Measuring CPU heat output.

[Incoherent]

I have been reading various threads in this forum regarding the difficulties of testing waterblocks.
One of the unknowns appears to be the CPU power output. "Radiate" is not accurate (at all) and the rated specs for the various CPUs are also worst cases or design parameters.
I have decided to throw this idea at you guys.
I think it is possible to measure the power output of a CPU with a relatively simple method using Fouriers heat transfer equation:
Q=-k*A*dT/L

If you took a small block of material for which you know the thermal conductivity, eg copper @ 392 W/m°K or aluminium @ 233 W/m°K and produced a die 10mmx10mmx10 with two (or more) holes for thermistors 1mm from each end, drilled to the center of the block. (none of these dimensions are important, only that they are known.)
A picture would clarify this:


So, mount this on the CPU with your waterblock.
Switch on and let it stabilise. (loaded, idle, overclocked, underclocked, whatever)
Monitor the temperatures.
k=392W/m*°K
A=0.0001m^2
L=0.01m
dT=measured
Calculate the deg/mm. e.g. above numbers (35.2-24.3)/8mm =1.3625 deg/mm
Extrapolate to find the highT and lowT and hence the dT. (35.2+1.3625)-(24.3-1.3625)=13.625 deg
Calculate the heat being transfered. -392*0.0001*13.625/0.01 = 53.4 Watts

As I see it, the thermal gradient is proportional only to the power generated by the CPU all other things being equal. Knowing the other variables lets you calculate it reasonably accurately.
What this is doing is basically making a heat current meter in series with the path from CPU to water.

OK now flame about:

1. the uncertainty of the thermal conductivity of copper/aluminium. (problem)
2. the TIM joint at each end. (irrelevant?)
3. the unknown water temperature. (irrelevant?)
4. the unknown CPU temperature. (irrelevant?)
5. the secondary heat paths. (constant for a given CPU temp? small impact?)
6. the disruption of the CSA by the thermistor holes. (small impact?)
7. the accuracy of the thermistors. (problem)
8. pick your issue.


Just a thought.
Please forgive me if this has been covered in another discussion and point me in that direction.

Cheers

Incoherent

[Since87]

Originally posted by Incoherent

4. the unknown CPU temperature. (irrelevant?)
5. the secondary heat paths. (constant for a given CPU temp? small impact?)

5. There is reason to believe that the secondary heat path is not so small. (Les may be able to provide a reasonably justified range of values.) The issue is also complicated by heat producing components on the mobo such as MOSFET's and NB. It's certainly possible (though not easy) to characterize these things though.

4. If you did characterize the secondary heat path, you would need to know the CPU temperature (et. al.) to make use of that characterization.

Easier IMO to measure the electrical power input to the CPU. You'd still have the issue of how much heat is flowing through secondary paths though. I'd suggest 'nulling out' the secondary paths by using chilled coolant so that the CPU substrate was at the same temp as the mobo below the CPU.

[Incoherent]

Originally posted by Since87
5. There is reason to believe that the secondary heat path is not so small.

True I think. But the secondary heat path is the same for the cpu running "normally" and the CPU running with this "measuring block" in the heat transfer path. EXCEPT that the CPU temp is higher by 10-20°and this changes the gradient for those paths and hence the amount transferred through those paths.
I think it would only be in the region of a few percent though. Significant, but not an experiment killer.
There are also ways to compensate. Adjust water temperature to various levels, recalculate Q for each temperature and extrapolate back down to "normal" levels and the delta is the secondary path "loss" difference. (Some wild assumptions here I will admit, but it would be in the ballpark. If you could reduce water temp so that the CPU is at a "normal" level, it would be quite accurate.)

Originally posted by Since87
Easier IMO to measure the electrical power input to the CPU.

The voltage is no major problem (although not completely straightforward) but how do you measure the current? I don't think it's possible without major motherboard modifications.

The point ultimately being, I want to know how much heat is being transferred by the waterblock, knowing how much heat is being transferred through this path simplifies the problem.

Cheers

Inc

[Les]

.A couple of examples of my guesses at the variation of W with heatsink efficiency (from http://forums.procooling.com/vbb/sho...5&pagenumber=2 )
.

[Since87]

Thanks for the graphs Les.

