More Liquid Cooling Diagrams

[wdlove]

I see your nit and raise you one

Yes that goes without saying however isn't fully accurate depending on what type of cooling system you are talking about (it is correct for the system in the G5).

The point I was making was that the system in the PM G5 doesn't contain a refrigerate like in the above and the cooling system isn't design to utilize/trigger the density/phase change of the refrigerate (at least nothing I see implies that). So the target of the cooling system cannot be cooled below the ambient temperature of coolant or the air used in the heat exchanger used to pull heat from the coolant. So it cannot cool a surface below the dew point of the air around it, so you will not get condensation. Which was the original question asked.

Thank you for the clear and concise explanation shawnce. It improves my confidence since I live in a high humidity area like Boston. The condensation issue could have been quite a problem otherwise.

 

[iris_failsafe]

The Dual G5 systems requiring a liquid cooling system need it to keep the machine quiet. No one wants G5s that sounds like G4 MDD. Unless you're in to sticking your head down the front end of a turbo jet engine. The Towers are "industrial" machines. Power over portability. I don't think anyone is planning on dual 2.5s in a 12" PowerBook. But one could imagine much like a 3.5 inch drive was made smaller for a laptop... An LCS or other interesting form of cooling in a compact and efficient set up for a single 2.5. Like encasing the board in a thermal diffusing hydro-polymer. Science fiction sure. But Apple is capable of innovation. There is a difference between what goes into an industrial cheese grater and one used on one of those cooking shows. Same idea, cheese grating, but size and application come into play. G5 Towers are an enigma of Industrial Power and carry a certain elegance. But the LCS within them I doubt will be used in what we see here today for smaller machines built around the G5.

Try a multicore power4 chip. It will occupy less area and I bet it can be insulated with aerogel

 

[wdlove]

Try a multicore power4 chip. It will occupy less area and I bet it can be insulated with aerogel

That aerogel is good stuff. I wish that Lawrence Livermore would be coming with some practical uses. Originally it was for homes. I'm not sure about the use in a G5, why would you want to keep the heat?

 

[jsw]

That aerogel is good stuff. I wish that Lawrence Livermore would be coming with some practical uses. Originally it was for homes. I'm not sure about the use in a G5, why would you want to keep the heat?

Excellent point - you certainly would not want to insulate the chip.

 

[Phinius]

Apple is probably overclocking the 970FX to reach 2.5GHz

I'll bet the 2.5GHz 970FX, that is used in the PowerMac, is overclocked and that is the main reason for Apple to use water cooling in the topend model. In other words, IBM is still having production problems with the 90-nm process and that has forced Apple to make a speed improvement for the PowerMac on only one model and to do that the company dramatically raised the heat output by overclocking the 970FX chip to 2.5GHz.

There are several reasons for my conclusion:

1) Apple kept two models of PowerMac at 1.8GHz and 2GHz, while moving the topend up 500MHz. Where are the 2.1GHz-2.4GHz processors? Why in the world would Apple release a 500MHz faster model without updating the speed of the other PowerMacs unless 130nm 970 chips are still being used for the 1.8GHz and 2GHz PowerMacs? Apple might do that as a last resort if there was a shortage of 970FX chips from IBM after several months of production (Apple did turn to overclocking the G4 to 1.42GHz, when they desperately needed faster chips for the PowerMac in order to keep PowerMac sales alive until the G5 arrived). If IBM is able to produce 2.5GHz chips, then there should be 970FX chips available at 2.1GHz-2.4GHz. Afterall, there should be a smaller supply of faster 2.5GHz processors than the slower speeds, since less chips would be able to run reliably at the highest speed.

2) The Xserve is backordered 4-6 weeks. Apple clearly stated in the last quarterly meeting with analysts that the reason for the delay in shipping G5 Xserves was the small supply of 970FX chips from IBM.

3) Apple had to include water cooling after having the G5 PowerMac out only 8 or 9 months and it is only available on a PowerMac that only runs 500MHz more than the slower models? That seems rather odd. It could be that Apple will drop water cooling if IBM is able to produce 3GHz 970FX chips in the next few months. If that is the case, then it would point decisively to overclocking of the 2.5GHz chip.

