Kelvin Varley Divider [and Precision Voltage Source]

[amspire]

I think easy calibration is the way to go for the average bear. What I don't know is the magnitude of the error sources that "come out in the wash" when the Analogic unit is trimmed. That trimming might have compensated for switch resistance, leakage and who knows what else. The unit did step in 10 uV steps on the 10V range. IMO, one of the nice things about the way they did it was being able to go a bit over 1 or 10V; It's very common to need that feature. I like the idea of modules for each decade. They could have Samtec or similar pin headers so you'd plug them together, creating a block.

Ideas for the construction definitely would help.

I would put the summing resistors on the main board as the design of the summing junction is the critical part of the design.  When you are summing currents from a  260Mohm resistor, then I think  shielding surrounding the main summing resistors would be a good ides - just the adjustment pots would be outside the shield.

So you would end up then with all the decade modules being absolutely identical and unless a offset zero pot is added, the modules would have no adjustment.

The way the modules are connected is critical. If there is 1mA flowing through the main reference divider chain, then a total leakage of 1pA causes a 1ppm error. It will be almost impossible to get 1pA absolute leakage. The 4051's are a big concern, and it may be necessary if you are chasing the 1ppm to add extra voltage followers on each of the divider outputs.  It also means that the divider must be calibrated with all the actual loads, so we will need to come up with a method of calibrating all the stages of the divider that does not involve matching resistors.  I think it can be done with a combination of a precision 1:1 resistive divider and a 10:1 resistive divider - ie the elements of a Hamon divider.

But also issues of PCB leakage currents and thermal voltages come into it. It may be best to pit right angle pins on the modules and solder them to the main board.

Doubt it makes much sense to go beyond 1 ppm in terms , as the resistor and reference requirements get way out of hand, not to mention thermal issues. Analogic did make one horrible mistake. They used quadrature switches on the front panel for each decade, that didn't hold up well. The old units can be really unpredictable when setting the voltage, though whatever they read out on the display is what you get. Sometimes you turn it up and it goes down, sometimes down and it goes up. Contact cleaner only helps for a year or two and then becomes ineffective.

In terms of the number of decades, you are right that trying to go for more that 1ppm would be a bit too much. However, if we find opamps that give a real 1uV offset or less - whether it is through chopper auto zero, or through zero calibration, then it may be possible to have useful resolution down to 1uV or less with a low voltage output. you can achieve something similar by having multiple ranges on the voltage source, but every time you add something like a range switch, it is another 1ppm voltage divider you have to consider.

[amspire]

Here is another approach to throw into the mix.

It is the BCD tree used by Fluke in its Programmable Voltage Source like the 4210A. These were actually programmable supplies with voltages up to +/-65 V and currents to +/-1A.



Its merits are the big choice of switch IC's for each decade, 5 resistors per decade, all resistors in the range of 25K to 200K and only 5 different precision values are required.

Also the ranges overlap - each decade can go up to 15 and so 10 on a lower range can be calibrated from 1 on a higher range.

Problems are the same as a R-2R tree - calibration changes in the lower ranges affect the higher range calibration, so the calibration cycle may require going around the calibration loop until everything settles down to the calibration spec.

Fluke were only after 0.01% accuracy in voltage sources, so they limited the calibration problem by limiting the number of calibration pots to about 5.  For 1ppm though, calibration may get too difficult.

Richard

[amspire]

I stumbled across these switches that look like they might be OK for a Kelvin-Varley divider. 2 pole 11 positions (which I prefer to 10 position).

Ceramic wafer, and it looks like silver plated.

$4 each with free delivery. Ebay Item 300562748033

[fmaimon]

In practice, any time you are chasing precision, it is always the detail that you have to be obsessive about.  You have to consider all the factors you normally do not have to consider in other designs. If you are using analogue switches you have to look at the switch resistance, the switch-to-switch variations of resistance, the temperature coefficient of the switch resistance, the OFF leakage current through the switch, and the leakage current from the supply rails to both sides of the switch.  The currents will probably rise exponentially with temperature, so you have to think of the maximum operating temperature.

