Calibrating a CPU diode

[pHaestus]

Ok there has been a fair amount of discussion on the forums of late regarding making diode readings a bit useful. Since I have been idle on the site for a while, I decided to quit screwing around this evening. I soldered some copper twisted pair wire to the DXP and DXN pins of my 1600+ CPU directly. Yes you read that right. I have some familial obligations at this moment, but later tonight I will see if the chip still works. If so, the next step is to calibrate vs a Digitec 5810 in water baths using my MAX6655 evaluation system (0.125C res). Here's an obligatory pic:




I used a rubber band to hold the CPU onto a heatsink to hopefully help it withstand the heat a bit better.

Update later tonight...

[pHaestus]

The diode still works:



It is external sensor #1

I'll have to cut up my socket to actually install the chip; that is a project for later tonight. Then I will do a calibration tomorrow. I wish I had a constant temp bath.

Thanks to BrianS and KnightElite in the Pro/Chat for giving me the confidence to do this

[KnightElite]

Good to see the diode lives, unlike the one in my XP1600+ .
Hope the CPU still works also.

Pretty hardcore mod, indeed.

[BillA]

hey !
good stuff

I'll lay you dollars to donuts you can't raise the coolant temp (inlet to outlet) 0.25°C at 1.0 gpm

[pHaestus]

The CPU still works! I tested it out by putting some more wire under the corners to hopefull keep it kind of level and then putting the heatsink back on. It booted up and then started crunching K7Burn without issue:



I am going to have to modify the socket to use this CPU since wires are sticking off of it, but it shouldnt be a huge problem.

More to come after a calibration attempt. I am guessing getting stable water temps will now be my next challenge...

[BillA]

for the purpose of calibration, I don't see the absolute constancy of the coolant temp as a big issue
unless of course you are trying to relate the CPU temp to the coolant

I would suggest putting the whole shebang in an enclosure, w/o circulation but filled, and plotting the readings over a range of temps
- the real difficulty with this is that the 'lag' of various parts will be diffferent so the determination fo equilibrium is crucial
(now the problem is holding everything at various temps to establish such)

you cannot have any temp differential between the CPU and the coolant temp/inlet sensor or you will have the whole host of temp offsets across the chain to deal with
- zero heat flux during calibration

I am assuming that everything is going to be 'calibrated' against (wrt) the coolant inlet sensor

[bigben2k]

Congrats pHaestus, I'm glad to see you were the first to take the plunge!

Now maybe you can write an article, as a follow up to your previous one, so that the rest of us can mod our rigs to measure temps more accurately (at least those of us stuck with a mobo that tries to read the CPU diode).

This could become a ProCooling standard, for anyone who posts a claimed temperature.

[BillA]

Ben

did you disconnect YOUR CPU ?
the one between your ears ?

did you read pHaestus' posts ? and mine ?
clearly you missed the nitty gritty

pHaestus is using a Digitec system with 0.01°C resolution
- please name a single OCer with that kind of capability
given that a 4 to 10 fold greater measurement capability is needed than the achieved accuracy,
we can assume that pHaestus will attain a solid 0.1°C accuracy

and thats the hardware end of it:
calibration - just who is going to do all the work that pHaestus is just beginning ?

or are you (NOT) thinking that the values derived by pHaestus will be 'valid' for everyone ?
arrrgh

I sure hope you never define "ProCooling" standards, we will all need Plexiglas bellybuttons

[pHaestus]

The big problems with others reproducing this are:

1) My diode reader is not of the DIY variety, but instead is the full package from Maxim on pcb connected to the parallel port of a second PC. This gives me the "extra bits" from the 6655 for 0.125C resolution. Not many will want to drop $100US for the package (though it is a very nicely made setup).

2) When you turn the CPU over and get out the soldering iron, you tend to lose your nerve. If you get the pins too hot, you will kill the diode. If you get the core itself too hot, you'll kill the chip. Your iron may not be grounded and so ESD is a concern as well. Not to mention, it is pretty ****ing insane to just start soldering copper wire to the CPU. To top it all off, when you are done then you have to hack up the top part of the socket to make room for the wires.

I can't really see anyone else doing something like this. The whole time I was getting the nerve to do it, I was thinking about how STUPID an idea that it was.

[bigben2k]

was refering to this article, which allows Joe user to measure the CPU temp.

Since then, diode reading has been added to most mobos, but this prevents us from doing that mod, because having two readers on the same diode gives an incorrect reading.