The upcoming July quarterly sales meeting with analysts should answer whether my hunch is right or not. If the backorder of Xserves was only due to a 970FX chip supply problem from IBM, then it will be rather obvious that there would be a even more severe severe shortage of 970FX chips for the much higher production PowerMacs.

 

[macsrus]

In the end, it isn't the liquid that makes the new G5s system so much more efficient, it's the radiator the coolant flows through. If the radiator didn't disperse more heat than a carved slug (heatsink), and if the coolant flow didn't allow better dispersion of that heat (energy),.. the ambient temperature of the coolant would rise too high and the water block wouldn't do anything at all.

Your whole post was right on the money

 

[johnnowak]

Who's to blame for all this? Powermac G4 owners. They whined and bitched constantly about a bit of fan noise. I had a DP G4 PM, and it wasn't that bad. So now Apple is noise obsessed, and we get expensive watercooling nonsense for 500 more mhz.

I want a loud-as-hell 3.5ghz dual G5. Bring it on!

 

[DWKlink]

There are several reasons for my conclusion:

1) Apple kept two models of PowerMac at 1.8GHz and 2GHz, while moving the topend up 500MHz. Where are the 2.1GHz-2.4GHz processors? Why in the world would Apple release a 500MHz faster model without updating the speed of the other PowerMacs unless 130nm 970 chips are still being used for the 1.8GHz and 2GHz PowerMacs?

Isn't this fairly stardard operating procedure for Apple product line updates? Increase the speed/features on the top level machine and just bump the other machines down a level while eliminating the bottom of the line.

 

[nsb3000]

This is supposed to fit into a PowerBook eventually? Wow, good luck...

Ya..right. I thought the 970fx were supposed to be cooler... Maybe we will see G5 powerbook in Jan 06? :-(

 

[shawnce]

Ya..right. I thought the 970fx were supposed to be cooler... Maybe we will see G5 powerbook in Jan 06? :-(

They do use less power at the same clock speed then a 970. If you use the numbers from IBM almost 50% the power (which is important not only because of heat generation but for battery life if used in a laptop). The change in die size difference however can require more rapid heat transfer then what was done before. In other words the parts product less heat and consumer less energy then before (however the static current is higher so better sleep/nap/etc. management will be needed) but they do so in a smaller area which could require changes in cooling to maintain system design goals.

Also as Apple has stated several times since the introduction of the G5 don't expect G5 laptops any time soon, not sure why some folks let their hopes get ahead of things when statements are made like that... this is Apple who usually doesn't say much of anything about future products, so that should be telling.

Anyway a 970FX G5 running at 1.4 - 1.6 GHz is not that far out of line with the power/heat characteristics of the current top end G4s being used in PowerMacs (ignoring the issues of the faster system buses that would be required).

 

[shawnce]

1) Apple kept two models of PowerMac at 1.8GHz and 2GHz, while moving the topend up 500MHz. Where are the 2.1GHz-2.4GHz processors? Why in the world would Apple release a 500MHz faster model without updating the speed of the other PowerMacs unless 130nm 970 chips are still being used for the 1.8GHz and 2GHz PowerMacs? ... If IBM is able to produce 2.5GHz chips, then there should be 970FX chips available at 2.1GHz-2.4GHz. Afterall, there should be a smaller supply of faster 2.5GHz processors than the slower speeds, since less chips would be able to run reliably at the highest speed.

Why not look at Apple system updating history? The often up the top end system and bump the prior top end system down to the middle or bottom slot. So introducing a 2.5GHz system and maintaining 2.0GHz and 1.8GHz system isn't a big surprise.

Also would it make sense to have systems in the line up that only differ 4% or 19% clock rate wise? Historically Apple usually has around a 40+% difference between the low end and top end CPU clock speeds for its Power Mac line (look at the Apple hardware docs if you want). The current line up falls at about 38% difference between top end and low end. So I believe this point isn't the most valid interpretation of things.

No doubt IBM had and is having issues delivering parts in the 90nm process (this is not news) and continues to be below expected yields in both volume and clock rates. They also had some issues at the beginning of the 130nm process because the Fish Kill plant was basically just getting up and running back then. It does however sound like they are catching up and getting towards the yields they want.

