CPU Diode calibration: Round 2

[Groth]

Prototype done. KX7 soon to get the treatment.

Late for work - details later.

[bigben2k]

I came across an interesting article, and for some reason, I kept thinking about this thread, even though the application described is not the same:

http://www.e-insite.net/tmworld/inde...e=6%2F1%2F2003

I thought it might inspire someone.

[pHaestus]

Would it be possible to get some more detailed info on how this was done Groth? I am guessing that with the nichrome resistors that since87 suggested and your design that I might have something doable?

[Groth]

Do-able, indeed. Actually, the whole thing was fairly easy.

For ye olde K6, I paralleled 5 generic 0.1 Ohm, 3% resistors. For the upcoming KX7 project, I have ten resistors from the same company (Ohmite), but I've chosen the 10 series for their higher precision. Their temp-co isn't as nice the nichrome, so they'll need a fan when doing important work. Depending on what I find when I pull the inductors, I'll use 6 or 9 for the sensing.

A length of twisted pair was attached to the shunt and led to the amplifier stage. I used a TI LF412 dual JFET-input op-amp, half as a unity gain difference amp, half as an inverting amp with a gain of 21. For the Abit board's turn, I'll add another LF412, so that I have 3 difference amps, one for each inductor, and an inverting summing amp with a gain of 75-80 or 110-120.

Next is the isolation stage. I didn't want the reading/recording of date to affect the target computer's workload, so the I passed the amp stage output and a Vcore reading through a pair of ISO124 isolation amplifiers.

Last, but not least, the isolated signals were digitized with a Nat Semi LM87. I used it because it is noise tolerant, easy to interface, and supported by Motherboard Monitor. It only has 8 bit resolution (around 0.2 A resolution for the K7's), but, I already had one in the house.

New brake rotors first; schematics soon.

[Since87]

Sorry to leave you hanging here pHaestus.

Nice work Groth.

Do you know of any good schematic capture shareware? I don't have any good way of creating schematics at home.

I hope you don't mind some suggestions...
Maybe you've thought of some of these things already, but I thought they were worth mentioning just in case.

Originally posted by Groth
I have ten resistors from the same company (Ohmite), but I've chosen the 10 series for their higher precision.

I'd go for minimal TCR. Actual resistance can be measured and compensated for in the following gain stage. Drift due to variations in ambient and airflow can be a real pain.

I don't know what value of shunt resistor you were going for, but keep in mind, the larger the shunt resistor, the more of your Vcore adjustment range you are giving up. On the other hand larger shunt resistors will give better signal to noise ratio.

Originally posted by Groth
A length of twisted pair was attached to the shunt and led to the amplifier stage. I used a TI LF412 dual JFET-input op-amp, half as a unity gain difference amp, half as an inverting amp with a gain of 21. For the Abit board's turn, I'll add another LF412, so that I have 3 difference amps, one for each inductor, and an inverting summing amp with a gain of 75-80 or 110-120.

I'd suggest adding a film 0.1 uF cap in parallel with the shunt resistors. This will bypass a lot of the high frequency transients that come through the capacitance of the switching inductor. These transients might screw up the the amplifier output somewhat, unless the amplifier is fast enough to track them.

I'd suggest using OPA227 Op-Amps. (OPA2227 or OPA4227 for duals or quads) With your present setup (20 mOhm shunt, 9A max current) the input offset voltage of the LF412 can cause a +/-1.7% error in the first difference amp alone. The main advantage the LF412 has is low bias current, and because of the extremely low output impedance of the shunt, bias current is a nonissue. The OPA227 has substantially better: input offset voltage, input voltage noise,
common mode rejection ratio, and unity gain bandwidth.

I can get these at my employer's cost. If you don't mind working with surface mount packages, (SO-8) I'd be happy to send you some. (I might be able to get the dip package in singles.)

I've got more to bring up, but it's late...

[Groth]

Woo hoo, my car stops now. Slide hammer - fun toy.

Do you know of any good schematic capture shareware? I don't have any good way of creating schematics at home.

I've been using LTSpice. Linear puts it out free to encourage the use of their switcher components. It isn't the greatest Spice implementation, nor super for schematics, but the price was right and the learning curve agreeable. I've also been trying to forge my way through the free version Vutrax, with little result.
At home, eh? What packages have you at work? Perhaps they have a cut down share-/free-ware version.

I hope you don't mind some suggestions...

Bring it on...I've more enthusiasm than knowledge and experience.

I'd go for minimal TCR. Actual resistance can be measured and compensated for in the following gain stage. Drift due to variations in ambient and airflow can be a real pain. I don't know what value of shunt resistor you were going for, but keep in mind, the larger the shunt resistor, the more of your Vcore adjustment range you are giving up. On the other hand larger shunt resistors will give better signal to noise ratio.

I lack the equipment to accurately measure resistances below a tenth ohm (&$^%# DMM), making 3% tolerances mighty ugly. Cheap upgrades? I could have fairly wide uncompensated temperature swings with the series 10 before error reached 3%.
On the K6 board, I discovered that the regulator was not compensating for the shunt - the nominal 2.4 V was appearing on the inductor side, not the CPU side. I'd like to think that KX7 has better feedback circuits, but, to be safe, I'll use a shunt in the 0.5 to 1.0 mOhm region. This will also decrease the self-heating and associated drift.

I'd suggest adding a film 0.1 uF cap in parallel with the shunt resistors. This will bypass a lot of the high frequency transients that come through the capacitance of the switching inductor. These transients might screw up the amplifier output somewhat, unless the amplifier is fast enough to track them.