Originally posted by Incoherent
True I think. But the secondary heat path is the same for the cpu running "normally" and the CPU running with this "measuring block" in the heat transfer path. EXCEPT that the CPU temp is higher by 10-20°and this changes the gradient for those paths and hence the amount transferred through those paths.
I think it would only be in the region of a few percent though. Significant, but not an experiment killer.
There are also ways to compensate. Adjust water temperature to various levels, recalculate Q for each temperature and extrapolate back down to "normal" levels and the delta is the secondary path "loss" difference. (Some wild assumptions here I will admit, but it would be in the ballpark. If you could reduce water temp so that the CPU is at a "normal" level, it would be quite accurate.)

I agree that by varying coolant temperature and doing various temperature measurements (including the ones you suggest) quite interesting data could be obtained.

Originally posted by Incoherent

The voltage is no major problem (although not completely straightforward) but how do you measure the current? I don't think it's possible without major motherboard modifications.

Some info on current measurement here. toward the bottom of the linked page and on. It would take some mobo mods, but not totally hideous ones.

At this point, I'd be happy to see data gathered using the THG method of putting a clamp on ammeter on the output(s) of the Vcore regulator.

[pHaestus]

At this point, I'd be happy to see data gathered using the THG method of putting a clamp on ammeter on the output(s) of the Vcore regulator.

Give me a list of parts and an idea of what to solder where and I'll do it. I am not all that gifted in electronics but I can solder and I have no fear of hardware deaths

[jaydee]

Originally posted by pHaestus
Give me a list of parts and an idea of what to solder where and I'll do it. I am not all that gifted in electronics but I can solder and I have no fear of hardware deaths

What he said. I am up for trying it aswell. I got a POS A7V8X-X that is just begging for an ass kicking.

[Since87]

All you really need is some wire and a clamp on ammeter that is capable of measuring DC.

You can get a DMM with a current clamp built in. Something like the Fluke 336 wouldn't be bad. It would give you a good general purpose DMM with AC and DC clamp built in. For DC current it's accuracy is spec'd at 2% +/- 3 counts. (I interpret 3 counts to mean 0.3 Amps, but it would be best to check with Fluke.)

A definitely cheaper option would be to get something like this Fluke 36 clampmeter on Ebay. I don't know the specs though.

You can also get clamp on probes that are meant to connect to a standard DMM such as the Fluke i410. (3.5% of reading +/- 0.5 Amps) The error of the DMM you plug it into gets added on to the accuracy of the clamp itself so the total accuracy would be substantially less, but you might find these much cheaper.

In case you hadn't noticed, I haven't mentioned anything other than Fluke products. I'm not aware of ANY other sources worth looking at for this kind of stuff.

If you (pH or jaydee) get a clamp on, I can characterize its error over a span from say 10A to 100A which should allow accuracy substantially better than the manufacturer's spec to be achieved. (Although, how long, and under what conditions such characterization holds true is open to question.)

The next step is to modify the motherboard. (The following is a lot of words, and pictures might be better, but look through this, and let me know if anything is unclear and I'll try to draw a picture to explain it.)

There will be several inductors in the Vcore regulator circuit which the current to the CPU flows through. (Two to four for mobo's made in the last couple years.) These are the little toroids (doughnuts) with wire wrapped through them near the CPU.

The basic idea is to connect the output side of these inductors to wires which go through the clamp and then back to the holes in the mobo that the inductor's output lead originally went to.

One end of the inductor will connect to a MOSFET (or two) and the other will connect to a capacitor (or two). The output side of the inductor is the side connected to the capacitor(s). Unfortunately you can't just use an ohmmeter to determine which end of the inductor is which. (The resistance of the inductor is so low that it's effectively a short to any handheld ohmmeter.) It's likely that you can visually look at the board and determine which side of the inductor connects to the cap(s) and which to the MOSFET(s). If not, you will have to unsolder at least one lead of the inductor, and use an Ohmmeter to determine which inductor lead is which.

Once the output lead of the inductors are identified, they need to be desoldered from the circuit board. Then you need to attach a fairly heavy wire to the lifted lead of the inductor. Ideally you would use the thickest solid copper wire that will fit into the hole in the motherboard that the inductor lead was pulled out of. (I'd say 16 to 18 gauge at a guess) Start with about a 1 foot wire for each inductor. Wrap one end of this wire around the lifted inductor lead and solder the connection. When you've got the wires soldered to the inductors, bring them all together so that you can get the clamp around all of them simultaneously. Then cut them down to where they are just long enough so that the loose ends of the wires can be soldered into the holes the inductors were pulled out of. (It's note likely to matter whether the wires go back into the hole for the inductor they are soldered to, but just in case, make sure to match the wire from each inductor with the correct hole in the mobo.)