With what pHaestus has just done, that article can now be updated/expanded so that we can get back into measuring CPU temps to the closest thing to accurate that the average Joe can get. As for calibrating, it would be possible to include a procedure for that too! It'll be rough, but it's still a vast improvement to the +/- 10 deg C variance we're encountering.

Now do you see where I was headed?

[bigben2k]

Oh, I don't know, maybe I'm just way off...

I for one certainly would jump at the opportunity to do this mod.

[pHaestus]

Someone in the other thread gave a MUCH more elegant solution to people with faulty diode readings or COP instead of a proper diode reader.

Use desoldering braid to remove all the solder from the diode pins on the backside of the motherboard. This will break the connection with the onboard reader. Then you solder the wires of your maxim reader to the remaining socket pin only. I would do this 10x before I soldered on a CPU directly once.

I don't want people to get the impression that this was a "safe" or "smart" thing to do at all. It was a retarded idea but drastic measures were required to get me where I needed to go with diode calibration.

[BillA]

never any doubt about your direction Ben, nor any argument with its desireability
but you must temper your 'hope' with a realistic technical assessment of all the tools at your disposal
vs. those necessary to achieve the task

flying involves more than flapping the arms

[bigben2k]

Well now see, I thought that the desoldering option was ludicrous, and I for one, preferred your wire-wrapping suggestion. Otherwise I agree, soldering directly to a CPU pin is something I'd avoid.

For the benefit of everyone else, here's the other thread.

[Brians256]

Congratulations, pHaestus! I'm glad I could lend a couple of kilograms of brass for your growing family jewels. Thankfully, the pins you were soldering to weren't in the middle of other pins!

[Since87]

Originally posted by pHaestus
Use desoldering braid to remove all the solder from the diode pins on the backside of the motherboard. This will break the connection with the onboard reader. Then you solder the wires of your maxim reader to the remaining socket pin only. I would do this 10x before I soldered on a CPU directly once.

I don't want people to get the impression that this was a "safe" or "smart" thing to do at all. It was a retarded idea but drastic measures were required to get me where I needed to go with diode calibration.

I'm not clear on how you are going to isolate the diode pins from the socket pHaestus. Cut them down so they don't reach?

Speaking from a lot of experience, I'd say using the solder wick on the MOBO is likely to be more challenging for most, than your technique. It's not necessarily easy to get enough of the solder 'wicked' out of the hole to have a clear separation between pin and hole. And, even when you do have a clean separation, unless you fill the gap between hole and pin, there is a risk of the pin making contact with the sides of the hole. (Pressure on the wire when you mount the motherboard.)

It looks like you used 24 gauge wire out of a Cat 5 cable. You could make it much easier on yourself with 30 gauge 'wire wrap' wire. I don't know if you can still get such at Radio Schlock, but any real electronics store would likely have it. (I'll send you 5 feet of it along with 36, 40 and 44 gauge magnet wire if you think you might want to try this on a different CPU.)

I think you are overestimating the risk of damage due to heating from the soldering iron. There is going to be a lot of copper to spread the heat between the pins and the die itself.

The difficulty in doing this is largely a matter of perspective. There are a couple ladies sitting less than 50 yards from me who wouldn't even bat an eye if I asked them to do either of these techniques. (Other than to give me a look of, "Oh, you screwed up and now we have to suffer the consequences." The answer to that is to come bearing chocolate chip cookies.)

[jaydee]

I am courious why you cannot just solder the wires to the backside of the mobo instead of the CPU itself?

Here is an example of the multiplier mod, but why couldn't you use the same mathod for this?
http://www.lostcircuits.com/cpu/amd_unlock/1a.shtml

[bigben2k]

JD, the problem is that when you add your own reader, your readings will be thrown off by the onboard reader (if your mobo has one).

[Since87]

The MOBO's temp measurement circuitry will interfere with the circuitry pHaestus is adding.

[jaydee]

Ok, so he cut the pins off the CPU so they will not connect to the mobo? Hummmm, I have a couple Epox 8K7A's that shouldn't read the internal diode. Would it work if I used the back of the mobo instead or are these connections still used for the diode under the CPU in the middle of the socket?

Now if you blocked this connection somehow will the mobo react wierd? Is there any monitoring in the bios that requires a connection from these pins or it will not boot, ect..? If not you could just cut the traces on the mobo, or dremel them out? I am up for destroying some stuff!

Sounds like a very interesting project. Might be cheaper than the die simulator I have in the works and you could test on a actual computer.

