S-TDX water block or What?

[Groth]

Yeah, the analogy is strained and the application is off. On the temperature profile picture, the lines through the copper an aluminum sections would be straight (if no lateral heat spreading was modelled), and would be curved in the opposite direction if spreading was modelled (greater dT per length nearer the die where the heat flux is greater).

[wijdeveld]

Ah well, strained or not; the model predicted the impact of flow behavior through the waterchill block within 0.5 oC from 0.1 to 10 times the reference flow (which is 109 l/hour and not the maximum stated Hydor L20 flow of 700 l/hour, also correctly predicted by the model) and was right on spot in predicting the effect of a R404a filling at different loads for the Vapochill SE almost a year before it came on the market (only the vapor properties were changed and no new calibration was made).

But no need to trust a model, I don't trust any model myself, only the (in)capability of the modeler .

It was/is usefull to describe a lot of proceses going on inside a cooling setup but I'm currently at the limit of this particular model since it doesn't include bed roughness, eddies&turbulance and non hydrostatic pressure impulses. So, looking at heat dissipation will require a new model setup, but I'm working on it .

By the way, the curving of the model is right since it is a finite element model, meaning that the dC/dX slope is dependant on different conditions for each element.

P.S. in post #60 from this thread the spreading was modelled in a 2-D (y-z slice with fixed x coordinates) model.

[Groth]

Originally Posted by wijdeveld

By the way, the curving of the model is right since it is a finite element model, meaning that the dC/dX slope is dependant on different conditions for each element.

Heatflow is directly proportional to the temperature gradient dT/dz. Your model gives a higher gradient as you move away from the die, implying greater heatflow in areas away from the die. Where is that heat coming from? You have either disequilibria or are generating heat inside the heatspreader and waterblock.

The curves may be right for the model, but neither the curves nor the model are right for reality.

Another question, did you ascribe the same properities to both TIM joints, or are they modeled on two different compounds?

[wijdeveld]

Originally Posted by Groth

Heatflow is directly proportional to the temperature gradient dT/dz. Your model gives a higher gradient as you move away from the die, implying greater heatflow in areas away from the die. Where is that heat coming from? You have either disequilibria or are generating heat inside the heatspreader and waterblock.

The curves may be right for the model, but neither the curves nor the model are right for reality.

Yes, you're right, I've just rechecked the input and this play model is from just before the introduction of 'heat' in term of production: I've just checked the diffusion profile based on a huge initial load in the 'die' segment. I'll upload the 3-d mesh figure; you can see the slow concentration decrease in the first segment, explaining the curve at the die to TIM/IHS interface. Sorry for the confusion, it was a long time ago, the model had a very limited scope (checking of influence of TIM) and was way before the much more detailed waterchill modeling

Originally Posted by Groth

Another question, did you ascribe the same properties to both TIM joints, or are they modeled on two different compounds?

Yes, but the first TIM layer (die to IHS) was only for completeness. Later on, I've made a 2-D schematic to calculate the impact of a less perfect solid to solid interface. As you'll know this has more impact then the exact composition of the TIM.

Model assumptions:
Model for temperature diffusion simulation (from CPU die to cooler)
Assumptions:
thermal conductivity copper: 380 watt/(m.K)
thermal conductivity aluminum: 235 watt/(m.K) (0.62 x Cu)
thermal conductivity Artic Silver 3: 9 watt/(m.K) (0.02 x Cu)
thermal conductivity normal TIM: 4 watt/(m.K) (~ half of ArticSilver)
(0.01 x Cu)
Thickness of layer:
1) die: 0,005 (one layer in model)
2) TIM layer die -> ALU heat spreader: 0,025 mm (0,001 inch)
3) alu heat spreader: 1,0 mm
4) TIM ALU heat spreader -> cooler: 0,025 mm
5) Copper core: 2,0 mm
6) cooling fluid: 0,005 (one layer in model)

Layers of: 0,005 mm: 102 layers in total
-1) die = boundary -1
1) TIM = 5 layers
2) ALU = varial thickness of layers, total of 40
3) TIM = 5 layers
4) Vapo = varial thickness of layers, total of 50
-2) fluid = boundary -2

[wijdeveld]

Ah well, I love to discuss models and there imperfections, but now a question from my side.