That's why I'm here asking all these questions. I don't have much experience in precision stuff. If I didn't thank you and Conrad yet, let me thank you both for your insight and knowledge.

The beauty of the totally passive resistive-based dividers is that as long as you start with great resistors and great switches, there is nothing else to go wrong. After a few years, it stabilizes and remain incredibly accurate. Put it in the cupboard, drag it out in 20 years and it will still work perfectly.

Part of the genius of the Kelvin-Varley divider is that the whole divider can be build from the built from one single batch of resistors, all of the same value. If all the resistors have matching temperature coefficients, then errors due to ambient temperature changes cancel out. Hopefully, they will age the same too. In the summing method, you have a very wide range of summing resistors, all that have to stay accurate. Is the 1Mohm resistor temperature characteristics going to match the 1Kohm characteristics?  Yes if you were Fluke of 50 years ago, and you make all the resistors from wire of the precisely same composition, and yes if you are HP or Fluke today, and you can make a laser trimmed thick film resistor network, where all the resistors are made from exactly the same deposited film. Probably not if you are a hobbyist getting whatever parts you can find.

Be sure I'll build one pure KVD because of exactly that. Actually this message just made me realize that we went a little off topic and started discussing a precision voltage source. I'll change the thread title to reflect that.

I think your maths is a bit off with this one. Top marks for getting the values to work successfully as a tree with each stage having the same impedance as the stage below it. That bit is perfect. Very elegant.

When I first design this 8R-56R circuit, it seemed that it would work, but I didn't made all the calculations. After your post, I've applied the thevenin theorem to the circuit and saw it really didn't work... Thank you for pointing out.

No doubt that was clear as mud!

Actually that was very clear! Thank you for sharing.

If the MSD op-amp has a 1K resistor to the summing amp, the others have 10K, 100K, 1M, 10M and 100M for the LSD.

Liked it! Nice design, but isn't the 100M resistor a bit high? Isn't a little bit better to use a 0.1 hamon divider with it´s output a little higher (about 0,1%) followed by a 10M resistor?

I think easy calibration is the way to go for the average bear. What I don't know is the magnitude of the error sources that "come out in the wash" when the Analogic unit is trimmed. That trimming might have compensated for switch resistance, leakage and who knows what else. The unit did step in 10 uV steps on the 10V range. IMO, one of the nice things about the way they did it was being able to go a bit over 1 or 10V; It's very common to need that feature. I like the idea of modules for each decade.

Easy calibration is everything I hope for!
On going a bit over 10V, Richard's design can do that too. If you select the 10V and  the 1V outputs, you get 11V. With 6 decades, you can go to 11.1111V.

Good ideas with the modularisation of the design. I'll keep it in mind.

I stumbled across these switches that look like they might be OK for a Kelvin-Varley divider. 2 pole 11 positions (which I prefer to 10 position).

Nice find! But why 11 positions are better? Don't you only need 10 positions for a KVD?

Thank you,
Felipe

[amspire]

If the MSD op-amp has a 1K resistor to the summing amp, the others have 10K, 100K, 1M, 10M and 100M for the LSD.

Liked it! Nice design, but isn't the 100M resistor a bit high? Isn't a little bit better to use a 0.1 hamon divider with it´s output a little higher (about 0,1%) followed by a 10M resistor?

100M is a bit high, and using a 10:1 divider or a 100:1 is a great solution. Accuracy is not important - you can calibrate that out. Stability is important. One thing you can do is to use the divider networks that have great temperature matching, so you get good stability.  Since it is a one-off build, it is perfectly reasonable to manually select resistor combinations until it is very close to correct, then use a pot to adjust for the last 1 part in 100 of accuracy. Also the high resistances are for low order decades, so an an accuracy of 1% is adequate. You do not need to go overboard. It is the first 4 decades that you have to be obsessive about.

I think easy calibration is the way to go for the average bear. What I don't know is the magnitude of the error sources that "come out in the wash" when the Analogic unit is trimmed. That trimming might have compensated for switch resistance, leakage and who knows what else. The unit did step in 10 uV steps on the 10V range. IMO, one of the nice things about the way they did it was being able to go a bit over 1 or 10V; It's very common to need that feature. I like the idea of modules for each decade.

