Measuring CPU heat output.

[Les]

Great
Numbers and numbers, it gets better.
Am I understanding correcty., that this is the Thermistor correction curve http://<br /> http://w1.863.telia.c.../cal_curve.jpg ?
Am questioning my head not its validity .
Surely interpolation is suffice without the 4th order poly
Am digesting numbers slowly,will pay more attention after "The Hennesy"(invested in Strong Flow)

[pHaestus]

You should be able to edit the bios registers for ANY motherboard with WPCredit and force CPU HLT to kick in. Not sure what register and value for NForce2. I'll google it now.

//edit: I think this page might explain the registers:

http://gathering.tweakers.net/forum/...sages/743765/8

but I can't read it without the Swedish chef muppet voice in my head making me laugh.

An english alternative:

http://www.xtremesystems.org/forums/...threadid=11026

Not sure about the 4th order polynomial's necessity either:



How much do numbers change with a linear eqn or second order polynomial for calibration?

[Incoherent]

Sorry guys, I was a bit misleading, that chart was a first attempt, I had not insulated the leads very well or the thermal paste was contaminated with some AS. It was giving quite a large error for T1
I redid the sensor insertion and capped with epoxy instead of nail polish, it was more in agreement with the in air behaviour after that so I am happy that this is now OK.

This looks pretty linear but when you look closely, removing the slope it does not look as good:

T1 has a 4th order polynomial trend fit, T2 a 2nd order.
You could probably use a second order quite happily, I am using the 4th because of habit, I tend to use this with some more complex curves than this, it can handle more curve inflexions. Both would work in this case, but you can see why a linear fit would not.

The effect of correction on T1 only (for clarity)



And the adjusted response curve. T1 corrected and uncorrected.

Les, re interpolation. It most certainly would work. In fact it would probaly be more accurate but in the interests of speed when dealing with lot of data I find it convenient to use this kind of correction. A little more complex if you are not used to it (I do this kind of stuff every other day) but with intelligent Excelling it is much faster.

Hope this clarifies a few things.

Cheers

Incoherent

[Les]

Off topic: NH racing
"Strong Flow" won "The Hennessy" in grand style
Definitely "Gold Cup" material but think connections have opted for the "Sun Alliance" this season.
On Topic:
Even Wattage -uncalibrated should give some informative CPU's MHz and V-core data.
So far appears to scale with MHz but V^2 looks decidedly iffy..
More data,please,please.

[Incoherent]

I have made some measurements that are showing a connection between CPU temperature and power output. This makes some sense of course, electronics is always dependent somewhat on temperature. Higher temperature could easily lead to higher current.
It has significance for anyone who wants to attempt an electrical power measurement of a CPU. It makes the die simulator approach a bit more reliable perhaps.

Cheers

Inc

[Since87]

Originally posted by Incoherent
I have made some measurements that are showing a connection between CPU temperature and power output. This makes some sense of course, electronics is always dependent somewhat on temperature. Higher temperature could easily lead to higher current.

Yes, I would expect power consumption to be a complicated function of:

Voltage applied
Clock speed
CPU temperature
Software loading
The characteristics of the specimen CPU

I would be interested in seeing more details on this.

[BillA]

to which you could add:

environmental temperature
air speed at the mobo level
direction of air flow

an inordinately complex interaction

[pHaestus]

that's what makes life fun though!

[Incoherent]

Originally posted by unregistered

an inordinately complex interaction

It sure is.

One thing is clear, my set up is not stable enough to get absolute numbers from, but I believe that they are representitive. Once I have got a bit more data I will attack it and try and remove some of these variables mathamatically. It will still only be correct-ish, but we might than have a better idea of what is happening.

Some more data, as usual here, the relevant file is "power_calc.xls"

I have measured power at stock CPU speed but varied the voltages. It looks linear enough but as you observed Les, V^2 looks a bit iffy. Secondary paths and changing Rs rearing their ugly heads perhaps.


I am going back to try and get some more MBM Voltage readings to see how they work relative to nominal voltages.

