CPU Diode calibration: Round 2
 

Ok many know of my original goal to solder wires directly to CPU and then submerge in water bath to calibrate with Digitec 5810 dual linear thermistors. Well the last two days I have been getting with the program.

First step: Soldering. I took care of this like the professional that I am; I used way too much solder on the wire and once I got a good connection with the pin I then used solder wick to clean the adjacent pins :)

http://procooling.com/~phaestus/progress.jpg

The key is to strip the wire nice and long and then wrap it like a noose around the pin. You can get a really good mechanical connection then and just solder the wire to the pin. Much easier to geta good joint.

Once the CPU was prepared, I had to set up a bath with stirrer. For this I used a Fisher Scientific digital hot plate magnetic stirrer with a stir bar:

http://procooling.com/~phaestus/testing.jpg

Supposedly accurate to 1C but I had problems with the hot plate part. It is just not designed for low temp precision work. I had very little luck getting the damn thing to warm the water at all. Instead, I ended up putting a pot of water on to boil and exchanged water in the beaker for hot water to get to a higher temperature. Oh I put the CPU in a layer of Saran Wrap and held it with the core completely submerged and not touching the side of the beaker. The other temp probe is that of a Digitec 5810 dual linear thermistor. For diode readings I am using a Maxim 6655 diode reader system that connects to parallel port and has 0.125C resolution. I was hoping to get close to that in accuracy after calibration against the digitec :)

Here's the result:

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

Looks great! BUT

Those are the average of about 50 readings (4/sec) of the maxim reader per point. The average has a very good correlation obviously. But the std deviation is quite large (0.45-0.55). I suspect that the magnetic stirrer was the exact wrong thing to use and that it was generating noise that was making its way into my diode readings. I got the CPU a little wet, so I will try the calibration again tomorrow (or maybe later tonight) sans stirring.

Your thoughts?

use a controlled oven

First of all, you need a link to the MAX6655 datasheet.

Here you go.

I'm repeating myself, but...

If that flux is still on there you definitely need to remove it. It's a leakage path and the amount of leakage will vary with humidity. Look at the "Remote Temperature Error vs. PC Board Resistance" curves. Major bad news. Also depending on which pin(s) the leakage path 'connects' to, the signal leaking in may vary depending on what the processor is doing.

Flush it with isopropyl to remove the flux. Depending on how long it's been on there, you might need to brush or scrape at it. A toothpick would work to scrub the flux off.

According to the specs, (in particular the "Remote Temperature Error vs. Common Mode Noise Frequency" graph) the setup should not be very sensitive to common mode frequency below 1 MHz. I'd be somewhat surprised if there is much signal coming from the magnetic stirrer that is above 1 MHz, but it's hard to say considering the somewhat chaotic motion those things undergo. In any case, if there is a common mode noise issue with the stirrer, those ferrite beads and the amorphous metal cores I sent should help dramatically. Just slide them over the shielded twisted pair cable. (Even if high frequency common mode noise is not an issue with the stirrer, it will be with the CPU running.)

Differential mode noise is more likely a problem with the stirrer. The 'loop' between where the red and white wires exit the gray jacket of the cable, and connect to the diode pins will be prone to injecting differential mode noise into the measurement. The red and white wire should be kept twisted together as tightly as is practical. It looks like the red wire is longer. If you don't want to redo the soldering to tighten the loop up, I'd suggest at least folding the red wire back on itself to minimize the loop.

The resistance of the connections is not that important because you will be calibrating out its effect with the Digitec. The issues above are substantially more important, because they involve things that can change the calibration of the system.

I design electronic instrumentation. I've been burned by all these issues multiple times. I hope you won't be.

Just looked at an individual file; it's definitely some electrical interference of some sort:

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

Just fired it up with no stir plate running (CPU on bare desk) to see what the log would look like:

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

That is just noise and on the order of +/- 0.125 not +/- 0.5. I will give the stuff you sent a go on tomorrow to see if I can eliminate some of that. I don't think I am going to resolder the pins, but it isn't too much trouble to solder onto the maxim pcb :)

Definitely most encouraging.
Taken the liberty of comparing Bulit's* calibration :-

http://www.jr001b4751.pwp.blueyonder.co.uk/Phaestus.jpg

Is your Digitech probe in the Saran Wrap bundle?

* http://www.ocworkbench.com/ocwbcgi/u...&f=37&t=000245 from http://www.jr001b4751.pwp.blueyonder.co.uk/Bulit1.jpg

No it wasn't inside the bundle. There were a few issues other than the noise from the stirrer that will have to be addressed on tomorrow. I am going to have to think on the best way to proceed. My highest temp point is a bit suspect; I noticed the water had leaked into the bag slightly after that point and as Sean mentioned the flux resistance will change with humidity.