Easy calibration is everything I hope for!
On going a bit over 10V, Richard's design can do that too. If you select the 10V and  the 1V outputs, you get 11V. With 6 decades, you can go to 11.1111V.

Good ideas with the modularisation of the design. I'll keep it in mind.

I wasn't sure if my explanation for the easy calibration method was clear, but if you followed it, then great!.

I stumbled across these switches that look like they might be OK for a Kelvin-Varley divider. 2 pole 11 positions (which I prefer to 10 position).

Nice find! But why 11 positions are better? Don't you only need 10 positions for a KVD?

Thank you,
Felipe

To calibrate the resistors in a KVD stage, you need to be able to break the connection between at least one leg of the chain and the rest of the KVD. Going all the way with the Fluke 720a system with the whole calibration for the first 3 stages managed by switches is the ultimate, but they can afford to buy great switches  - and great switches can easily cost hundreds of dollars.  Having an 11th position that breaks the chain free of the next lower divider, combined with a calibrate procedure that perhaps using plugable wire links rather then Fluke's switches,  is a workable solution for a home builder. There would be other ways to do it I am sure, and some people will just hate having that 11th position.

There still need to be another switch that can disconnect all the resistor in parallel with the dividers, or maybe not with guarding during calibration? Exactly how how to make it all work will need some thinking.

Calibrating the KVD is all about keeping the box sealed and untouched as much as possible. If there was a removable panel that gave access to the adjustment pots, and the pins that you could link for calibration, that would be ideal.

You can safely match resistors to 0.01% before assembly, so like the Fluke, it is best if the lower 4 ranges have no calibration at all. That means they really only need 10 position switches. Just have to find good ones at a decent price. The supply of affordable good quality rotary switches is not what it used to be.

On the first decade, an 11th position is essential, as it means you can you can have a "1.1" input like the Fluke 720A. This allows you to do plus-minus variations around the 10.00000 mark. But you also need to be able to break the connection to the second divider, so does that mean we need a 12 way? Very possibly.

It is very annoying if you can only go to 9.99999 and no higher.

Richard.

[amspire]

I have ordered 10 of those switches, and I will test them when they arrive from Hong Kong.

If you order them in lots of 100, the price comes down to $1.75 each. (ebay item 290571818531)

Looks like they have good stocks, so I am hoping they are good switches.

Richard.

[Flavour Flave]

Linear techs Kelvin Varley Voltage divider 



You can get the schematic here:

http://www.edn.com/article/459869-Video_Design_Idea_Build_your_own_laboratory_precision_voltage_reference.php

[amspire]

There are some good ideas in that video.

Unfortunately, those dividers are a problem.

http://www.ietlabs.com/voltage-divider/esi-dp-1211-dp1311-voltage-dividers.html

He used a 4 decade divider. Price for 3 decade concentric dividers on the IET Labs website is $3940 to $4895, and they don't mention the 4 decade version at all. I think the 4 decade Tegam coaxial resistors are available from IET, but you have to contact them for details.

They are impressive though:

5ppm/C temp coefficient, 0.3ppm/mW power coefficient and short term switch repeatability of 1mOhm.

Richard

[Flavour Flave]

This would be way cheaper "How to make custom resistors"

http://www.all-electric.com/schematic/res_trim.htm

This could be a cheap way to make a divider?
Using a file, nail polish and cheap carbon resistors oh and metal film aswell.
There is good tip for tuning filters on that page as well.

[codeboy2k]

This would be way cheaper "How to make custom resistors"

http://www.all-electric.com/schematic/res_trim.htm

This could be a cheap way to make a divider?
Using a file, nail polish and cheap carbon resistors oh and metal film aswell.
There is good tip for tuning filters on that page as well.

Ouch.  That webpage hurts my eyes.  1999 called, they want their web back.

