Apogee from Swiftech...

[Marci]

There are TTV-to-CPU correction formulas that can be applied to all TTV-obtained results to produce die results, and those exist for EVERY cpu on the market (AMD and Intel). The correction factors account for heatflux differences etc etc. Problem is, I think Intel are the only folks that KNOW what those correction formulas are for each TTV - and with never being in the position of acquiring one, don't know if they share that info with the TTV purchaser etc...

If correction formulaes for each CPU bed-type (eg: Socket A/ Sempron bare die, A64 IHS, A64 Dual Core) are applied and graphed, and a simple confirmatory test done to see whether realworld matches corrected TTV data, then battle is solved... yes? Provided results become more useful... (?)

Logic would say Swiftech could easily contact them and get this info and apply it (but likely won't)... which would give the customer something more of use to base comparison from...

If not, p'raps attempts need to be made to get it... (in progress already)

[Les]

Finding even a simple analysis a can of worms

[bobo5195]

Ah Les, surely a man of your talents could work out a solution to the fourier equation for an IHS + heat die. Might want to try STRAND7, its an FEA (doesn’t do fluids) package the demo goes up to 100 elements which might just do for your tasks

There is no true (as in perfectly true) calibration factor for cpus as there is two sets of data that need to be know. The absolute heat flux and the distribution of this heat flux. Give that the dim sims mentioned on this board are based on constant surface temperature / heat flux distribution there is going to be some difference that may or may not be significant. The distribution of the heat energy “flow” is a unknown for cpus and will probably remain that way. One solution might be to ignore it and say that heat distribution to the WB or a at least a short distance inside it becomes constant. To model the increased resistance because of the longer heat path some kind of constant can be used.

Generating this constant other than clicking the random number generator; guesstimation in your head might be a little tricky. The proper solution would be to look at statistical distributions in real cpus and pick an appropriate value. This might help with heat distributions as well if you could include them. Further more with knowledge about cpu heat output distribution you could include that as well. I have a feeling that intels formula is empirical using stats from a large number of cpus. Get some chips test the output, put a resistance in. From these results find a conservative value for what the factor should be(or indeed the form of the factor. Could say that thermal paste thickness is increased by X or ideal resistance multiplied by some number).

Obviously DIYers can’t do this. A statistical model using FEA could easily be developed with perfect knowledge about the manufacturing (FEA all possible paths and get a result). We don’t have a perfect data but we can guess.
The unknowns are:
CPU total heat output – Guess able or non dimensionalise.
CPU heat distribution – hard. Constant heat flux might be appropriate. I think being a little smart and saying the cache is cooler than the processing units might be valid.
Distribution of thermal compound – This again is hard. We’re going to need some BC’s here to help.
Geometry of HIS – this is obviously some manufacturing variations, indeed are we even going to say the same about the core (given the process for making a core I would say that saying its perfectly flat and level is a suitable assumption).
Material properties – some silicon to model the core, thermal paste is doable, I H S is doable as well. Interface properties if we chose to include them might be hard.
Block properties – Do we include the interface there, I assume just use Bills data, but the distribution might prove troublesome.

My suggestion would be:
Assume we’ve got material properties and say the I h s conforms exactly to available data. Data on heat output from cpus is around. The temperature distribution can be made up by some guy with a beard (preferably long, with pizza in it :P), even have it as a changing distribution(ie manufacturing difference between cpus, could do further test to show influence on OCing etc. Thermal paste distribution is then the troublesome bit. We can however say what the limiting value is as we can guess when the core can shut down due to heat with the bare minimum craptastic heat sink. We can also guess what perfectly applied thermal paste would look like. Easy option then is run FEA on this millions of times to show what a random distribution of the variables would look like. This then gives you an approximate model of heat distribution on the surface of the I h s (so you can tell if its approximately constant or not) and a good GUESS at thermal resistance of the rig. This kind of procedure of empirical FEA is carried out by a few companies to check that parts as a group won’t fail etc. It will take awhile however on computers. Its not SETI but a good evenings work and computer time.

[Albigger]

http://www.systemcooling.com/swiftech_apogee-01.html

[jaydee]

Originally Posted by Albigger

http://www.systemcooling.com/swiftech_apogee-01.html

Subject: Liquid Cooling
Date: Novermber 05, 2005
Author: Lee "RoboTech" Garbutt
Source: Swiftech

Posted for a month already? Err, the month not only is wrong but spelled wrong to. Glad I don't care about spelling.

Interesting page: http://www.systemcooling.com/swiftech_apogee-13.html

[Les]

Good review Lee.

[nikhsub1]

Seems to me Swiftech should rename the block - Apologee?

[Incoherent]

Another nice review Lee.

Rather damning.

[Les]

Bobo,Strand7 appears to be 21st century and I don't go beyond 20th. I will leave the field clear for you

Originally Posted by Incoherent

.....Rather damning.

The only "damning" I see in the review is the inadequacy of present techniques That is simple uniform-flux heat-die measurements,SpeedFan 4 reported temperatures and side-of-IHS temperatures.
Bronze Age shows promise if coupled with 21st century analysis.

Models in attachment are all single TIM inclusive

Edit: Removed "alien point" from graphs

[Incoherent]

Originally Posted by Les

... simple uniform-flux heat-die measurements,SpeedFan 4 reported temperatures and side-of-IHS temperatures.