Cheers

Incoherent

[pHaestus]

Ok I spent today running a flow test to see how well I could estimate W from deltaT across waterblock. Numbers are tabulated HERE

At low flow rates, I think W is reading low because the temp probe on outlet side is about 2' from the wb outlet. I am going to replumb so that it is only a few inches away. For really high flow rates then you have a different problem: 0.001C res is needed as 0.01C translates to roughly 9W at 2.25GPM. If you average 1.25-2.00GPM W then you get 81.53W. That's in reasonable agreement with your 2200MHz and 1.85V results (I am running 2200MHz and 1.85V Bios/1.81V in mbm).

[Since87]

I have found, with my Abit IC-7, that the sensors that MBM reads are less accurate than the value set in the bios.

For good accuracy, solder wires to a power pin and a ground pin of the CPU socket and use a voltmeter.

[BillA]

the other day I compared 4 cheap to moderate handheld dmms using a source and calibrated "good" dmm

would you believe off by 1.8V @ 12.000VDC on the 20V range ?

ck (calibrate) your meters guys

[Incoherent]

Originally posted by pHaestus

81.53W. That's in reasonable agreement with your 2200MHz and 1.85V results (I am running 2200MHz and 1.85V Bios/1.81V in mbm).

I think that is a very good match if you trust that the MBM voltage is the more correct reading.
This begins to call for ever more accurate measuring equipment, BillA, your DMM accuracy comparison is a horrifying result. My work Fluke is going to need a check up I think. You've scared me away from buying another cheaper type for personal use.
I have not checked voltages with any meter yet, but I have confirmed that MBM is simply reading the Bios hardware monitor readings. For now I will assume that they are correct, subject to verification later.

I am now going to concentrate on low power and frequency settings to try and quantify the secondary path losses. I am having a little trouble getting my head around semiconductor resistance vs temperature behaviour but I think I need to see if it's significant on paper, and hence definitely contributing.

Cheers

Incoherent

[pHaestus]

If I can actually win an ebay auction for a lab chiller then I may be able to help by producing some results with chilled water such that the CPU die temp and motherboard temperatures are the same. There are a lot of tests we could (and should) do...

[Incoherent]

pHaestus, I have been looking at your table of flow rates vs dT. That is very good clean data. How was it obtained, meaning how many measurements per data point were averaged etc?
Great work, it appears to me to be behaving very predictably, the dT is close to an inverse function of flow rate, actually it is almost exactly the inverse of flowrate minus an error which goes into your power calculation as you noticed.
If it is, as you suggest, caused by the output sensor being too far from the output, it (the error) should be also an inverse function of flowrate. And it is.
This begs the questions, what was the ambient Air Temp?, what is the hose/pipe material? and what is the exact length of the hose between the WB and output sensor?
I think we have a very clear situation of heat going into water at the WB and heat leaving water from the hose and you are measuring the sum of the whole system. Your data is so clean I think we can get very decent numbers from it.

Cheers


Incoherent

[pHaestus]

Temperatures taken with Digitec 5810 and YSI 700 series dual linear thermistors (0.01C res). I get better delta T numbers than you'd expect because I am lucky enough to have two probes that read EXACTLY the same over the 20-30C range (maybe a longer range but I haven't tried). Air temperature was monitored but not tabulated here because (a) the waterblock is going to have its springs swapped out this week, and (b) I could tell that there were some secondary losses at low flow rates so I wasn't as religious about writing down all the other temps.

I can take a picture tonight of the loop. I have it currently set up for maximum convenience in plumbing and so I could use one probe for measuring both rad out and wb in (plumbing is res -> pump -> T1 ->rad ->T2 ->wb -> T3 -> flowmeter -> valves ->res). I have another probe though and so it is trivial to replumb so wb in and out are measured very close to the wb. I may need to insulate the hose in that region too.

On how was data obtained:
System was allowed to equilibrate for exactly 60 minutes, at which time I would record the wb in and out temperatures every ten seconds over a one minute period. While I was writing these down in Excel my MAX6655 diode reader was recording CPU temperatures at 4Hz sampling rate. By 60 minutes the numbers track exactly and the delta T was never changing over the measurement period (the absolute value of wb in and out would go up or down by 0.01C in some cases though).