[Conrad Hoffman]

Carbon won't get you where you want to go. It's not hard to make precision resistors, at least for low frequencies. Get yourself a roll of manganin wire and wind them on mica squares, just like they did years ago. If you want to rig up a simple spot welder, you can get "800 series" wire. It can't be soldered (easily) but has very low TC, better than 5 ppm/C. I think many people used to trim the value (trade secret revealed here) by abrading the side of the wire using an electric eraser- the rotary drafting kind with an ink eraser tip. If you look inside things that use card wound resistors, you'll see this frequently.

When Jim Williams wrote on KVDs (not sure if app notes or his books) he talked about getting Julie Research KVDs for something like a few dollars at a surplus market. They were tagged "bad" because someone tried to measure them with an ohm meter, not knowing that you can't pull current from the pickoff point. I bought a used KVD directly from Julie. It was the prototype of the 307, and proudly wears serial number 1.

[Flavour Flave]

Carbon won't get you where you want to go.

Why is that. Is it to do with temperature?

I thought that I could knock up a Wheatstone bridge and file the carbon resistors to the same value or use metal film, but they are bit harder to file. Well the aim is get them to the same value, not an exact absolute value, to use in the KVD for a precise voltage reference.

Also that mica seems expensive(show me a source please) and wouldn't I need a lot of Manganin wire to do resistors for KV divider ?(for 10k 2k etc..). And a lot of winding.

Thanks
Richard

[alm]

Filing won't do anything about tempco or drift. I think there's a table in the Art of Electronics that indicates several percent change in value after soldering, load cycling and other kinds of stress. Don't remember if this was carbon film or carbon comp. The important spec for precision equipment is its drift (short term, long term) and tempco, not so much the actual value. This is why precision components are sometimes aged by manufacturers. You can cal out any change in value (eg. put a label with 1.010 ohm on it), but you can't compensate for variations, apart from indicating that the instrument has +/-2% tolerance.

I think there might be a job for you in China somewhere where you can use some blue paint to turn carbon film resistors into metal film, although I doubt they actually bother tweaking them to 1%.

[Flavour Flave]

Filing won't do anything about tempco or drift. I think there's a table in the Art of Electronics that indicates several percent change in value after soldering, load cycling and other kinds of stress. Don't remember if this was carbon film or carbon comp. The important spec for precision equipment is it's drift (short term, long term) and tempco, not so much the actual value. This is why precision components are sometimes aged by manufacturers. You can cal out any change in value (eg. put a label with 1.010 ohm on it), but you can't compensate for variations, apart from indicating that the instrument has +/-2% tolerance.

I think there might be a job for you in China somewhere where you can use some blue paint to turn carbon film resistors into metal film, although I doubt they actually bother tweaking them to 1%.

Well that's knocked it on the head for carbon film resistors then. But on that webpage I linked to about filing resistors it also said you can do it to metal film reisistors. In the Art of Electronics it says that with metal film you can get stability of 0.1% under normal conditions. Well, howabout if you file them roughly near  to each others value, then solder them to those switches like the ones you just bought( well if they are any good) and then remeasure them  and trim them with file. And then pot them with thermally conductive potting compound. Wouldn't that be abnormal condition for them and you might get better than 0.1% ?

Using this KVD with the Linear Tech circuit it will only have 10mA (I think?) going through it if that helps with stability.
Well I'm learning a lot about the humble resistor. 

[alm]

Filing removes the lacquer and might make them more susceptible to oxidation, but feel free to try it and find out. It will take a few years to get results about long term stability, though, and you need a stable reference (eg. DMM) that will drift little over that time span.

10mA through 1k is 0.1W dissipation, which for 0.25W resistors will likely result in some heating, since it's fairly close to the max. It would also drop 10V over the resistor, so it's likely a few orders of magnitude off for a buffered voltage source.

The precision wire wound resistors often used for these applications are often physically large compared to their power rating to reduce heating.

[Flavour Flave]

Filing removes the lacquer and might make them more susceptible to oxidation, but feel free to try it and find out. It will take a few years to get results about long term stability, though, and you need a stable reference (eg. DMM) that will drift little over that time span.