[Since87]

Originally posted by jaydee116

Now if you blocked this connection somehow will the mobo react wierd? Is there any monitoring in the bios that requires a connection from these pins or it will not boot, ect..? If not you could just cut the traces on the mobo, or dremel them out? I am up for destroying some stuff!

I've wondered about this as well. I wouldn't be surprised if some BIOS'es wouldn't let you boot with the diode removed from the circuit.

One advantage of doing things as pHaestus did, (I'm assuming he'll cut the CPU pins down.) is that a diode can be attached to the MOBO to keep the onboard stuff happy.

It might be possible to cut the traces on the MOBO, but there are multiple layers of traces in the PCB, and the odds that the connections to those two pins are accessible from the bottom are low. You might have better luck finding and cutting the traces near the diode reader IC. But then, if you do have to hook a diode to the diode reader to keep the BIOS happy, you may have a difficult time of it. (The pin spacing on the QSOP type packages these diode readers come in are are 0.025". Very tricky to solder a wire to an individual pin at that spacing.)

[murray13]

BillA's correct in one respect that the accuracy needed to measure temperature with certanty to less than one degree is beyond any average person. In water cooling a difference of 1deg C is a big deal. But most of us don't have a way of measuring a 1deg C difference. PERIOD. I'm really not going to get into all of the technical difficulty in doing it. Just say that temerature measurement is one of the least accurate basic things we measure. There are things less accurate but most are derived from the basics.

As an example we can measure:

Time +/- .000000000000005 sec
Voltage +/- .00000002 V
Mass +/- .0000002 kg
Temp +/- .00004 C

From NIST the people who calibrate the standards.

I personally don't care to know if I'm running colder than someone else. I agree it would be nice to be able to compare.

The only thing I can think of is using a ice bath. An ice bath is a insulated cup (styrofoam) filled with ice chips then filled with enough water just to cover. The ice and the water used should be pure not tap. Then while using keep stirring it. You can get 0 deg C +/- .1 to +/- .01 deg C using this method depending on your technique. In theory one could use this reference to see how much your temp sensor is off at 0 C. This correction can then be applied to all values. Realizing that most PN junction (diode) thermometers have a non-linerarity of about 1 deg C you can get close to being within 1 deg C.

Getting beter than 1 deg C readings is just well beyond what most enthusiasts will do.

[pHaestus]

The issue is not dealing with the motherboard's circuitry in my case.

Some observations:

The use of a diode reader tapped into SMBus just doesn't work for testing. I am seeing 5-10C increase in temperatures by booting the test PC into DOS mode and running K7Burn while monitoring the CPU temperatures with a second PC.

And what of calibration? What value is a diode reading without knowing how the diode deviates from the true temperature? So my major concern was:

Once you have a way to measure the diode temp, how can you calibrate it?

One could in theory put the whole system into a temperature-controlled chamber, adjust the temp, and then record what the diode reader reports vs. what the actual air temperature is. I lack this equipment or the energy to pursue this avenue.

Another idea is to calibrate the diode and reader vs. a much better thermometer in a water bath. By adjusting the temperature of the water over a fairly wide range one can MEASURE the error associated with the exact diode in the CPU you are using and with all your wiring. This is no problem since the CPU works fine as a diode without power. What IS a problem is that if you calibrate via one method (maybe wire wrapping to the pins) and then use a different method to actually connect the reader to the motherboard (soldering onto the socket pins) then your calibration is worthless. We are talking about resistances in the ohm range substantially affecting the diode reading. This is incidentally wh

Based on all of this, I decided to permanently affix the wires to the CPU pins. Now I can use the exact same setup (wires, CPU, reader) in calibration and in testing.

This was the only way I could devise to approach reasonably useful numbers for a diode reader.

Hope that clears things up...

[jaydee]

Originally posted by pHaestus
What IS a problem is that if you calibrate via one method (maybe wire wrapping to the pins) and then use a different method to actually connect the reader to the motherboard (soldering onto the socket pins) then your calibration is worthless. We are talking about resistances in the ohm range substantially affecting the diode reading.

I get it. The calibration process is what through me off.

[pHaestus]

Oh and jaydee I have put a diode reader on an 8K7A before via soldering onto the socket pins without incident. I DID have to tap into SMBus via the pins from a DIMM slot, but it wasn't a huge deal. I then just grabbed 5V and ground from the backside of the ATX connector and fired everything up no worries.