Measuring CPU heat output.

[Les]

Originally posted by Les
possibly the most important factor - the CPU Diode temp is effected by secondary mb cooling

Yes you are right.It is crap.
As you say "the overall affect of second paths is simply to adjust the gradient".This being accompanied by an adjustment of heat-flux.
However think I score a point by being first to be stupid.

[bigben2k]

Not Swedish; French Canadian!

Another Q: did you use the same "flux block"?

One reason I ask, is because sometimes, a testing method has to go to "extremes", to prove or disprove a point. Not saying that your figures are wrong, and I recognize that gathering 500 data points involves a tremendous amount of work, but I'd like to see it done over a few more "flux blocks".

I also have to analyze the method here, a little bit: one possible source of error, may come from miniature temp probe movements, no? Would you be able to find any trend (up or down) in the sequence of data? (assuming that that would be relevant/applicable)

Just trying to tighten up the results...

edit: here's another one: How about calibrating the flux block, by setting each end at different temps, on purpose?

[Incoherent]

Originally posted by Les

However think I score a point by being first to be stupid.

LOL
A bold claim Les. There are entire societies which have sprung up in this data which have no other purpose than to build monuments honouring the enormity and awesome power of my stupidity. But I will concede, subject to recount. These kind of competitions have a way of switching leaders very fast and I am on your tail.

bigben2K, the same flux block was used for all these measurements but other parameters were changing. Tests with another would be very interesting.
Re calibrating the flux block as you suggested. Tricky. If you have a heat source at each end the temps just explode. With a calibrated heat source and a calibrated heat "extractor" (peltier?) it could be done but then you are moving outside the purpose. The "calibration" it has is a "known!?" conductivity, "known" dimensions and many measurements. It is inherently resistant to probe movement because the range of possible movement is small as a percentage of the distance between the probes. This also leads to it being insensitive to probe errors as well. Analogy: Measure the thickness of a coin to 1% you need a micrometer. Measure the length of a room to 1% you only need a tape measure. The flux block is the tape measure in this case, and the kind of work pHaestus is doing is the micrometer, far more accurate by necessity to obtain the same accuracy. Ultimately his data will be far more valuable but consider this a poor mans method for some of the same problems.

Cheers

Incoherent

[bigben2k]

Ok. What would you list as the "other parameters"? Just curious.

I think that your work has potential to be more accurate, I'd just like to see how it could be done. Would a diode reader, similar to pHaestus be of any help?

[Incoherent]

Originally posted by bigben2k
Ok. What would you list as the "other parameters"? Just curious.

I think that your work has potential to be more accurate, I'd just like to see how it could be done. Would a diode reader, similar to pHaestus be of any help?

Other (varying) parameters are:
Water temperature
CPU voltage
CPU Frequency
Ambient air temperature
Motherboard temperature
CPU Loading.
TIM joints settling.
Time between taking CPU diode reading and taking FB T readings.

Constant parameters are:
Flux block dimensions
Flux block conductivity
Waterblock/Pump/Radiator (and presumably flow rate)
Thermistor calibration (hopefully)

An accurate calibrated diode reader would certainly be helpful. The extrapolation of the MBM reading down to 0W works I think and I trust it for now but it would be nice to be a bit more sure.
Same goes for all the sensors, I think I will move to a more highly calibrated setup in rev 2.

Cheers

Incoherent

[pHaestus]

If it would help I have a Duron (Morgan) 1100 upon which I'd be happy to solder twisted pair and a diode reader. The problem is that unless you can find some way to hack the SMBus that it'll only report with 1C resolution in MBM. There was someone on overclockers forums that had written a program for NetBSD that gave the 6657 the ful 0.125C res; might want to search there. The added resolution really IS needed.

I use the 6655 evaluation system from maxim that includes a SMBus to parallel adapter and software. It was around $100US shipped from them though and so may be more than you'd care to invest in the package.