2) The Xserve is backordered 4-6 weeks. Apple clearly stated in the last quarterly meeting with analysts that the reason for the delay in shipping G5 Xserves was the small supply of 970FX chips from IBM.

Actually Apple expects to be caught up with the back orders in the next week or so (they may even be used some of those 2+ GHz 90nm parts to catch up on orders).

3) Apple had to include water cooling after having the G5 PowerMac out only 8 or 9 months and it is only available on a PowerMac that only runs 500MHz more than the slower models? That seems rather odd. It could be that Apple will drop water cooling if IBM is able to produce 3GHz 970FX chips in the next few months. If that is the case, then it would point decisively to overclocking of the 2.5GHz chip.

Note 500MHz is 25% more not a bad jump. Apple didn't _have_ to use liquid cooling, they _chose_ to use it because of the design intent for the systems. They want them to be quiet systems and using liquid cooling allows faster more efficient heat transfer. If they didn't use it they would have to pull more air thru existing heat sinks to insure that a great enough temperature delta existed to insure that heat transfer maintained at a sufficient rate to avoid hot spots that exceeded what the die could withstand (they need capacity in the system to deal with peak loads). They decided that they would use a 2.5GHz CPU as the threshold for this (it is also likely that had hoped to have faster CPUs in this time frame, ones that would more likely need liquid cooling to keep the system quiet and within spec, they didn't come up with this cooling is a few months). I doubt that Apple will be dropping liquid cooling when we start to get 3GHz parts, etc. It will only be dropped if Apple decides not to strive for quiet towers or they utilize a different more efficient heat transfer system. The future is about heat management and the used of liquid cooling is a good step towards dealing with the issue as was/is the air flow management in the G5 case.

The upcoming July quarterly sales meeting with analysts should answer whether my hunch is right or not. If the backorder of Xserves was only due to a 970FX chip supply problem from IBM, then it will be rather obvious that there would be a even more severe severe shortage of 970FX chips for the much higher production PowerMacs.

Fabrication of chips are done in batch runs with a little downtime between runs for adjustments, refinements and maintenance (more if they have to fix things). If prior runs had issues but those have been resolved they can have little affect on future runs. So they may have stumbled with the first set of 90nm runs that targeted in the ballpark of 2.0GHz chips. If current runs are going well (we will have to wait to see what is said by Apple and IBM) the yields could be good across the range and so no major supply issue may exist. Apple appears confident enough that next month they will have the volumes needed to ship the 2.5GHz systems, a 90nm part (given recent misses I doubt they made this statement without sufficient belief in supply).

Of course without exact tested specs (aka bining) for the CPUs being dropped in the systems some of the parts could very well be "overclocked" (all chips these days are in effect overclocked since active cooling is needed). All this means is that the CPU has to be maintained below a particular temperature threshold (factoring in hot spots) to run at a particular clock speed safely, stably and with sufficient longevity. The liquid cooling system will no doubt give Apple wiggle room to insure that temperatures can be maintained as needed. Supply issues may have pushed Apple to start using the liquid cooling system they had been developing sooner then they original planned but regardless such a system was coming.

The system Apple is using is far from the extreme over-clocking tricks used by some folks which often involves AC units and things like nitrogen baths.

Anyway my general point is that Apple appears to be playing the cards the got dealt and if they are using engineering to help make the best of it good for them. It also shows that have been forward looking enough to have such a cooling system ready (it would take many months to get something like this ready and things tooled to make it).

 

[Phinius]

Lack of chips forcing overclocking?

Isn't this fairly stardard operating procedure for Apple product line updates? Increase the speed/features on the top level machine and just bump the other machines down a level while eliminating the bottom of the line.

Apple tends to do that when there is a second round of speed bumps from a given process size. However, with a move to a smaller process size, IBM should have several 970FX speeds from 2.1GHz-2.5GHz. My question is where are those chips except for the 2.5GHz? That and the backorder on the Xserve points towards a continuing production problem with the 970FX. Apple seems to be betting on enough increase in supply for July to at least boost the speed of one PowerMac model by the use of overclocking and water cooling. If it was simply IBM achieving 2.5GHz from higher production, then where are the 2.1GHz-2.4GHz 970FX chips? There should be more 2.1-2.4GHz 970FX processors produced than there are 2.5GHz 970FX chips.