Agree, all are flawed. Not to fault Robotech, best possible analysis under the circumstances I'd say.

Originally Posted by Les

Bronze Age shows promise if coupled with 21st century analysis.

Think so too, becoming more and more aware that understanding these problems/behaviours are impossible without good models. I might want to move into early iron age with the hardware though. Some ideas germinating.

[Les]

Originally Posted by Incoherent

..1) Not to fault Robotech, best possible analysis under the circumstances I'd say.
2)Think so too, becoming more and more aware that understanding these problems/behaviours are impossible without good models. I might want to move into early iron age with the hardware though. Some ideas germinating.

1) Absolutely
2) Will probably have a look at free thingies like Strand with view to application to an Iron Age die.

[Roscal]

Originally Posted by Albigger

http://www.systemcooling.com/swiftech_apogee-01.html

And what I was saying in the start of this post, eh ? Some heavy considerations must be taken here using dies in comparison of true systems and Lee shows well this fact. Reactions are (and will be) always very disparate and say "WB blabla is better" don't mean nothing without giving testbed conditions (far from reality or not)... But some are always reluctant to understand, I see . Data and more (controled) data from others are a necessity to enlarge perf view.

Nice job Lee !

[Albigger]

*deleted <brain not awake> *

[Les]

Originally Posted by Albigger

Based on Robotech heat die data, average effective convection coefficient can be graphed. Area used was area of the waterblock based on my modeling inside a 32mm or 14 mm square (i.e. assumes no heat spreading through IHS or baseplate bottom).

No
1)spreading is a most important factor.
2)missing (a) TIM (b) the influence of furniture area .

[Albigger]

Originally Posted by Les

No
1)spreading is a most important factor.
2)missing (a) TIM (b) the influence of furniture area .

1) agreed. not saying it isn't.
2) (a) true. was not attempting to model at this time
(b) the rest of the block?


Edit: realized NEED to model TIM based on temp taken (should get some sleep before posting). Will delete erroneous post above.

[bobo5195]

Once i learn STAR CD or ANSYS CFX (commercial high end/ moderately high end CFD), does that mean that i will then move into Future tech? :P

OR hatch a cunning plan to start rapid prototyping blocks in my bedroom via mail order. Storm G7 alike with less pressure drop (3d porting) and slightly lower performance (copper ploymer matrix, only moderate thermal conductivity) for half the price (lloking at maybe $100 a block to manufacture) with your name and maybe something funny in 3d on the top (skull or something as that would be well l33t).

[Les]

Originally Posted by Albigger

.. (b) the rest of the block?...

I only consider the cup base and cup wall area:-
My estimates
Effective Cup Base Area=330.8391314 mm^2
Effective Cup Wall Area=1694.330101 mm^2

Consider separately to apply "efficiency factor" to wall area.
See Post330

Edit That is the Storm G4 Model
Same applies to Apogee but use Pin Area and "Channel base" area.
Apogee
Total Pin Area=5936.4 mm^2
Channel base Area=672.0 mm^2
on 1141.2 mm^2 's square of bp

[bobo5195]

doesnt quite work like that for storm, area is a non-linear function of performance.

Send me an email via pm, i think i can show you how to model it near precisely (stew has a few performance tricks up his sleeves). You better know how to do geometric scaling though (have data for 4mm jets).

[Albigger]

Thanks. will take a closer look at previous post.


re: pin area and channel area I have apogee modeled if you need them. My numbers (mostly) agree with the ones Lee quoted in review, although he says .9mm channel and I measure ~1mm (will have to verify with gauge pins when I can stop by work again).

[Les]

Originally Posted by bobo5195

doesnt quite work like that for storm, area is a non-linear function of performance.

Send me an email via pm, i think i can show you how to model it near precisely (stew has a few performance tricks up his sleeves). You better know how to do geometric scaling though (have data for 4mm jets).

Never suggested "area" was a "linear function of performance".
Post Model in thread,several wish to learn.

As you should have gathered(eg here and link within) I only modelled the Storm as a last resort because nobody else would.

Edit: Added link

[bobo5195]

Since i just read it i thought that you might as well use this les. Some one modeled a storm type blocks heat transfer (using the nusselt number as)

(Nu / Pr^0.53) * (D/d)^2 = 32.49 * Re^0.523 [jet impingment effect] + 0.266*(D/d -8) * Re^0.828 [wall region]

This is for gases and high Re's (2000) but it shows a model of the effect.

[Les]

Show models, both the "unnamed some one" and your own.

[bobo5195]

k, but i dont have my own, as im not that good.

source for the one above is Sitharamayya S; Raju KS (1969) Heat transfer between an axisymetric jet and a plate held normal to the flow.

The distance between the jet and the plate surface is not given so its a little simple. Probably should use DIYer units instead of scientists.

[Orkan]

My own testing shows the apogee performing quite well on my X2. Only problem;The weak ass top. After seeing it X-sectioned... I'm pissed. No way, can that thin of plastic support any type of load.

I'm afraid to push my hoses on.

[bobo5195]

The thin point is at a stress concentration (a corner) as well, they could of easily curved that section and got rid fo the problem. If they've doen the injection moulding dies now then they can't really change it.