I think this is on the right track though (as Hammers are released and make it irrelevant)

That is nice looking data. (With the stirrer off.)

Standard deviation for the CPU sans stirrer is 0.085C according to Excel. That isn't bad; maybe it could become a little more stable with more shielding or removing that flux from the chip. I am loathe to mess with it too much though; don't feel like soldering it again and I am all out of spare CPUs at the moment.

With the stirrer on std dev is 0.488!

The flux won't have an effect on short term stability. (At least until the CPU is running, and switching signals are leaking varying amounts of current to the diode pins.) It'll definitely be a long term drift problem though.

The biggest problem is the loop. Perhaps you could cut the red wire and splice it back together shorter, so that the red and white wire can lay closer together?

It might be possible to pull a bit more of the red wire through the shielding; I'll look into it tonight.

More!
 

Managed to get better data today with a styrofoam ice chest and no stirrer. It isn't as easy as it sounds to get water to stabilize within 0.01C for a minute to record temps...

Here is the obligatory graph (I regraphed yesterday's results as well because numbers are VERY close)

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

And for Les:

MUCH better this time. I have a suspicion calibration will continue to improve as I practice. That oven sounds like a good purchase though I have to say...

looks pretty good
be interesting to see the distribution run over several days

looks like Les quietly updated his graph. It seems that some deviation from linearity is beginning to occur above 40C. Both times I had to stop at that point because some water had gotten into the "CPU baggie" though. I will give it a go again tomorrow and see if I can't get numbers up closer to 70C. I assume that with enough reps and points that even if the line curves a bit I can reproduce it with an equation...

Ta for raw data.
Looking even sweeter.
Yes quietly added second calibration data to previous graph.

Hoping have used correct interpretation of "Std Dev" the following may interest:-
http://www.jr001b4751.pwp.blueyonder.co.uk/Cal.jpg

http://www.jr001b4751.pwp.blueyonder.co.uk/Cal1.jpg
http://www.jr001b4751.pwp.blueyonder.co.uk/Cal2.jpg

EDIT: Edited graph at 11.47 am

Thanks Les. Of course a polynomial will fit the data better though :)

Perhaps another few points from 40C to 70C will sort things out. I'm tempted to take the whole damn thing to work and use a constant temp water bath. Would be a lot of crap to move though. In the short term I will fire it up again tonight and see if I can get another run or two in.

Ignoring (4.55, 5.56) get the interesting :
http://www.jr001b4751.pwp.blueyonder.co.uk/Cal3.jpg
http://www.jr001b4751.pwp.blueyonder.co.uk/Cal4.jpg

interesting indeed. I am eager to get back on this today; maybe I will head home early :)

A big load tonight...
 

With a polynomial fit (not so sure it is warranted over linear):

http://www.procooling.com/~phaestus/cal3.jpg

Also for Les again the raw data:

Didn't do std devs as I stupidly closed Excel 2/3 of the way through the data reduction but they ranged from 0.06 to 0.11.

I need something better than a bowl to get any better than this.

I plotted last night's data and tonight's all on one graph. Check THIS out:

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

The ONLY outlier is the 40.62C point from last night. It was at that point that I quit because water had leaked into my baggie. If you throw that point out the two runs are pretty much perfectly reproducible:

http://www.procooling.com/~phaestus/40cremoved.jpg

(edited to put in the second graph with outlying 40.62C point removed)

I'm just curious, what exactly is the rationale for not having the water touching the CPU? It's not going to hurt it, I've washed one of mine under water to get thermal paste off before....

The *package* is able to absorb water.

(edit: not the core)

The core absorbs water? Wow, I didn't know that. Okay, it makes more sense now :D.

good work pHaestus
I think you have nailed down one leg
any way at all to get better resolution ??

KnightElite
all epoxies are slightly permeable (weight gain and swelling)
this is why direct-die cooling may work short term, and will never work long term

Not using the MAX6655.

The part only has an 11 bit ADC and one of those bits is a sign bit.(positive or negative)

The way they have it setup, the part can output data representing -128C to +127.875C. (Spec'd operating range appears to be -55C to +125C.)

Looking only at the positive range, and dropping the ADC's sign bit, the resolution can be calculated at:

127.875 / ( 2^10 - 1 ) = 127.875 / 1023 = 0.125C

Maxim would either have to use a higher resolution ADC or shrink the range of temperatures that can be encoded. Either's unlikely due to marketing considerations. (It's too bad they didn't leave off the negative part of the temperature range, and double the resolution on the positive range.)

Replacing the Maxim IC with a homebrew setup is a possibility, but it would be very expensive by comparison.