The 970FX seems to be designed as a low cost chip since the L2 cache was not increased as it normally is with a process shrink. The 90nm process size 66mm2 970FX chip is about 70% smaller then the 112mm2 Pentium 4 Prescott and yet it was about the same die size as the P4 on the 130-nm process size. That points towards a likely scenerio of Apple moving the 970FX down to the iMac and perhaps the eMac when the bigger and faster Power5 derived PowerPC 9XX is produced in the next few months.

 

[Phinius]

Apple didn't _have_ to use liquid cooling, they _chose_ to use it because of the design intent for the systems.

How do you know that Apple didn't have to use liquid cooling? It's only been about 9 months since the 2GHz 970 PowerMac first came out with a brand new cooling system that has 9 fans.

They want them to be quiet systems and using liquid cooling allows faster more efficient heat transfer. If they didn't use it they would have to pull more air thru existing heat sinks to insure that a great enough temperature delta existed to insure that heat transfer maintained at a sufficient rate to avoid hot spots that exceeded what the die could withstand (they need capacity in the system to deal with peak loads).

What makes you so sure that Apple would have to use the existing 970 heat sinks for air cooling the 970FX? Seems to me that if Apple can come up with a water cooling system then they certainly would have the ability to design more types of heatsinks.

They decided that they would use a 2.5GHz CPU as the threshold for this (it is also likely that had hoped to have faster CPUs in this time frame, ones that would more likely need liquid cooling to keep the system quiet and within spec, they didn't come up with this cooling is a few months).[/quote

It will be at least six months from the time that Apple probably knew about 970FX production problems until the water cooled 2.5GHz PowerMacs are available.

Of course without exact tested specs (aka bining) for the CPUs being dropped in the systems some of the parts could very well be "overclocked" (all chips these days are in effect overclocked since active cooling is needed).

Name one major PC manufacturer that is using water cooling with either the AMD Opteron or Prescott Pentium 4. I did notice that Sony has a desktop model that uses water cooling, but that seems to be a rare exception.

The 66mm2 970FX is 41% smaller than the 112mm2 Pentium 4 Prescott chip. Since a 3.4GHz Prescott process uses up to a peak of 127 watts, then that means a 2.5GHz 970FX would have to use a maximum of 75 watts to have the same power dissipation per mm2 as the Prescott. 90 watts is about the maximum power use of the 130-nm 2GHz 970 chip. Which means that a 3GHz 970FX would have about the same maximum power use as the 130-nm 2GHz 970 chip. Judging from the average power use figures that IBM has released for the 970 and 970FX, it would seem that the average power use has gone down considerably for the 970FX compared to the 970.

 

[jsw]

How do you know that Apple didn't have to use liquid cooling?

As has been mentioned by others, it's likely that the liquid cooling will be truly needed in future G5's, and so Apple has chosen the 2.5 as a trial system which can benefit from the technology, regardless of whether it is literally essential. That gives them some time to work out the problems.

This is just a theory. Maybe the dual 2.5 absolutely required liquid cooling or horrendously load fans. But I suspect that it's a trial run at perfecting an LCS.

And also a cool new thing to get cutting-edge Mac geeks to buy the 2.5 as opposed to the dual-2.0, which is almost as fast in real-world terms.

 

[Makyz]

So now Apple is noise obsessed, and we get expensive watercooling nonsense for 500 more mhz.

If its dual processor 2.5 isnt it 1000 more mhz?

I like 1000 mhz more.

 

[Maxx Power]

Ethernet76 reply

Now considering the following. Take 1000 people put them in a 130 by 130 foot room. Then, take the same 1000 people and put them in a 90 by 90 room. Which room is going to be hotter?.

This is completely misleading and incorrect.

Two basic problems with your analogy

1) by reducing the conductor size and distance of electron travel, you are really looking at a 90 by 90 room with half sized people, or children, if you prefer.

2) At equilibrium, assuming you have the same heatsink for each room that draws a constant amount of heat away, you have the same room temperature for each room. It will take longer for the bigger room to reach equilibrium. This is basic newton's law of cooling, solve the differential equations....

 

[Maxx Power]

Negative

If its dual processor 2.5 isnt it 1000 more mhz?

I like 1000 mhz more.