Ideally, where the wires are to pass through the clamp, they would be close together, and pass through the exact center of the clamp. You want to avoid having any of these wires running near the 'gaps' where the jaws of the clamp come together. There is the obvious 'gap' where the clamp opens, and the less obvious gap at or near the hinge point of the clamp. One of the major sources of uncertainty in clamp-on measurements, is the location of the wire with respect to the 'gaps'.

You've got mutually exclusive goals here. Keep the added wire as short as possible, yet make it long enough that you have a nice big loop that you can easily clamp your meter to while keeping all the wires away from the 'gaps' in the clamp. Use your best judgement. The manufacturer's spec includes uncertainty for where the wire is located wrt to the gaps, so you should get better accuracy than spec. (But who know how much better?)

Once you've figured out how you want to route the wires, solder the loose ends into the holes left in the motherboard and you are ready to go.

Everything should run fine regardless of whether the clamp is used or not. The Vcore regulator will be somewhat less efficient due to the added wire, and it will radiate more EMI, but neither of these issues is likely to have a substantial impact. (In pH's case the need for twisted leads from the CPU diode to the diode reader will be a bit greater, but hopefully the leads were already twisted anyway.)

It would probably be easiest to work with magnet wire, (The 'enameled' wire that is used to wind motors and transformers.) but I'm not sure where you could get a small quantity. If you find out the ID of the holes in the mobo that the inductors solder into, I can send you the gauge of wire that fits.

[bigben2k]

Did everyone forget Groth's excellent work?
http://forums.procooling.com/vbb/sho...&threadid=7124

[Since87]

Originally posted by bigben2k
Did everyone forget Groth's excellent work?
http://forums.procooling.com/vbb/sho...&threadid=7124

Not at all.

[pHaestus]

Ok that was an extremely detailed reply Since87 and I completely understand what must be done. Guess I will start the search for a fluke clampmeter, send it to you for calibration, and then bust out the iron.

Ben I certainly haven't forgotten about Groth's ammeter, but as far as I know it never got to the point of properly working?

[jaydee]

Originally posted by pHaestus
Ok that was an extremely detailed reply Since87 and I completely understand what must be done. Guess I will start the search for a fluke clampmeter, send it to you for calibration, and then bust out the iron.

Glad you understood it. I tell you what, let me know when you find the clampmeter and I will pitch in on the cost. Better than me buying one and not knowing WTF to do with it.

[Since87]

Originally posted by jaydee116
I tell you what, let me know when you find the clampmeter and I will pitch in on the cost.

Same here. I can Paypal you some towards the cost.

If you find something used in the US I could also just purchase it myself, calibrate it and ship it to you. Save a bit on shipping. (If you buy something new, you'd want the warranty, so this wouldn't be such a great idea.)

[pHaestus]

I would prefer used. Think that Fluke36 is tolerable? Not a bad price if it would do the job...

[Incoherent]

Originally posted by bigben2k
Did everyone forget Groth's excellent work?
http://forums.procooling.com/vbb/sho...&threadid=7124

Now that is interesting. Excellent work being done all over the place.
I am kind of sceptical of doing any kind of current measuring at these low voltage levels, small errors tend to take on obscene proportions in the final result, Groths work shows the precision required I think.
Since87, I think that your suggested method is really the only option not requiring specialised techniques but from experience with these kind of current levels, where the solder joints are critical, I would be worried about the R losses and hence the reliability of the measurement. (not to mention the solder joints themselves, I^2R, they can melt )
Basically, I don't think it is practical but I would be more than happy to be shown otherwise. (pHaestus, Jaydee116, if you do this, make sure your solder joints are flawless.)
Hence my suggestion for a thermal approach.

Originally posted by Les
.A couple of examples of my guesses at the variation of W with heatsink efficiency (from http://forums.procooling.com/vbb/sho...5&pagenumber=2 )
[img]

Les, thank you for these.
Assuming the numbers I used above the "measuring block would add about 0.25C/W to the "waterblock C/W", on your left chart example (guess?) about 60%, on the right (from Billa's Die Simulator?) about 95-97% through the WB path.
I am assuming the left chart is an example illustrating a point or are those numbers from some actual data?
See, I think you could measure this loss pretty exactly, as I suggested above.
There are a few holes in the concept though.

