Convection Coefficient: A Fanstasy in h

[Groth]

Following some recent talk about coldplate geometry and other's simulations, I've decided to work on a few idea that have been sitting in the back of my mind.

The equations for thermal and electical conduction are analogous, so it ought be possible to use the readily available software for electical simulation and turn it to the task of thermal simulation. To that end I've been assembling arrays of resistors, capacitors, and voltage and current sources. In simple tests, like uniform heat flow through a rod, or a point source on an infinite plane, simulated results match expectations.

So now it's time to take it up a notch.

What I'm looking for is estimates/theories/guesses for h, for a fairly uniform waterblock. Something like a #rotor block or the MCW5000A. I want to see how a coldplate-TEC-waterblock system reacts to changes in coldplate geometry, and how the system reacts to heat transients.

Off topic -- I've been using LT-Spice. At 8:10 this morning, I sent and e-mail to Linear reporting two bugs. At 3:30, they sent a reply indicating that problem was solved and updated binaries were available.

[Incoherent]

Originally Posted by Groth

What I'm looking for is estimates/theories/guesses for h, for a fairly uniform waterblock. Something like a #rotor block or the MCW5000A. I want to see how a coldplate-TEC-waterblock system reacts to changes in coldplate geometry, and how the system reacts to heat transients.

I'm currently working on a model. One version uses these formulae. I am not certain about them actually, but it does match what I have been generating myself, fairly well so I am accepting it for now.
The problem is, fitting theory with pHaestus' testing gives values for h which are very high in my opinion. For the Cascade and MCW6000 we are talking values between 50,000 and 150,000W/m^2*C depending on flowrate. This feels wrong to me. But Cather has demonstrated numbers well over 100,000 in his testing so I don't quite know what to think. For the MCW5000 and #rotor blocks the numbers are in the region of 10,000-30,000, showing that they are a fundamentally different type of block.
Guesses all but somewhat grounded in the region between reality and theory.

Incoherent

[Groth]

Good stuff, keep it coming.

I'm not going to even think about the Cascade or MCW6000 yet. Too much point to point variation in h, requiring much more model detail than I'm ready for.

[Les]

Originally Posted by Incoherent

....
The problem is, fitting theory with pHaestus' testing gives values for h which are very high in my opinion. For the Cascade and MCW6000 we are talking values between 50,000 and 150,000W/m^2*C depending on flowrate. This feels wrong to me. But Cather has demonstrated numbers well over 100,000 in his testing so I don't quite know what to think. For the MCW5000 and #rotor blocks the numbers are in the region of 10,000-30,000, showing that they are a fundamentally different type of block.
Guesses all but somewhat grounded in the region between reality and theory.

Incoherent

A couple of thoughts.
I guess you are referring to "h(eff)" and not "h (convection coefficient)"
This is very similar to "h(bp/Fin Interface)" that I have used, for example, here.
In this case I do not disagree with the numbers IF "C/W(TIM)"=0.1
The suggestion that "C/W(TIM)"=0.1 is very iffy
It is based on the unsatisfactory suggestion that "tc offset in the bp ~ 0.001c/w " in the measured "C/W(TIM)" ~0.15c/w obtained by Billa
The most inadequate analysis is :-
"The 0.15c/w consists of three components:
1) The tc offset in the die ~ 0.05c/w ( 2mm thick, 10x10mm Xsection
channel)
2) The tc offset in the bp ~ 0.001 - 0.05 depending on the the amount of
heat-spreading.(2mm thick, 80x50mm Xsection channel(surge) - 2mm thick,
10x10mm Xsection channel)
3) The Die/Goop/Bp interface."
This has not been modified since off-line discussions following this.

Changing "tc offset in the bp ~ 0.001c/w" to "tc offset in the bp ~ 0.02c/w" would ,perhaps, mean "Sieder -Tate calculated h(convection coeff)" values are adequate :
From here


The "Side to Side" uses "Sieder- Tate"

[Groth]

My brain hurts whenever I try to fathom your posts, Les.

As for h(eff), I could tolerate an effective value for the fin/pin as a whole, but when it includes the baseplate it loses value, as the lateral heatspread of the baseplate is included/integral to my modeling.

Off to read about "Sieder- Tate"....