Only very very very well coded software that uses two CPU's or threads independently will you see performance increase roughly equals to 90% of 1000Mhz increase.

Dual CPU is capable of running multiple tasks more efficiently, but not as good at running one task more efficiently, for that, a faster single CPU does the trick.

Consequently anyway, for most people, dualies are a waste, more of a marketing thing than actual usefulness.

For servers though, that's another story. The fact that most people don't want to pay 3000+ dollars for a new computer which is once again, just another tool of life, something most people wouldn't get emotionally attached to, is why dualies account for a very very small fraction of the market. The same argument also is true for why apple hardware accounts for a small fraction of the market.

 

[jsw]

...Consequently anyway, for most people, dualies are a waste, more of a marketing thing than actual usefulness....

I respectfully disagree. Multi-processor machines seldom do one task much more efficiently (depending on the app, of course), but they almost always do multiple, simultaneous apps more efficiently. A dual-2.5 isn't the same as a single 5GHz one, but it's a lot better than a single 2.5, or even a single 3 (again, aassuming you don't spend most of your time running only one non-parallel app).

 

[Sun Baked]

I respectfully disagree. Multi-processor machines seldom do one task much more efficiently (depending on the app, of course), but they almost always do multiple, simultaneous apps more efficiently. A dual-2.5 isn't the same as a single 5GHz one, but it's a lot better than a single 2.5, or even a single 3 (again, aassuming you don't spend most of your time running only one non-parallel app).

I also agree that DP machines aren't a waste.

If people think dualies are a waste, then they've plain forgotten about OS 9 and it's ability to do ONE thing at a time -- try to do anything else and it crashed or acted like a slug.

OS X handles multiple threads and CPUs more effectively that OS 9 ever did.

Under OS X DP machines are great.

---

And with Power5-UL machines with SMT on the way, the virtual CPU each real CPU will make the DP machines look like Quads to the OS. (aka, gives the Power5-UL about a 35-45% boost over the 970, due to their ability to run a second thread on the CPU)

 

[Phinius]

A 3.4GHz Prescott P4 uses an average of 103 watts and the 2.5GHz 970FX uses an average of 50 watts. At 66mm2, the 2.5GHz 970FX typically uses .75 watts per mm2 and a 112mm2 3.4GHz P4 Prescott chip uses .91 watts per mm2. So, the 3.4GHz P4 uses 21% more power per mm2 than the 970FX and yet PC desktop manufactures are able to use air cooling for 3.4GHz Prescott chips and yet Apple turned to water cooling. I still suspect that Apple is overclocking the 970FX to hit 2.5GHz.

 

[Sun Baked]

A 3.4GHz Prescott P4 uses an average of 103 watts and the 2.5GHz 970FX uses an average of 50 watts. At 66mm2, the 2.5GHz 970FX typically uses .75 watts per mm2 and a 112mm2 3.4GHz P4 Prescott chip uses .91 watts per mm2. So, the 3.4GHz P4 uses 21% more power per mm2 than the 970FX and yet PC desktop manufactures are able to use air cooling for 3.4GHz Prescott chips and yet Apple turned to water cooling. I still suspect that Apple is overclocking the 970FX to hit 2.5GHz.

You are missing the actual design criteria for the 970FX which is the Maximum Power Dissipation (about 100-110W @ 2.5GHz).

Sort of like design the A/C system for a home in Arizona and using the Maximum Summer Temperature instead of the Average Yearly Temp.

If you design for the Maximum, you're living in comfort.

If you design for the average -- you'll be one sweaty human when the temps his 119°F a couple/few days in a row.

 

[Maxx Power]

This is what i said

I respectfully disagree. Multi-processor machines seldom do one task much more efficiently (depending on the app, of course), but they almost always do multiple, simultaneous apps more efficiently. A dual-2.5 isn't the same as a single 5GHz one, but it's a lot better than a single 2.5, or even a single 3 (again, aassuming you don't spend most of your time running only one non-parallel app).

