How's this?

Watercool a system with chilled water so that the CPU diode indicates the same temperature as the motherboard in the CPU area. Now you know that no heat is moving through the pins. (There will be a temperature gradient across the motherboard under the CPU due to voltage regulators and other heat sources. The gradient will change too depending on how much current is flowing through the regulators etc., but you get the general idea. Epoxy copper sheet to the back of the motherboard to spread the heat around the CPU? Heatsink the MOSFETS of the voltage regulators?)

Insulate as for a TEC cooled CPU, so that air convection can't carry any heat away. 'All' the heat is going through the waterblock.

Since you know the C/W of the waterblock under the operating conditions, (+/- the error due to the CPU to Block TIM joint) you know the Heat output of the CPU.

As usual TIM joint error is the major problem. Regardless of how consistent you are when mounting to your die simulator, the curvature of the CPU die means that you won't get the same C/W through the TIM. I don't have a solution to that problem.

Ugly test with high probability of trashing a motherboard for questionable results.

Much more cleverness required.

BTW, WRT to the temperature diode stuff I mentioned earlier:

It occurred to me that the voltage output of the setup will be around 0.65 Volts with a +/- 0.01V swing over the entire range of temperatures you're likely to care about. It would be highly desireable to have a stable 0.65V source to 'null' against the diode voltage. This would allow the voltmeter to operate on the 0.1V range. (or even the 0.01V range if the voltage source was tweaked to match the CPU.) The diode voltage changes about 200uV/C, so this would help a lot with resolution. However considering the noise environment of a diode on a CPU, I'm not sure how much resolution can actually be useful.

Anyway, if you decide you want to do something with a CPU diode, let me know. I'll set something up for you.

Excellent work but I suspect it is not nailed down.
If I am understanding correctly this is a calibration of a CPU Diode connected through the CPU's pins to a Maxim 6655/ Reader . This shows the combination is capable of giving accurate temperatures.
Is it legit to assume the same calibration will apply when there is an intervening Ziff socket?
Or is a further motherboard in oven calibration required?

I soldered to the base of the pins so that I can just cut away part of the top piece of the ZIF socket and run the wires out through the hole. As so:

http://www.voidyourwarranty.net/revi...etmodded_s.jpg

Yes my concern with dumping CPU into water was the fact I was applying and measuring 100 uA and then a 10uA current through the diode pins rapidly. I could see that the measurements where water had gotten into the baggie were "off" so I think it is a necessary step. If I cared much about the long term viability of the CPU I wouldn't be soldering wires onto its pins!

Re: heat flux

Nevin from arcticsilver brought up an interesting point. If the AMD diode isn't near the hottest bits of the CPU then there will be a significant difference in the peak core temp and the measured core temp. He was concerned about this re thermal paste testing I guess, but it would also mean I will tend to overestimate the importance of secondary heat losses because my measured diode temp is less than the real core temp. More stuff to sort out always. Nonetheless this is a pretty good start.

Substantially lower currents than I'd recalled.

pHaestus, would you give me a link to the in depth info on how these diode readers work?

http://www.procooling.com/articles/a...de-amdspec.gif

is specs for AMD internal diode, and you SHOULD be able to dig the pdf for the 6655 from Maxim (it also as I recall uses 10 and 100 uA as the forward sourcing currents). I am on dial up AOL in a remote location (dialing to server 45mins away) and it is taking me approximately a minute to even load a procooling page. The maxim pdf is beyond my attention span...

MAX6655 (PDF, 18 pages)

The section from the AMD spec must have been what I'd skimmed before. (I'd been looking for that info on the Maxim site and having no luck.) Thanks.

I'd already linked the MAX6655 spec in my first post in this thread. (I'd missed seeing the diode currents in there, because they were listed on a page, where all other specs related to the digital half of the chip.)

Modified the motherboard today. CPU still works properly! Tried running some preliminary tests with a heatsink for giggles (and to write the progress thus far up as an article and came across something interesting:

http://www.procooling.com/~phaestus/prime95.jpg

Slick what resolution plus sampling rate can do, eh?

The data looks noisier than it actually is because the maxim software allows for 1/4Hz sampling but only adds timestamps with seconds. To further complicate matters, the software doesnt log successive points which do not differ to save space. The graph would improve in oscillatory pattern if I just set the reader to 1Hz sampling.