[Incoherent]

I have removed the heat current block first to test this and then totally removed watercooling for my impending move from Texas back to Sweden. Meanwhile I have become totally distracted by pH's data. Wrestling with it I think it might be possible to characterise the secondary losses.
Assuming measurements are correct three assumptions must be made;
1. That the real CPU power output is unchanged throughout the measurements,
2. That the real CPU power output equals the sum of measured power output, secondary (via mb and air) losses and hose losses. i.e. MB is cooler than CPU and we are not adding heat from other sources and therefore measured/calculated W must always be less than real CPU W.
3. The properties of the heat paths (primary and secondary) are not changing throughout the measurements i.e. conductivities, lengths, surface areas.

Predicted dTs can be calculated for different wattages and compared to the measured data:



From this chart and obeying 2. above I have interpolated that actual power output is 88W. This number is based on the fact that we should always measure less power than the CPU is producing no matter what the flowrate, if the motherboard T is less than the CPU. It is an estimate. That said, even if it is wrong the behaviour is valid I think.
The difference between ideal and measured is the dT error caused by the sum of all losses (MB and outlet probe position)

By using the measured power vs flowrate values and the measured CPU/inlet water dT we can generate a curve representing the lengths of an equivalent piece of copper (or any material) which would behave the same as the CPUdiode/silicon/TIM/WB/water path. Assuming that the properties of this path are constant (3 above), it is possible to extrapolate a value for the length and hence the apparent Q and thereby from the real "known" W the secondary path loss and Outlet sensor error (hose losses) which does not violate 2 above. (measured W+ 2nd loss+Outlet sensor error= "known" or real CPU power output.)



A bit flakey I know, but some of the results look right to me. It's hugely dependent on the assumed value of W. pHaestus, if you manage to replumb your setup it will show whether I am in the ballpark or off chasing wild geese.


Cheers

Inc

Edit: Oops, killed the charts. I will reinstate them tomorrow.
Edit 2, replaced the pics after an unfortunate FTP mouse click
Edit 3, corrected a mistake, "real CPU power" was "measured CPU power".

[BillA]

reasonable, but - as always - the question is "W"

[pHaestus]

setup WILL be replumbed and lab chiller will be used to control wb inlet temperatures as soon as it gets here.

[Les]

Like it.
Think we are on a similar wave length: You are treating similarly to this illustration?

from here

[Incoherent]

Very very similar Les.
Which is excellent. I am very pleased to see that you have been taking similar approaches, it makes me think I might not just be making this up. (it also confirms my chronic affinity for wheel re-invention)

Originally posted by pHaestus
setup WILL be replumbed and lab chiller will be used to control wb inlet temperatures as soon as it gets here.

Fantastic, that will be massively interesting data when looked at relative to what you have produced already. I am looking forward to it.

Originally posted by unregistered
reasonable, but - as always - the question is "W"

I've looked at some of the data I've been collecting, and produced this chart:



I am perplexed. Very noisy data, but the trend is clear and with over 500 data points the statistical probability that the trend line is correct is very high. This says that my diode error is 0.0336°C, which I can not believe, I was estimating about 5°C too low. I am a little confused, the only possible answers I can see are either that four separate thermometers (water, T1,T2,Ref) have a linearity/compression problem or the measurement noise is too excessive. probably the latter, standard deviation is 4.3°C. But there is some very interesting behaviour in the W normalised to water T. I need to look at this closely and have a bit of a think.
If this is correct, the chart also suggests a C/W of 0.509 for the water/WB/TIM/measuring block/TIM/silicon/diode heatpath. It might be possible, assuming the measurement noise is mostly secondary path losses, to put a figure on it (the loss) and hence get a fairly good figure for "W"
Great, if my diode reading is correct (which obviously it is) then my latest waterblock is blowing the Cascade out of the water performance-wise at 0.07C/W </OPTIMISM></IRONY>

Cheers

Incoherent

[Incoherent]

I have continued the analysis, I have come to the (unexpected) conclusion that my CPU diode is in fact telling the truth, relative to the water sensor. The (I have started calling it) "flux block" readings are 1.925°C out. This conclusion is reached by looking at the Calculated power vs T graph, normalised to water temperature and extrapolating down to zero CPU power output. Here are both unnormalised and normalised