I reiterate, for Most People, and by this, go to your local popular computer store, probably WorstBuy or HistoryShop, and inquire what is the fastest selling computer currently, or observe what type of people most frequently comes in and makes a purchase. They barely know the difference between a Pentium and a Computer, what makes spam and e-mail attachment viruses thrive is because your AVERAGE typical computer user is under-literate when it comes to computing. These targets of pepsi commercials and nike shoes are not your power users, they have no need for a multi-gigahurtz processor, much less, a dualie, if downloading pr0n and music is your cup of tea, what i said earlier still stands.

Of course, you are correct when it comes to people who are not the majority, for people who needs to produce things on computers that have multiple threads, a couple of CPU surely beats one. Something i can point out is gaming, on Macs. UT2004 on Macs renders its sound entirely in software, and all done on the CPU. This slows the game to a abysmal 20 fps on a G5 1.6. THis is because there is no hardware accelerated sound for Macs. On a dual 1.8, with the same ATI9600 graphics card, you are looking at doubling in frame rates because the game allocates one of the CPU's to do sound processing and the other for game AI, networking, etc....

PS. Try UT2004 on a PC with hardware 3d sound/EAX, it sounds so much better, even with just two speakers.

 

[Maxx Power]

I also agree that DP machines aren't a waste.

If people think dualies are a waste, then they've plain forgotten about OS 9 and it's ability to do ONE thing at a time -- try to do anything else and it crashed or acted like a slug.

OS X handles multiple threads and CPUs more effectively that OS 9 ever did.

Under OS X DP machines are great.

You are correct about the OS 9 situation. I never liked OS 9. Compared to OS 9, i much prefer (don't flame me for saying the truth) windows xp or windows 2000. The windoze series have been able to multi-task efficiently since, i guess, doze 98 ? But many other aspects of that os is pure crap, so let's not get into that.

Regarding dualies, please read my other response a few replies up. To multitask efficiently, you really only need a capable OS, I also regularily use a Doze EKS-PEE machine, and this thing multitasks beautifully on one CPU. I can encode MP4's while watching an episode of this or that, or have some DVD encoded in the background while I listen to music and browse macrumors and share my music online legally (i'm in Canada, it's legal by law) at the same time.

 

[wdlove]

Who's to blame for all this? Powermac G4 owners. They whined and bitched constantly about a bit of fan noise. I had a DP G4 PM, and it wasn't that bad. So now Apple is noise obsessed, and we get expensive watercooling nonsense for 500 more MHz.

I want a loud-as-hell 3.5ghz dual G5. Bring it on!

I'm not one of those that complained. Don't think that the noise is a problem at all.

If as some are mentioning the overclocking of the G5 ship, will that shorten its life or cause problems in other ways.

Until Apple says so, I'm not sure which camp is correct.

 

[shawnce]

How do you know that Apple didn't have to use liquid cooling? It's only been about 9 months since the 2GHz 970 PowerMac first came out with a brand new cooling system that has 9 fans.

So? They designed and shipped a system which matched their design criteria (their solution was to divide the system into independent zones with independent adjustable rate fans, etc.) and around that same time frame (if not before) they likely already had a liquid cooling system starting to be worked on, to build on this thermal management. They saw the road ahead. In fact the cooling framework setup in the Power Mac G5 chassis and now this liquid cooling is well positioned for the future.

As you state, others don't use liquid cooling... yet they arguably use CPUs with near equivalent and often not greater power dissipation densities then the 970FX (or 970) and with greater over all dissipation per CPU. So if they don't have to use it then Apple didn't have to use either but because of Apple's stronger desire for quiet systems Apple is taking extra steps (just go look at the various silencing products that exists in the market for PC rigs).

What makes you so sure that Apple would have to use the existing 970 heat sinks for air cooling the 970FX? Seems to me that if Apple can come up with a water cooling system then they certainly would have the ability to design more types of heatsinks.

This presumes a more efficient heat sink could be designed and fitted then what is already in the PowerMac G5s while still maintaining one of the design goals of a quiet system. It also presumes that such a system, if possible, would be cheaper then using a liquid cooling system (thermoelectric cooling device comes to mind but it still requires more efficient heat sinks to make that work since thermoelectric cooling devices contribute heat themselves as the work). Anyway I fully believe that Apple could have used air cooling for the 2.5GHz, even if they are "overclocked" as you imply, but it likely would have required additional air flow and hence noise (IMHO not much in the case of the 2.5GHz 970FX). Also note that man in the PC overclocking crowd do so without having to use liquid cooling, they often can do it by adding more directed air flow.