10mA through 1k is 0.1W dissipation, which for 0.25W resistors will likely result in some heating, since it's fairly close to the max. It would also drop 10V over the resistor, so it's likely a few orders of magnitude off for a buffered voltage source.

The precision wire wound resistors often used for these applications are often physically large compared to their power rating to reduce heating.

Nail varnish would stop oxidation(maybe?)
Anyway  wirewounds is probably the way to go then. But the precision ones seem very expensive. Well making your own might be an option as Conrad has written.

I just thought that I could just buy one stage at a time. So 0.01% wire wounds and  10ppm/C are the ones to use?

I think there might be a job for you in China somewhere where you can use some blue paint to turn carbon film resistors into metal film, although I doubt they actually bother tweaking them to 1%.

One Hung Low Precision Resistors Inc 

[fmaimon]

Step 2 is you adjust each decade, starting from the second most significant so that "10" is exactly equal to "1" on the range above.

If you are using electronic switches, then to match the ranges, you can switch between the "1" on one range and the 10 on the next lower range at rate of 133Hz (or any frequency that is not a mains harmonic). You will get a square wave out when they do not match, and when you adjust the summing resistor for a match, the amplitude of the squarewave will go to zero.  If you built an sensitive AC amplifier, along with a tuned 133Hz bandpass filter to eliminate everything but the 133Hz Ac, and put your multimeter on the output, then you can match the decades with extreme precision without needing anything expensive or precise.

So calibration comes down to matching resistors within a decade, and zeroing an AC value. These are two steps that can be done simply and very accurately without expensive equipment.

Rereading this, I just can't understand how to calibrate the decades. If you sense the output of the summing amplifier, you will get exaclty 0.2 x the AC voltage, not 0! If you remove the summing amplifier and sense the output of the pots, you will get 0 allways! Am I missing something?

One more thing, the first 2 decades need a transistor (something like Art of Electronics figure 4.21) in it's output to source enough current, as the opamps can't source the 10 mA (first decade) or 1 mA (second decade) needed. As the idea is doing as a module, it is needed in every decade.

And instead of using 2 4051, it may be better to use a single 4067. BTW, wht are you connecting the 1V and 0V taps to the X3, X6 and X7 of the second 4051?

Thank you,
Felipe Maimon

[amspire]

Step 2 is you adjust each decade, starting from the second most significant so that "10" is exactly equal to "1" on the range above.

If you are using electronic switches, then to match the ranges, you can switch between the "1" on one range and the 10 on the next lower range at rate of 133Hz (or any frequency that is not a mains harmonic). You will get a square wave out when they do not match, and when you adjust the summing resistor for a match, the amplitude of the squarewave will go to zero.  If you built an sensitive AC amplifier, along with a tuned 133Hz bandpass filter to eliminate everything but the 133Hz Ac, and put your multimeter on the output, then you can match the decades with extreme precision without needing anything expensive or precise.

So calibration comes down to matching resistors within a decade, and zeroing an AC value. These are two steps that can be done simply and very accurately without expensive equipment.

Rereading this, I just can't understand how to calibrate the decades. If you sense the output of the summing amplifier, you will get exaclty 0.2 x the AC voltage, not 0! If you remove the summing amplifier and sense the output of the pots, you will get 0 allways! Am I missing something?

In this case, we have electronic range switching, and I proposed that lower decades could go to "10".

Now each decade has to fist be calibrated so it is linear to the final spece. All that is left is to calibrate each lower decade to the higher decade.

So you have your precision 10V DC +/- 0.001% as the reference voltage. Switch between these two settings at, say, 133 Hz:

Decade1 = "1" - all other decades = 0
Decade2= "10" - all other decades = 0

When calibrated, both these two voltages will be exactly equal and you will DC out of the summing amplifier. If they are not equal, you will get a 133Hz squarewave out. When you adjust the calibration resistor of the second decade so that the 133Hz squarewave disappears, then the second decade is calibrated. The nice thing is that no precision equipment is needed to do this adjustment, just a simple audio preamp (probably with a low pass filter to remove switching transients) connected to the AC volts on any multimeter.