The CPU diode temperature crosses 0W very close to average water temperature (15.6°C), the 0W temp for T1 and T2 is 1.925°C higher, in the normalised chart I have corrected them.
This check of the measurement system is open to debate but I believe it is valid.
Trusting this means a lot of other things fall into place. For example I have a number for the CPU diode/TIM joint, based on over 500 data points, statistically significant, of:
0.063C/W
and for the WB+TIM
0.162C/W
These are unexpectedly low perhaps and I am uncomfortable in stating them as fact, it is sure to draw fire. But, I am only working the numbers, and that is what the numbers say. (Les, if you can be arsed, I would welcome some cross checks)
Secondary path losses I am having trouble quantifying, they appear to be an odd function of water temperature and W, getting higher (actually might be lower ) at both low and high W and dipping in the middle. I cannot figure out a mechanism for that behaviour. I think I might have to remeasure with more controlled (although varying) water temps in order to track that mechanism through the noise. I don't think I have all the data points needed.

Be gentle

Incoherent

[Les]

Morning, it's bloody early(4.15am)
Been glancing thro whilst had coffee.
Not exactly sure which cross-checks.
However do confirm that the Cu shows K~390W/m*k (reassuring but not surprising since you fed in 392)
Assume you only want rough checks?
Confirm a "nearest CPU" (C/W)TIM =0.06203c/w (assuming 1mm offset)
Possibly a tad lower than the " 'til we know better" ~0.09 for a 1.1sq cm
But (1) Bill's value of 0.1(1sq cm) only allowed for an offset in the die t/c.There probably should,in my opinion, be an off set allowance for the wb's t/c; somewhere between 0.002 and 0.05(numbers are from memory but 0.002 is correct).This the point where simple analysis floundered.(2)Possibly accounted for by higher clamping pressure.(3) Forgetting(silly me) possibly the most important factor - the CPU Diode temp is effected by secondary mb cooling(added after more coffee)

Again for WB +TIM get similar 0.16643c/w ( 1mm t/c offset)
Have no comment on this value.

Earlier sums for Heat Flux seem correct(or at least the only example I checked) so, like you, I have to believe the numbers until it is found we are being stupid.
.
I do find the 1st graph inordinately pretty.
Not sure if these are the kind of checks you want, but I only do simple sums.Let me know

EDIT : Added : (3) Forgetting(silly me) possibly the most important factor - the CPU Diode temp is effected by secondary mb cooling(added after more coffee)

[bigben2k]

Here's a stupid question: what if the "flux block" was isolated with say, neoprene or such?

I keep looking for a curve, as the higher temps would allow for a higher secondary heat loss, but it's not visible (duh!).

Just babbling, don't mind me. Still, nice work.

[Incoherent]

Originally posted by bigben2k
Here's a stupid question: what if the "flux block" was isolated with say, neoprene or such?

I keep looking for a curve, as the higher temps would allow for a higher secondary heat loss, but it's not visible (duh!).

Re neoprene. It is currently insulated with some stick on foam rubber tape. I am intending to use neoprene for the next round of tests. I presume you don't mean as a TIM?

Re looking for a curve for secondary losses, I was too, but I am realising that in this data, (T plotted vs W) the secondary path losses/gains will not manifest themselves as a curve but as a different slope. I believe they are the source of most of the noise.
The problem is to isolate them.
Les,

the CPU Diode temp is effected by secondary mb cooling

The reason I think this is not relevant is also related to this. Although the CPU diode is affected, so are all the other sensors in the path to the water, thus the overall affect of second paths is simply to adjust the gradient. If I had kept the water temp stable for these measurements I believe that there would be less noise and no chance of quantifying the losses.
Bed-time for me now.
Re checks, your "simple sums" are perfect, more chance of spotting me "being stupid". Which I might add does not occur infrequently.

It is a gorgeous chart isn't it?

bigben2K, are you Swedish by any chance? I'm noting your use of "isolate", common amongst Swedes in this context I have noticed. Now I'm babbling.

Cheers

Incoherent