[Les]

Originally Posted by Groth

As for h(eff), I could tolerate an effective value for the fin/pin as a whole, but when it includes the baseplate it loses value, as the lateral heatspread of the baseplate is included/integral to my modeling.

Off to read about "Sieder- Tate"....

To my understanding h(eff) does not include the baseplate.
My "h(bp/Fin Interface)" is calculated using a 0.1mm bp.
It uses Kryotherm to connect "h(bp/Fin Interface)" and "h(conv. coeff)
Kryotherm uses Sieder-Tate

[Cathar]

I can't lay down any really solid evidence for it other than my emperical testing aligning with projected theory.

My best estimate for TIM(C/W) for a ~100mm^2 die is actually around 0.06-0.07 for a fully cured interface. 0.07 is what I tend to use for average h "predictions". I am having very great troubles rationalising that TIM(C/W) can be as high as 0.08 at all, except for an immaturely cured mount. I base this around modelling effectively infinite values of h for very simple designs that I can emperically measure and using that to back-correlate to what TIM(C/W) would need to be to fit the proposed model. At 0.06-0.07 is where it all starts to make sense in my private world.

However, I do use somewhat higher mounting pressures that perhaps is standard, typically being at around 16kgs (~35lbs) for said 100mm^2 area, and that may very well explain my observations.

[Incoherent]

Originally Posted by Cathar

. At 0.06-0.07 is where it all starts to make sense in my private world.

That is around about what I have measured too. If not a bit lower. I am almost certain that pHaestus is getting even lower than this in his setup. Practice perhaps.

h eff. Yes indeed Les, I am talking only about the effective coefficient.
As far as "The tc offset in the bp ~ 0.001 - 0.05..." goes, I have measured a value for copper of, from memory, 0.0245 ish C/W/mm*100mm^2. It should be 0.0255C/W/mm for a parallel heat channel with xsectiom 100mm. With spreading effects, lower?

Where oh where is that magic formula for effective "h "? with the simple terms flowrate, surface area.... The dimensionless variables are blowing my neural circuits.

[Les]

Originally Posted by Incoherent

.......
As far as "The tc offset in the bp ~ 0.001 - 0.05..." goes, I have measured a value for copper of, from memory, 0.0245 ish C/W/mm*100mm^2. It should be 0.0255C/W/mm for a parallel heat channel with xsectiom 100mm. With spreading effects, lower?
....

Yes, according to my thinking.
There is a change in dimensions of the Flux Channel at the TIM/wb interface.
The c/w between the TIM/wb interface and the tc(wb) would be that of the larger Flux Channel of the wb with infinite heat-spreading(zero heat-spreading resistance), ie ~ 0.001c/w for 2mm of a 80x50mm(dimensions of Surge wb) Flux Channel.
For the case of zero-spreading c/w~ 0.05 for 2mm of a 10x10mm(Die dimensions) Flux Channel.

Edit: I do like your C/W(TIM) = 0.063c/w for 11x10mm CPUs

.

[Cathar]

Originally Posted by Les

Edit: I do like your C/W(TIM) = 0.063c/w for 11x10mm CPUs

.

Well that does basically agree with my 0.07C/W value for 100mm^2 CPU's that I've arrived at over the last couple of months.

Originally Posted by Incoherent

If not a bit lower. I am almost certain that pHaestus is getting even lower than this in his setup.

Hard to tell really. I (firmly) believe that there is some degree of temperature compression going on in pHaestus's test CPU, but that this is out of his control to compensate for and peculiar to the individual CPU he's using. This is a bit of a tussle that Derek and I have wrangled over privately. It's not important with respect to the differentiation of block performance though, just that I feel that the scale of the Y-axis is compressed to a certain degree.

[Les]

Do you mean incorrectly calibrated?
Or some obscure meaning?
Have been impressed with pH's calibration and would be interested in any flaws.

[pHaestus]

I am not convinced it's a traditional temperature compression either Les; I think it's more that the CPU is in thermal contact with the motherboard's traces and all that associated copper kinda smooths things out (doesn't show wbs performing as badly as Bill's die sim at really low flow rates and doesn't show wbs performing quite as well at high flow rates).. One could alternately conclude that the diode is in a "cool spot" on CPU and that this results in temperature compression.