Another option would be to send me a "flux block" and let me run a few tests on my setup. The advantage would be that I could get a lot better temperature measurements. The disadvantages would be that I'd be out of waterblock testing for a week or two and that I am sorta leery about mounting this thing on my test CPU.

[Incoherent]

Secondary Path Losses.





I'll explain later.

Cheers

Incoherent

[Les]

Like the colours.
May I start thinking after the Tripleprint Gold Cup at Cheltenham - small investment in Telemoss

[bigben2k]

Higher dT's at lower differences between system temp and room temp... right?

But at 90% efficiency, doesn't that mean a 10% secondary? I thought that actual secondary loss was 1 to 2 %?

[BillA]

"I thought that actual secondary loss was 1 to 2 %?"

ONLY with an inordinately well insulated heat die
socket/board/trace losses will be vastly greater
10x is vast, eh ?

[bigben2k]

Yeah... makes me thing about insulating the edge of the heat die, but then I'm not reproducing an actual PC anymore. Don't mind me, just babbling some more (shooting for the cup, still )

[Les]

Where is the 100% derived? An extrapolation?
You are talking "mb secondary heat paths" not "flux block"?
Please excuse the questions- I still have on my "bewilderment head "
However numbers look eminently sensible.

Telemoss was a creditable 3rd(2m 6f) - difficult to access - tis 2nd time has looked good at top of Chelt hill and not won(previously over hurdes - Championship 3m ) - Dunno - but was handicapped to win today- maybe still prone to BB.

[BillA]

jeez Les, betting on hurdlers ?

[Les]

Chasers, but high class(within 25lb of Arkle).
Odds are odds

[BillA]

an insight
now I understand how/why you are so adept with numbers
I understand odds are are odds, but is not the scatter much higher than with flat racers ?

[Les]

Originally posted by unregistered

but is not the scatter much higher than with flat racers ?

Would guess so ,particularly dung July and Aug in the uk , but gives better opportunities or disaster pits
I only play with the horses like I do with the quasi-theoretical cooling
An fully drunkenly retired

[Incoherent]

I am very close to gaining a stupid point.
The chart above showing secondary losses is reliant on the measured W tracking water temperature. I established the 100% level by plotting all W measurements vs water temp and finding the linear which fit. This curve I used to recalculate W adjusted for water temperature.
The problem is that I do not have enough data. I need many more measurements at fairly constant power levels but varying water temperatures. Essentially I think it is mostly luck that the trend I used agrees fairly well with what I believe will happen. Which is bloody dangerous. It is too easy to make the experiment results match expectations.
Consider the above chart a picture of the tendency, but be suspicious of the absolute Efficiency number. That said, I am in the process of looking at this from the bottom up by isolating temperatures, it is in reasonable agreement, losses are possibly a bit less. (the mental contortions this approach requires are painful). Ballpark.
Another chart of the same data, this time relative to absolute CPU temp.


Oh, Les. Yes, I'm trying to extract the MB losses. I don't think the flux block loses too much. Other issues instead as touched on earlier (heat flux uniformity etc)

Cheers

Incoherent

[Les]

Finally got round to studying the Excels.
Must commend .There are thousands of numbers! Your organisation and manipulation puts my efforts to shame.

Calibration I will leave to "measurement accuracy questioners".
Have no problems with your Wattage calculations other than puzzling over L=0.008m now appearing (rather than the original 0.01m).

As I understand you have set the 100% Efficiency as that at a Water Temp of 0c( W=Wo - 0.2635Tw)?
This seems somewhat arbitrary and needs justification - I am not happy.
However using the technique show trends is probably OK.

Maybe worth a recap to you 1st graph ,here
Although this describes the behaviour I would expect from a "Case-enclosed-MB", I do not think it is adequately described by constant "die-MB DT" and constant R.They indicate a change of heatsink temperature(case or room).

Got completely lost by your "poly" manipulations - however will probably dawn.

Just ramblings not a point scoring mission.