Cheers

Incoherent

[Since87]

Fluke doesn't list it anymore, but it looks like several used dealers snagged the Fluke web page for it. Like this.

Relevant specs are:

Range: 1000A, 200A
Resolution: 1A, .1A
Accuracy: 0-600A: ±(1.9%+4), 600-1000A:±(3%+3), 0-200A: ±(1.9%+7)

My interpretation is that on the 200A range, the resolution is 0.1A and the accuracy is 1.9% + 0.7A.

Assuming 70 Watt dissipation on a CPU at 2V, gives 35A. So likely less than 4% error. (With the CPU loaded.) Probably substantially less error than that when the error data I can generate is applied.

The effect of that '+7' could be reduced by looping the wire through the clamp several times, and dividing the reading by the number of times the wire was looped through the clamp. I don't think I'd actually advise doing this though.

There is still the issue of how the reading is affected by high frequency components of the current going through the clamp. (My main objection to the THG article) A set of readings of idle and load current at different Vcore settings ought to provide an indication how much error this can cause. If it appears that error due to high frequency components of the current is substantial, additional steps can be taken to improve the test setup.

I think the Fluke 36 would definitely provide interesting data. It's certainly not going to get the results that can be achieved with a die simulator with 'inline' current measurements, but I don't think that's necessarily the goal here.

[Les]

Originally posted by Incoherent

................
(1)I am assuming the left chart is an example illustrating a point or are those numbers from some actual data?
(2)See, I think you could measure this loss pretty exactly, as I suggested above.
There are a few holes in the concept though.

Cheers

Incoherent

1)Yes the (C/W)mb=1 used on the left is illustrative .
However it is chosen as my guess at reality using "indirect proof." arguments based on(1) the limited comparative "Insulation-calibrated Die simulation" tests and "cpu in motherboard",(2)Ambient and sub-ambient wb-base/cold-plate temperatures.*
2) Sounds a beast of a set-up and procedure. However do go ahead with some exploratory work;all experimntal data have some meaning and I await yours with interest.

* The results used to form the argument have been gleamed over a period of ~ 2 years and have not documented it.The best I can offer is : (1) is primarily based on Billa's and Hoot's work(eg http://forum.oc-forums.com/vb/showth...threadid=75649 ) (2) on Peltier work (e.g http://forums.overclockers.com.au/sh...5&pagenumber=1 ) and chiller work by PlayerO .

Edit: added 2nd link. Edit 2 : removed accidental smilie

[jaydee]

Interesting stuff. Maybe move it to the appropriate section of the site?

[Since87]

Originally posted by Incoherent
I am kind of sceptical of doing any kind of current measuring at these low voltage levels, small errors tend to take on obscene proportions in the final result, Groths work shows the precision required I think.

I do R&D on equipment that can typically measure a 50 Amp AC signal with 25 ppm accuracy with less than a 15 mV drop across the measurement input. There are no fundamental problems with doing these kind of measurements, just implementation details.

Originally posted by Incoherent
Since87, I think that your suggested method is really the only option not requiring specialised techniques but from experience with these kind of current levels, where the solder joints are critical, I would be worried about the R losses and hence the reliability of the measurement. (not to mention the solder joints themselves, I^2R, they can melt )
Basically, I don't think it is practical but I would be more than happy to be shown otherwise. (pHaestus, Jaydee116, if you do this, make sure your solder joints are flawless.)
Hence my suggestion for a thermal approach.

I'm not sure where you see R losses having an impact on accuracy.

In the case of the clamp, it is the magnetic field generated by the current that are being measured and series R doesn't matter. (Unless the clampmeter itself is being heated substantially so that its calibration drifts.)

In the case of the inline measurement Groth was working on. The use of low tempco nichrome shunts was discussed. Easily available shunts could provide much more temperature stability than is needed.

You bring up a good point about the power dissipation in the wire. Assuming a total CPU current of 50A and a two phase Vcore regulator, we are talking 4 Watts of dissipation per foot in 18 gauge wire. Not likely enough to melt the solder joints, but you certainly wouldn't want to touch the wire. With 16 gauge wire the dissipation would 2.5 Watts per foot.