One more thing, the first 2 decades need a transistor (something like Art of Electronics figure 4.21) in it's output to source enough current, as the opamps can't source the 10 mA (first decade) or 1 mA (second decade) needed. As the idea is doing as a module, it is needed in every decade.

And instead of using 2 4051, it may be better to use a single 4067. BTW, wht are you connecting the 1V and 0V taps to the X3, X6 and X7 of the second 4051?

Thank you,
Felipe Maimon

Of course you choose the first decade summing resistor so that the opamp can handle the current. You can easily get opamps that can do 10mA or more, but if your chosen opamp needs to work below 2mA for best accuracy, you use a summing resistor large enough so that the current is below 2mA. This will make all the lower resistors proportionally larger. Or  as you correctly suggested, you just add a transistor or two to the opamp output on the first decade, and then the 10mA is then easy. Keeps the heat out of the opamp, and that is a good thing.

You can use the 4067 - I was thinking of 2 4051's just because it might make calibrating within a decade slightly easier - I had an idea.  But as I also said, precision voltage sources are all in the details, and it may be the 40XX just have too much leakage for great precision. you have to do the calculations and see if it works or not. Once the numbers are right, you have a design.

Richard.

[fmaimon]

I'm about to buy the resistors for the KVD at digikey and thinking on trying to optimize the cost/benefit ratio, instead of buying 250 10k 1% 50 ppm resistors @ $20, I was thinking on buying 50 10k 0.1% 25 ppm resistors @ $24. They are both 0.5W or better. The problem is that my probability skills are non existant and I don't know the minimun number of 0.1% resistors that I need to get 2 sets of resistors that match better than 40 ppm. Can anyone help me with this?

Or can use the idea of using 3 3k3 resistors, each of 0.25W, but it will probably make my head hurt when matching.

I'll do the 10k-10k-4k-1k-1k-1k KVD as the fluke 720 and the iet kvd-700 as I'll need less values to make it all work. The shunts will have the same specifications of the previous decade and for the 4k and 1K, I'll buy 100 of the 50 ppm.

Thank you,
Felipe Maimon

[amspire]

If you are going to all the trouble of making a KVD, it is worth trying to source some lower temp coefficient parts for the first decade at least. Perhaps the first two decades.

You really notice the drift of 25ppm parts with 10V or more applied. The trouble is if you are using a KVD and you see the voltage drifting a little, it is easy to loose confidence in the accuracy. If you see it sitting rock solid at the set voltage, then you have confidence.  It is really worth investing that extra bit of effort and make something really good.

The negative is that if you buy a small number of 10ppm or better resistors, you definitely cannot select a matching set as well - you will need a trimmer pot on each one.

Now, back to your question on matching,  40ppm is 0.004%.  You need one superb set for the first decade. The second can be a little worse. The third worse again.  So it comes down to how many resistors do you need to get 11 resistors within 0.004%. The chances are good with 100 resistors. With 150, you probably can do it easily.  With 50 0.1%, you may be able to find 11 within 0.01%. 

I wouldn't be overly concerned about using all 10K resistors in the KVD which means you just buy lots of the one value.  When the 720a was designed, they didn't have the pico-amp fet input opamps that are available now to make a output buffer amplifier. So to minimize the output resistance, they used lower values in the lower decades.  They were often relying on very sensitive mechanical galvanometers that can't show lower currents then a few microamps.

You cannot put a 10Mohm multimeter across a KVD output - it draws too much current and will change the output voltage, so to use it, you either need a very high impedance input, or you use the KVD to compare with a second voltage, and you are nulling the current between the two.

Richard

[fmaimon]

If you are going to all the trouble of making a KVD, it is worth trying to source some lower temp coefficient parts for the first decade at least. Perhaps the first two decades.

You really notice the drift of 25ppm parts with 10V or more applied. The trouble is if you are using a KVD and you see the voltage drifting a little, it is easy to loose confidence in the accuracy. If you see it sitting rock solid at the set voltage, then you have confidence.  It is really worth investing that extra bit of effort and make something really good.