Has anybody visualized the eddies/turbulence at the solid/liquid interface and calculated the impact of such small scale water turbulence on the heat dissipation? I’ve some test equipment which uses a laser sword (2D light screen of just o.5 mm thickness) for large scale visualization of eddies and a laser Doppler for x-y velocity measurement (and used for calculating the Reynolds stress), but both are too crude to use for measurement between pins in a water block (mm scale). Just curious

[BillA]

far FAR outa my league, both
$$$$$$$$$$$$$$$

[wijdeveld]

Sorry Bill, it's good that this forum has a wider mix of expertise then most other forums. Modeling is just a tool and as such only useful to gain some insight in how and why performance between designs differ and can be improved.

I think Asetek made me a mod (I’ve absolutely no other bonds with Asetek then some unpaid ‘free time’ projects and giving some user assistence as a mod!) because I was loosing interest in most of the forum posts, not much new to learn in there right now (still waiting for the Bowman review on the new evaporator).

So keep developing and testing and some of us will try to find an explanation behind your excellent water block performance

[BillA]

oh I know the 'why' well enough, its the quantification thats a bear
thats why I'm called an experimentalist (label by an experienced CFD modeler, lol)

[Groth]

I wish I had your equipment. Closest I've seen your suggestion is some papers looking at vortexes and such in impingement situations. Never seen between the pins and fins stuff. You might want to check into temperature sensitive encapsulated liquid crystals.

[Incoherent]

Just a chart to illustrate the point:



From measured data. The delta T of the block is greater than it would be if there was no "barrier" to the water.
The TIM is significant. (from this data, 0.06C/W)
Wijdeveld, keep modelling, it's very interesting.

Cheers

Incoherent

[BillA]

Incoherent
how did you measure the temps at the various interfaces ? (I count 4 of interest)

[Incoherent]

Bill, they are extrapolated. The solid blue points are measured data averages.

There is some uncertainty due to the unknown thickness and conductivity of the silicon die. The gradient CPU-TIM1 should (probably) be steeper, making the TIM C/W even less. The CPU side TIM C/W has been applied to the WB side TIM, might not be the same, but I believe it is in the same ballpark.

Cheers

Incoherent

[BillA]

bah
change the graph title
you did good work, but present it correctly
grrrr

[Incoherent]

lol, sorry Bill.
I'll do that tomorrow, bedtime for me.

[wijdeveld]

Originally Posted by Incoherent

From measured data. The delta T of the block is greater than it would be if there was no "barrier" to the water.
The TIM is significant. (from this data, 0.06C/W)
Wijdeveld, keep modelling, it's very interesting.

Cheers

Incoherent

Nice, more data
I know TIM is important, but after some model testing I don't believe the exact compound used is that important. It looks like that increasing the contact area between both solid interfaces by using molecules of different sizes (for better stapling) is much more important. I’ve borrowed a graph from dpanter:

[Incoherent]

Originally Posted by wijdeveld

...after some model testing I don't believe the exact compound used is that important. It looks like that increasing the contact area between both solid interfaces by using molecules of different sizes (for better stapling) is much more important. I’ve borrowed a graph from dpanter:

Have you been modelling this? Very interested in seeing the data if so, it is one of the big parts of the ΔT, it would be nice to see how far it is possible to reduce it. I tend to agree on the idea of compound being somewhat irrelevent, having seen some results using motor oil or even water not substantially different from say Arctic Silver. Do you mean particles or actual molecules?

[wijdeveld]

What you can see in #67 is that the slope between the die, TIM and copper block is fairly fluent. As can be seen under the assumptions (#97) the thermal conductivity of normal TIM is ~0.01 x that of Cu.

So, from a non restricted heat transport point of view (perfect die contact and no heat dissipation limits) the TIM compound doesn’t influence the performance that much.