[pHaestus]

My chiller ended up in another country (resolution is imminent though) and so I am left playing with W estimates from water temps. I replumbed the loop so that temperature probes are closer to wb inlet and outlet, insulated the probes, and insulated the hoses. The result was a W estimate that is substantially higher than before, but that STILL isn't constant over a very large range of flow rates:



I am using a Cascade now instead of MCW5000A, and CPU temperatures were quite a bit lower than in the first run. I shall try to get another run in tonight; consider this very preliminary.

[Les]

pHaestus
Maybe,just maybe, consistent with

from here

[BillA]

better presentation Les
but I would call it 'Heat Source Efficacy'

[pHaestus]

I discarded the 2GPM point as a difference of 0.01C equates to nearly 10W. The rest of the data does indeed look familiar:



I've generated straighter data though...

Back in same old boat; C/W vs. W means more variables than one can sort.

[Incoherent]

I am in a frenzy of packing so I must be brief. I will be off line for a wee while now as well, due to a trans-Atlantic move.

Briefly. Les: Your ramblings pack a punch, I realise I need to start over, I'm very impressed at you being able to make sense of my meandering excellery. I find it almost impossible to follow my own stuff after a few days, let alone anyone elses.

puzzling over L=0.008m now appearing (rather than the original 0.01m).

This is the effective L for FB-Water, ie WB+TIM. I was checking the apparent power for all the sensors as sanity check and to try and get a better power average. You will notice that this value is only used for the T2-Water power calculation.

As I understand you have set the 100% Efficiency as that at a Water Temp of 0c( W=Wo - 0.2635Tw)?

Correct but at average water temp, not 0C. This is the elusive and as you say somewhat arbitary relationship between W and water temp. As I said above, the curve it is derived from is W vs water temp. This is very dependent on what measurements (at what water temp, voltage, frequency etc) have been taken so I concur, I am not happy either. In fact checking myself I would go as far as to say trash this, it can't be done this way. (Although I think the curves show pretty close to what the behaviour would be, but the offset is the unknown and unfortunately exactly what we're after)
The Efficiency vs W chart. There's another Excel sheet I've put up ("conductivity tester.xls" which a possible situation closer to what is likely to be reality. I agree, a constant CPU-MB dT is unlikely.
I'm missing data. Need MB temp but too many variables in that.
Some approaches show promise, but I think the uncertainties may be too much. Shame, I had some hopes for this approach.
The poly manipulations, a good trick, the Excel command INDEX(LINEST(y,x...x^2...x^3),1...and..2...and...3 ..) can give the constants for any order polynomial. Help is worthless. Check it on the sheet to figure it out. It gives the same numbers as a chart trendline equation but to far more significant figures.

pHaestus. Great data. I wish I had time to look at it immediately. Curious that W is so high.

Summary of the Flux block so far is that it gives a value for W to water, a value for TIM joint C/W, and a value for WB+2nd TIM joint C/W. All dependent on how much one trusts ones sensors.
Secondary paths need more work. Perhaps readings down to very cold water (Swedish winter here I come) would show a "knee" in the power readings which might be the point at which there are no secondary losses i.e. CPU temp<MB. the relationship of CPU temp to "electrical" power needs to be known, distinguishing from secondary loss is difficult.

Cheers guys

Incoherent

[Les]

Originally Posted by pHaestus

........
I've generated straighter data though...

I do suspect something is amiss and results are only partially explained by change in "Heat Source Efficacy " with Waterblock's "C/W"
Analysis gives:

[pHaestus]

I think I might have an explanation for the high W with the Cascade. When I removed it the other day I accidentally touched the northbridge hs (passive) and burned my finger (ok not to the point of a blister but it was red for a bit). The cascade and other 4 hole blocks are really close to that nb heatsink while the MCW5000-A is not (it's square rather than rectangular). So my guess is that extra heat from nb (as much as 10W though?) is being dumped into the cooling loop, and that the better the wb the more heat goes in (from copper traces on mobo and possibly from air as well. I'll put a fan on that hs and see if the numbers are a little more sensible. My cascade sprung a leak though and currently my test system is being air dried.