I used a RC low pass filter on each side heading into the difference amp. 1 Mhz corner frequency versus 3 Mhz unity gain bandwidth. I'll add a shunt bypass cap, and see if it anything changes.

I'd suggest using OPA227 Op-Amps. (OPA2227 or OPA4227 for duals or quads) With your present setup (20 mOhm shunt, 9A max current) the input offset voltage of the LF412 can cause a +/-1.7% error in the first difference amp alone. The main advantage the LF412 has is low bias current, and because of the extremely low output impedance of the shunt, bias current is a nonissue. The OPA227 has substantially better: input offset voltage, input voltage noise, common mode rejection ratio, and unity gain bandwidth.

A case of using what I had in stock. As for the input offset voltage, a bit of hand trimming while using known inputs killed it. Power supply noise/regualtion was problem; cured with 10V zeners and filter caps.
I'm more than willing to pop some OPA227s in to compare performance. If nothing else, I'd be a fool to turn down parts at bulk prices. Surface mount is wonderful, saves tedious drilling.

[Groth]

This is what I used for the difference stage. I actually used a 200 Ohm trimmer in series with a 9.9k resistor where it shows the 10k's. With precision resistors/amps, you could get away without the timming. Plus I'm just lazy with the drawing...

[Since87]

Originally posted by Groth
I've been using LTSpice.

I'll check it out. I use Orcad at work. There is a "student" version that's downloadable but it's extremely limited, and would be unethical for me to use anyway due to the 'student use only' language.

Bring it on...I've more enthusiasm than knowledge and experience.

Great, I've got more knowledge and experience than enthusiasm. Maybe if pHaestus has more money than brains, the three of us can get something pretty cool put together.

Actually, it sounds like your general electronics knowledge is very good. I've been working for a power measurement instrumentation company for the last 10 years. I've got too much experience at this for it to be much fun, but I want to see more graphs out of pHaestus, so I'll put the effort in.

Perhaps the three of us could do something like:

Groth and I work out a design that produces a voltage output scaled for use with the MAX6655's 3.3V A/D input. (Need an accuracy target. I'd suggest +/-1% max from 10C to 50C. The MAX6655 will add +/-1.5%.)

pHaestus pays for the parts to build two of these. One for himself and one for Groth.

Groth builds the two circuits.

I calibrate the two circuits, and do any mods required to get them in spec. (I'll even temp chamber test them.)

Are you two interested in an arrangement along these lines?

[pHaestus]

How many bananas are we talking about for parts?

[Since87]

Originally posted by Groth
This is what I used for the difference stage. I actually used a 200 Ohm trimmer in series with a 9.9k resistor where it shows the 10k's. With precision resistors/amps, you could get away without the timming. Plus I'm just lazy with the drawing...

I'd suggest the circuit below. (Largely because I have access to laser trimmed 20K center tapped resistor networks, with very good ratio accuracy and ratio tempco specs. I like to avoid trimming as well.)

I think it would be a good idea to up the capacitance to roll off at 10kHz. (or even lower depending on the A/D used) It's a good idea to filter out the current ripple as soon as possible.

I'd also swap the two input signals to put the 'noisier' signal into the noninverting input. (The way I've redrawn the circuit, the noninverting input will do a better job of filtering at high frequency because the cap goes to ground rather than the output. The amp will have a significant output impedance at high frequencies.)

Anyway, once again it's late. I hope I'm actually making sense. Enough for tonight.

[Since87]

Originally posted by pHaestus
How many bananas are we talking about for parts?

Depends.

If the circuit provides an output voltage proportional to the CPU's current draw, I'd say parts cost could be kept between $50 and $100.

If the circuit provides an output proportional to Wattage, I'd guess more like $100 to $150.

But at this point, these are really rough guesses. I think those are conservative numbers, but not enormously conservative. I realized today that I probably have a pile of obsolete circuit boards I can pull some, if not all, of the OPA-2227's from. That could make a significant dent in the parts costs. There may be a lot of other potentially useful parts I can scrounge. I'll have to take a look around at people's junk stashes at work to see what I can get for free.

[Groth]

Did a bit of playing around. I tried various values and types of capacitors across the shunt. The biggest change I saw was with a 1 uF polyester cap; the 2000 sample average for CPUburn's current draw dropped by 0.014%.

Connecting the inverting input bypass cap to the output is a good concept to know. One of my early versions, with the inverting input bypassed to ground, had some ugly feedback problems. I moved/replaced the filter caps as per your suggestion, using 4.7 nF ceramics (largest I had in 0805, which is all there was room for). Results: average current reading down 0.03%, population standard deviation down 0.72%. :shrug:

As for the rest (up too late to quote), I'll be making multiple prototypes and versions, anyway, so, I'm open to doing the building. I don't think there's any need for a watt output - it's easier and cheaper to do the multiplication in software.

Mmmm, laser trimmed resistors...

[Groth]

OK, trying to get a handle on accuracy...

Quite by accident I discovered that when a floppy is left in the drive and the computer complains "non-system disk", it uses a steady current. A current that varies with Vcore and multiplier, but is boot-to-boot reproducible within a 10 mA range. All the switcher noise, non of the software variability.

Using the known gain of my circuit and the nominal LSB value of my ADC, I computed the expected byte values for a series of multimeter readings. I then divided the actual value, as recorded by MBM, by the expected value.

For the statistics crowd, variance comes up as 0.3%. Not bad at all for a 2% ADC and 1% (in this range) multimeter. Unfortunately, the shunt has a 3% tolerance, so overall accuracy still sucks. I wonder how it will fair versus a fully calibrated, precision Since87 design.