That is slick to see though isn't it? Now I wonder if I can see temp differences between "slowest ram timings and low fsb" and "most aggressive ram timings and high fsb" for same MHz?

CPUBurn is without these oscillations; much tighter loading.

That's pretty cool, how you can actually see the CPU temperature varying as the load varies.

now we want more
quite interesting

Yeah, really cool how much the temperature fluctuates with load and without delay.

Is the time in seconds?

Are those readings after the calibration? :drool:

Maybe a stupid suggestion but,

Why didn't you use some form of oven to calibrate and all?

As I was saying the time x axis is in seconds but the data is in 1/4 seconds so the .log has up to 4 points with the same x value. I will rerun at 1 Hz (1 per second) sampling rate tonight.

An oven would work better BUT I don't have one that can reliably be set in the 25-70C range stably

Cool data.

Is the PC that is communicating with the temp sensor different than the one the rigged CPU is in?

Can you show a graph of the temperature when you have the rigged CPU fully loaded, and suddenly power off the machine? (Using CPUBurn if it gives a 'cleaner' signal.)

This might give some indication of how significantly RFI is affecting the reading. If there is a sudden step in measured temperature followed by an (exponential?) decay, the step may be attributable to noise causing an offset in the reading. Depending on the steepness of the exponential decay, it might be difficult to tell. (A case where the specific heat of copper does matter, and would, in this case, be advantageous.)

Try toast, it raises the temperature more than cpuburn or prime95. Also I think it is more stable.

Since87:

Yes there are two PCs: the desktop PC with the rigged CPU, and my notebook that is connected to the maxim reader via a parallel port. At the moment, I have the maxim's power going though the desktop box, so my temp monitoring loses power upon shutdown. I can easily enough add another PSU into the mix to supply 12V and 5V. For reference, here are the CPUBurn test results from last night:

http://www.procooling.com/~phaestus/cpuburn.jpg

See how tight and how little variance with this code? So my guess is that the erratic temp swings when Prime95 is running are NOT just noise and rather representative of very rapid fluxes in temps.

Wow!

I'm very impressed.

Any ideas on how one can actually measure the current draw of a CPU? THG reported that they were doing this once, but gave zero details of the method for measuring current draw. I think I can figure out the vcore measurement (vcc pins eh?), but unsure about getting the current...

The only way I can think of to measure the current, and not totally compromise the CPU's ability to run at high speed, is as follows:

Take a sheet of copper that is thin enough that it can fit between the CPU and socket without interfering with the socket connections, yet thick enough it can be drilled without folding up completely. (Maybe 0.030")

For each Vcc pin location, drill a hole with an ID just slightly larger than the CPU pin.

For all other pin locations drill a larger hole so that there is no risk of shorting the pin to the foil, but there is still a substantial mesh of copper.

Place the drilled sheet of copper over the CPU pins.

Solder all of the Vcc pins to the copper sheet, and cut them off flush with the sheet.

Remove the electrolytic capacitors on the motherboard which are connected to the Vcc regulators. Reinstall them on the backside of the board.

Install the cpu with attached copper sheet into the socket.

At this point you need to prepare a very short heavy wire to connect between the Vcc pin of the electrolytic capacitor, and the copper sheet. This heavy wire has to have two smaller wires attached to it near either end. This gives you a 4-wire (Kelvin) connection to a length of wire having some resistance. (1 inch of 10 gauge wire would give you 83 uOhms, so with 50 Amps going through it you would measure about 4 mV.)

Before soldering the heavy wire into place. Measure the voltage drop between the two smaller wires with 10 Amps flowing through the heavy wire. This allows you to calculate the resistance of the heavy wire between the two smaller wires.

Now solder the heavy wire into place. (Don't allow the connections of the smaller wires to change while you do this.)

Connect additional 10 gauge wires between the Vcc pins of the other electrolytic capacitors, and the one to which your 'resistor' is connected. (These additional wires are to keep the full current of the CPU from being pulled from the pin of one electrolytic capacitor.)

Do this, and there's some chance the setup will actually run and allow you to measure the current. I think there is likely to be a problem of inadequate decoupling capacitance between Vcc and ground that prevents stable operation. Space constraints make it hard to do much about this though.