I thought about this when writing my earlier post, but it seemed like enough detail at the time. Ideal would be to rewind the inductor with heavy gauge wire like 12 Gauge, and leave one of the leads long, to go through the clamp-on. Short smaller gauge wire (16 gauge or so) would need to be soldered to the ends of the 12 gauge to connect the inductor back to the motherboard. The 12 gauge wire would act as enough of a heatsink that there would be no meaningful risk of the connection becoming unsoldered. (Assuming the solder joints were done well in the first place.) It might take a 45 Watt soldering iron to make the connections.

[Incoherent]

I have been playing a bit with some numbers and thinking about these secondary path losses. I come up with something slightly different to what you did Les.
The secondary path "losses" are dependent on the temperature delta between the CPU and the motherboard/surrounding air.
In an environment like a typical case, I believe that the normal situation ie, HSF, the CPU has a temperature somewhat above the motherboard temperature.
With water cooling however it is not at all impossible for the situation to be reversed, i.e, the CPU is now cooler than the surrounding motherboard, with it's other heat sources (MOSFETs, regulators etc) and poor heat removal systems.
In this (my) case it is not physically possible for the secondary heat paths to be reducing the efficiency of the waterblock heat transfer. Indeed the reverse is now true. The Waterblock MUST be transferring (via the CPU) more heat than the CPU is producing
I think the function is an inverse exponential proportional to CPU power output, assuming a constant dT, which of course will never be the case but the principle stands. This chart probably exaggerates the effect. Assumed values for this chart:
R (2ndary path thermal resistance)=1C/W
TIM R= 0.08C/W
WB C/W ~-0.26C/W


To quantify this we need to know the CPU temp of course. And that's why I want the CPU power output. Electrical methods aside, I think it can be done, with reasonable accuracy and inexpensive components. Remember, the bigger the difference between the measured values (ie the higher the dT), the less the accuracy and precision of the equipment matters to the accuracy of the end result. Hence the proposed approach.

But I could be wrong.

Cheers

Inc

[Incoherent]

Originally posted by Since87

I'm not sure where you see R losses having an impact on accuracy.

Only this:
If you change the resistance of the wires feeding the CPU, you change the voltage at the CPU input, thus changing the conditions which you were trying to measure in the first place.

However, that is probably moot. You would measure the voltage at the CPU pins anyway so the power output would be presented as e.g:
@ X input voltage and Y freq, the current was z so P=IV. (Are we sure a CPU is a purely resistive load BTW? I guess so. )

So you are right. Ignore that statement.

Cheers


Inc

[WAJ_UK]

This is all well above my head I'm not a sparky but I have something that may help. Its a PDF titled:
"Modeling and Analyzing CPU Power and Performance: Metrics, Methods, and Abstractions"

written by

Pradip Bose
Margaret Martonosi
David Brooks

I didn't understand it myself but I can send it to any of you if you wish or you could search for it on google as that is where I got it from

[pHaestus]

Also as an aside, BillA did something kinda similar to this with a thin piece of copper. His goal was to characterize the TIM joint resistance. I think he stuck to thin thicknesses of copper and a single temperature probe though. There are some differences here (the insulation around the outside mostly) but same general experiment.

A practical concern may be that drilling holes into this little piece of copper to measure its temperature may introduce some "heat shadow" effects that will affect the results. No idea of magnitude of this effect, but since it's a pretty small piece of copper it could be an issue.

I am wondering about the relative advantages of this technique vs. another indirect measurement like insulating around the die and the waterblock and calculating W from the change in water temperature across the block. I can think of some pros and cons to both.

[BillA]

the smaller the device(s), the greater the effect of extraneous influences
that work with the shims was inordinately difficult, therefore was done only once
with the resulting lower confidence in the results
keep it simple

I do it now with the coolant temp, but at high flow rates (3gpm) the error is higher due to the limited resolution (0.01°C) and the proportionally greater effect of system measurement uncertainity (±0.03°C) on a value of 0.12°C

EDIT: had it bassakwards

[Les]

Originally posted by pHaestus
Also as an aside, BillA did something kinda similar to this with a thin piece of copper. ....

http://forums.overclockers.com.au/sh...threadid=67401

Analysis was never satisfactorily concluded(IMO).