The calibration's fine but what is the calibrated diode measuring? Is I think the major question

[Cathar]

Originally Posted by Les

Do you mean incorrectly calibrated?

No, not meaning that at all. Have been most impressed with Phaestus's calibration work.

Meaning that the readings are peculiar to the CPU (and its diode) itself, and to a lesser extent the effect of the motherboard.

The 11.33x7.37mm (83.5mm^2) T'Bred B die, with the calculated ~75W heat load producing ~9-10C delta's resulting in calculated C/W's of around 0.12-0.13 seems to me to be a bit low, and has nothing at all to do with Phaestus's exemplary calibration work. Would personally feel more comfortable with derived C/W's of around the 0.20 mark given those dimensions. Even accounting for secondary heat paths, I am still grappling with accepting that the motherboard's secondary heat paths could be half as effective again as the waterblock is at removing heat from the CPU.

[Les]

Originally Posted by Cathar

Well that does basically agree with my 0.07C/W value for 100mm^2 CPU's that I've arrived at over the last couple of months.

.

Have been getting a bit lost with AMD die sizes and Flux-block size.
Re-reading Incoherent's work, I now read as using an "Athlon XP 2500" with 7.8x14mm(109.2sq mm) Flux-block.
This would give TIM area as ~ 83 sq mm; not ~ 110 sq mm.
Becomes most confusing

[Incoherent]

Originally Posted by Cathar

The 11.33x7.37mm (83.5mm^2) T'Bred B die, with the calculated ~75W heat load producing ~9-10C delta's resulting in calculated C/W's of around 0.12-0.13 seems to me to be a bit low...

One thing could be the calculated 75W is too high.
The heat capacity of the coolant used in the power calculations may not be right. I don't think the glycol additive (25%?) is being accounted for. Should be 10% ish lower I think. Could be wrong, pHaestus?
Doesn't account for all of it though (or even much of it actually), but it can not be the secondary losses, the power calculation has already accounted for them. (Tin-Tout)*Flowrate*Cap. H2O ? only measures the heat moving through the block.

[Cathar]

Originally Posted by Les

Have been getting a bit lost with AMD die sizes and Flux-block size.
Re-reading Incoherent's work, I now read as using an "Athlon XP 2500" with 7.8x14mm(109.2sq mm) Flux-block.
This would give TIM area as ~ 83 sq mm; not ~ 110 sq mm.
Becomes most confusing

Barton die (512K L2 Cache) is 13.61 x 7.42mm =~ 101mm^2

Page 48 of this document

[Cathar]

Originally Posted by Incoherent

but it can not be the secondary losses, the power calculation has already accounted for them. (Tin-Tout)*Flowrate*Cap. H2O ? only measures the heat moving through the block.

Indeed, you are quite correct, so the ~75W is being calculated after secondary heat losses.

[pHaestus]

'tis true I use 25% ethylene glycol mix. I favor 71.3W not 75 too

[Groth]

Any other [on topic] thoughts?

As it stands, I'm looking at building 5000 model. Starting with a Sieder-Tate based h and tweaking it to better fit published pH, JoeC, and BillA data.

[Les]

To fit pH's data one first have address pH's comment "The calibration's fine but what is the calibrated diode measuring? Is I think the major question" Relation between the temperature at the, probable, cold-location of the diode and that at the Hot-spot(s) is elusive(tried, briefly reported here )

Would probably restrict to fitting to Billa's and JoeC's data

[Groth]

Yeah, I'm aware of the uneven heat production and unknown thermal diode location. Pesky, but not necessarily insurmountable.

I've never been completely satisfied with JoeC's testing, so I hate to give it too much weight.

As always, BillA reigns supreme.

[Groth]

I think I need more memory or less model detail. My hard drive LED hasn't blinked off in the last 18 hours....

[Incoherent]

Hows the model going Groth?
Still crunching?

[Groth]

At 37 hours, it locked up solid.

Started again with half the detail, it finished in 16 hours and then hung trying to display the results. Data was all there, but in a compressed format I don't understand.

Started again, remembering to set it to 'ASCII data files' and 'don't compress results' (a setting it refuses to remember). Was playing with the case fans (summer, 35C room), bumped the wrong wire, broke my raid-0 stripe.

Fixed my stripe-set, repaired the injured files, and I'm running it once again. <sigh>