It will be at least six months from the time that Apple probably knew about 970FX production problems until the water cooled 2.5GHz PowerMacs are available.

6 months is a rather short time line for the design, testing and tooling of a system like this. I personally doubt this was a reactionary solution as you imply.

Name one major PC manufacturer that is using water cooling with either the AMD Opteron or Prescott Pentium 4. I did notice that Sony has a desktop model that uses water cooling, but that seems to be a rare exception.

Name one that is using the divided zone cooling to the extent seen in the PowerMac G5 chassis (yeah rack mount stuff). Also name one that is as consistently concerned about making quiet systems as Apple currently is.

The 66mm2 970FX is 41% smaller than the 112mm2 Pentium 4 Prescott chip. Since a 3.4GHz Prescott process uses up to a peak of 127 watts, then that means a 2.5GHz 970FX would have to use a maximum of 75 watts to have the same power dissipation per mm2 as the Prescott. 90 watts is about the maximum power use of the 130-nm 2GHz 970 chip. Which means that a 3GHz 970FX would have about the same maximum power use as the 130-nm 2GHz 970 chip. Judging from the average power use figures that IBM has released for the 970 and 970FX, it would seem that the average power use has gone down considerably for the 970FX compared to the 970.

Well lets look at things this way using numbers from IBM... a 970FX running at 2GHz typically uses 24.5W while a 970 running at 2GHz typically uses 56W (likely a little more but I am doing a linear scaling of values from a 1.8GHz part). Now for max power lets double things, so 49W 970FX at 2Ghz and 112W 970 at 2GHz. The 970FX has a die size of about 66mm^2 and the 970 has a die size of about 118mm^2 (note the chip package is the same size so the heat sink attachment has the same surface area). That yields power densities of 0.74 for the 970FX and 0.95 for the 970 at 2GHz (note the power density for the 970 is 28% more then that of the 970FX at the same clock rate). One factor for the radical power difference is the switch to 1 volt for the core logic, which is thanks to the process shrink as well.

So the existing air cooling system is at least capable of dealing with dissipating 112 W overall while maintaining a rapid enough heat flow to deal with a power density of 0.95 (per CPU). First thing to note swapping in a PPC 970FX for a 970 at the same clock speed in no way requires a change in cooling (why you don't see the liquid cooling across the board and I do believe we will see / are seeing the 970FX across the board in the new systems because they will be cheaper however in the short run they may be using 970 for the 1.8 & 2 GHz versions to clear inventory).

Now what does a 2.5GHz 970FX burn? Well I don't know (no numbers from IBM) but we can likely bound it. The low end can be guessed at by linear scaling of the power dissipation as you vary the clock rate (for CMOS it generally scales linearly with clock rate). So the low end estimate is 62W. Now a high end estimate can be done by assuming that the 2.5GHz 970FX requires the use of 1.3V for its core voltage (the higher core voltage may be needed to insure enough current flow to allow switching at 2.5GHz). That would put the power dissipation at about 105W (for CMOS it generally scales to the square of the voltage). One final guess would be to note the delta in dissipation between the 1.4GHz (12.3 typical) and the 2.0GHz 970FX and assume that delta scales linearly with clock rate. That comes out to 85W and that is just above the average of the other two numbers. I personally believe it is in the 80s.

So the low guess puts the power density at 0.93, the mid at 1.29 and 1.6 for the high estimate (in one of your posts you seem to believe 50 W average which can be guessed to imply about 100W max or a density of 1.51... which is more then the P4 Prescot at 3.4GHz but anyways...). So if the power is at the low end then the existing cooling system is sufficient without additional airflow. If it is towards the mid (I believe it is around the mid) or high estimate then the existing cooling system may not be sufficient (who knows exactly what it tops out at) and hence an alternate system would be needed if one wants to avoid requiring additional air flow (which obviously Apple is trying to avoid). In other words a good enough reason exists to utilize liquid cooling for the 2.5GHz CPU to help keep the system quiet by improving its ability to transfer heat to air (the liquid flowed radiator will be far better at it then a passive heat sink) that one doesn't have to invoke a reasons such are overclocking (however recall my prior wiggle room comment).