer they were litrally forced to port Hspice from SPARQ to Itanium. Itanium is a good product but it needs far too much effort for the performance which you get in the end.
Itanium pricing starting from $5000
where does it end Ill tell you id say around $34000 for a fully loaded system with 16GB of RAM and TOP spec DUAL 1.5GHz CPUS.
Expensive for a machine that only has 1 major EDA application what a joke. You can get 2 TOP SPEC SUN BLADE 2500's and run them using the distributed networking.
Using DUAL cores could be a way out in terms of thermals but would require complicated power regulation (5/6 phase) on the MB depending of the power requirements but that wouldnt be a problem id worry about the yield thats a 120M transistor chip @ 90nm running at more than 2.2-3GHz thats gona be no small feat and how much $$$
I would say the 970 should be able to reach 2.6GHz at 130nm (I struggled to get my Opteron to 2.4Ghz on air cooling and thats a 100M transistor chip so I would assume the G5 could clock higher as its only 59M )considering the size of the HSF's in the G5 although Iam not certain of the noise???
I would say the 970 should be able to reach 2.6GHz at 130nm (I struggled to get my Opteron to 2.4Ghz on air cooling and thats a 100M transistor chip so I would assume the G5 could clock higher as its only 59M )considering the size of the HSF's in the G5 although Iam not certain of the noise???