So THG claims to have measured the current draw of an operating CPU? Unless AMD provided them with the test setup, I don't believe it.

on my very best day I'd not be up for that procedure
jeez

I can think of one other way which is more likely to work.

Put a small shunt resistor (again, a short wire with Kelvin connections) in series with each of the switching inductors of the Vcore regulators.

Attach a high bandwidth differential amplifier to each of these Kelvin connections to get a signal that gives decent resolution on a scope. Look at all three or four of the amplified signals on a scope, and gauge the current from these signals. These signals are going to be very noisy so it's going to take some interpretation to know what is relevant.

Not going to be that accurate...

Okay, yet another method. Most practical of all.

Just remove the Vcore regulators. Hook a seriously beefy bench supply up in their place. Measure the current coming out of the bench supply. The biggest problem with this (other than finding the supply) is getting the bench supply turned on at the right time. I have no idea what the power sequencing requirements of a PC motherboard are. Probably have to rig something up to get the bench supply switched in at the right time. Hopefully, not too much hardware gets destroyed in the process of getting it right.

[I was supposed to be working on a spreadsheet tonight. It will have to wait.]

Nor I. Here's the link I was referring to:

http://www17.tomshardware.com/cpu/20...ooler5-03.html

They're doing a variation on the second method I mentioned.

Instead of shunt resistances in series with the inductors of the Vcore regulators, they have loops of wire going through a clamp on current probe.

Accuracy most likely sucks, because of the high frequency 'noise' I mentioned earlier. Clamp on ammeters generally have lousy accuracy anyway. High frequency signals, on top of the DC component that the ammeter is attempting to measure, is not going to help the accuracy any.

Still, it is a way to get some kind of number...

Another issue, is that not all of the power dissipation is going to occur in the CPU itself. Some fraction is going to be dissipated in the Northbridge. This is a wild ass guess, but I wouldn't be surprised if 5% of the Vcore regulator's output power, is dissipated in the Northbridge during very memory intensive operations.

I ask because I have 3 issues against me at the moment:

1) what is the Real CPU temp
2) what is the REAL power applied to coolers from CPU
3) what are the importance of the secondary cooling paths

I have (1) under control. So 2 unknowns left. I have access to Bill's test data for a liquidCC surge waterblock, and I have all the tools necessary to test it as a function of flow rate and generate CPU die, wb inlet, wb outlet, and baseplate temps. So I can cross correlate to get ONE of the other two unknowns. Pick your poison; I am guessing that the power is easier to determine than the amount of secondary cooling that occurs.

So that is why I was asking so everyone is clear.

Thoughts?

Well, if you chill the water so that the CPU is at the same temperature as the motherboard in the socket area, and you insulate as per a TEC to eliminate convection, secondary paths should at least be substantially reduced.

Running a program that stays in the CPU cache without RAM access will minimize Vcore power dissipation occurring in locations outside the CPU.

It seems to me that the biggest problem by far, is getting a trustworthy current measurement.

Can you identify the manufacturer and part number for the Vcore regulator IC on your motherboard? Maybe looking at the datasheet for it will give me some idea how to measure the current accurately.

Using an Epox 8K7A+ at the moment, which has the Semtech SC 2422 voltage regulator on it for Vcore:

http://www.semtech.com/pdf/sc2422a.pdf

Weren't you the guy that said he did enough electronics crap at work and didnt want to take it home with him?

I am WAY out of my element here so any help is much appreciated.

LOL

Yes, but it's a lot more fun when all I have to do is think about it and someone else does all the work.

Well not so sure you are gonna convince me that drilling 462 holes to match up exactly with socket pins is the right sort of work to be doing :)

Nice try though; I thought seriously about it for a minute or two.

Nah, the way to do it is to get some 0.010 foil, some spray on photoresist, and ferric chloride.

Print a cad drawing of the hole pattern on mylar.

Use the mylar to create a photoresist image on the foil.

Then etch the holes into the foil with the ferric chloride.

Something like that, anyway.

Better living through chemistry.

While we blue sky, here is some data (taken this time at 1Hz so the x axis is reliable):

http://www.procooling.com/~phaestus/prime95b.jpg

Somewhat easier to see the oscillations in temp. Still WIDE variances in FFTs where the FFT size is larger than the cache, eh? Would be interesting to calibrate a Morgan and a Barton (tho I am sadly without a Barton) and see how they differ.