I’ve tested the use of AS3 versus the default goop versus no TIM (ah well, didn’t clean the die that well, just a tissue and some acetone) to evaluate the load behavior of the Vapochill PE (temperature on y-axis = CPU temperature). This time no calculations but measurements

[wijdeveld]

I've done a series of model calculations in a slightly more advanced 2-D model (including heat spreading) to take into account the possible effect of imperfect die contact. In this case for the Vapochill evaporator head:

[Incoherent]

Re first graph. How are you calculating CPU power output? Interesting result, perhaps not totally unexpected if well lapped.

Second set I am not following totally. What is vapour chamber length? I'm not up to speed with phase change stuff.

[wijdeveld]

Hi, I'll follow up with some new results soon (they're running right now ).
In the second graph I've removed 12% of the connecting segments between the IHS and copper block; this results in a much more restricted heat flow from the die to the cooling medium (and therefore you need a higher dC/dx to get the same net flux).

CPU power output is simulated by the production of the compound HEAT. The heat concentration can be translated to a temperature. I've calibrated the production term one time against a know delta T versus load (it helps if Asetek provides the necessary data ) and from there I'm able to apply different loads myself. The model can only be considered adequate if the calculated load/temperature relation corresponds rather well without further tinkering in the settings.

Ah well, water of phase change cooling doesn't matter that much here since I assume perfect (instant) evaporator behavior and heat dissipation. You'll get a different inlet/outlet delta T and differences in the mixing behavior of the cooling medium. But for comparing the TIM effect this doesn't matter.

[wijdeveld]

O.K. it's done!

For easy of comparison I HAVE EXAGGERATED THE TIM LAYER IN THE CALCULATIONS!!! It’s now 0.25 mm instead of 0.025 mm

Contour plot:
- Just for illustrating the isolines from the CPU/die to the cooling medium in a 2-D slice (scale = 10 x 10 mm)
- Don't mind the concentration scale in the contour plot, I haven't calibrated the calculation on a temperature scale here.

x-y plot:
- All three TIM compound have the same results (within 0,1 oC) and plot as one line

So, what you can see in the x-y plot (which is temperature corrected) is that the TIM compound doesn't matter (I’ve tried goop, AS3 and pure copper conductivity), it's the heat flux across the solid/solid boundary that's limiting. I've also done a calculation without TIM layer (direct HS to copper block contact); this improves the delta T with ~10% (-1.1 oC for the CPU/die).
So, my hypothesis right now is that:
1) the exact compound composition doesn’t matter that much
2) ensuring a perfect contact between solid interfaces is the most important aspect of a good TIM
3) even in a perfect fit situation TIM adds ~10% to the deltaT (+1.1 oC on a total of +14 oC)

I’ll try to evaluate the thickness of the TIM layer on the slope later on.

[EDIT] I've applied a fairly low load: 60 watt

[Cathar]

Originally Posted by wijdeveld

So, what you can see in the x-y plot (which is temperature corrected) is that the TIM compound doesn't matter (I’ve tried goop, AS3 and pure copper conductivity), it's the heat flux across the solid/solid boundary that's limiting. I've also done a calculation without TIM layer (direct HS to copper block contact); this improves the delta T with ~10% (-1.1 oC for the CPU/die).
So, my hypothesis right now is that:
1) the exact compound composition doesn’t matter that much
2) ensuring a perfect contact between solid interfaces is the most important aspect of a good TIM
3) even in a perfect fit situation TIM adds ~10% to the deltaT (+1.1 oC on a total of +14 oC)

I’ll try to evaluate the thickness of the TIM layer on the slope later on.

[EDIT] I've applied a fairly low load: 60 watt

Yep, you've basically shown mathematically through modelling that which has been verified experimentally.

It is the very nature of the material disjunction that is the limiting element here. Most of us, probably too casually, refer to it as the "thermal goop layer effect", when really it is 3 different stages, being the disjunction as the heat leaves the die, as it travels though the TIM, and the disjunction as it enters the cooling device.

Most of all, I do agree with 3). However, I put it that during an initial thermal paste "settling" period that there can be a significantly larger delta from die to cooling device as the thermal paste settles and any air in the joint gets squeezed out, and perhaps more often than not, the characteristics of the paste that allows this to occur is perhaps the major difference that characterises the performance of various pastes. Also a paste's resistance to thermal pump out is another important factor.