That's why I'm thinking on using 25ppm instead of 50ppm. With 10k resistors and 10V, you get about 100uW dissipation in each one and that's about 0.015C temp rise (150 C/W, according to the resistors datasheets). At 50ppm/C it is less than 1 ppm of error if my math is correct. The lower thermal resistance is why I choose 0.5W resistors instead of 0.25W or 0.125W.

This 10ppm resistor looks nice as, according to the datasheet, its thermal resistance is only 110C/W, but it's expensive, as a bag of 250 is nothing less than $165! 
 

The negative is that if you buy a small number of 10ppm or better resistors, you definitely cannot select a matching set as well - you will need a trimmer pot on each one.

Now, back to your question on matching,  40ppm is 0.004%.  You need one superb set for the first decade. The second can be a little worse. The third worse again.  So it comes down to how many resistors do you need to get 11 resistors within 0.004%. The chances are good with 100 resistors. With 150, you probably can do it easily.  With 50 0.1%, you may be able to find 11 within 0.01%. 

These numbers are for 1% resistors, right? I did check Dave's 1% resistor data (400 resistors) and it had 4 sets of 40 ppm. So I assumed (ASS-U-ME?) that 0.1% needed 10 times less resistors to get about the same numbers of sets.

I wouldn't be overly concerned about using all 10K resistors in the KVD which means you just buy lots of the one value.  When the 720a was designed, they didn't have the pico-amp fet input opamps that are available now to make a output buffer amplifier. So to minimize the output resistance, they used lower values in the lower decades.  They were often relying on very sensitive mechanical galvanometers that can't show lower currents then a few microamps.

That's a good idea. I'll just need 10K and 25K (24.3K + 1K pot) resistors.

You cannot put a 10Mohm multimeter across a KVD output - it draws too much current and will change the output voltage, so to use it, you either need a very high impedance input, or you use the KVD to compare with a second voltage, and you are nulling the current between the two.

I won't do that. The output will only go to my Fluke 8840A (>10Gohm) or a LT1050 or similar...

[Conrad Hoffman]

Based on experience with supposedly good expensive wire wound resistors, and with the bags of 200 plain metal films I'd get from Digikey for $9 (a while back), I'm going to suggest that you go through the exercise with the cheap metal films first. You're going to learn things about matching and TC that will serve you well if you try to do it with better parts. You can even compare TC by immersing resistors in warmed mineral oil. One thing I learned was that the resistors in a bag didn't necessarily have the same TC. The other was that soldering was very risky business and I always needed to have a few extra matched resistors to fine tune a decade after it was all assembled. This was true for the wire wounds as well, and I did use heat sinks. Trying to do a 1 ppm first decade is remarkably difficult and there's a lot to be said for trimmers, if you choose wisely. The Fluke uses trimmers, as does the Analogic source. Loebe Julie didn't, nor did ESI, at least on the one I've got. General Radio didn't either, but theirs was only 4 decades, though surprisingly good decades if one measures one in good shape. Of course you could also look for their precision ratio transformer that was good to better than 1 ppm and considered a primary divider by NIST. It was a close relative of the one used in their capacitance bridge.

[HLA-27b]

I was pondering the idea of constructing a KVD on the cheap a while ago. My line of thinking was, instead of purchasing expensive resistors, to buy a spool of low tcr wire like evanohm and wind it around a glass rod in order to obtain a whole decade. Then I could tap that at exactly determined intervals to obtain decade divisions.

The best way to tap leads seemed to be to weld them instead of soldering them. Perhaps by discharging a cap through the junction. But I have never tried this so I don't know if it would work in high precision resistors.

Another point I considered was the winding method. Normal helical wining would make the resistor inductive. There are low inoctance windings like bifilar and Ayrton - Perry windings but these are hard to tap precisely because the exact point may not be exposed. So my idea is to wind the wire around two glass rods simultaneously in a figure of eight fashion which would cancel out inductance and at the same time expose the wire enough for easy tapping.

[amspire]

The negative is that if you buy a small number of 10ppm or better resistors, you definitely cannot select a matching set as well - you will need a trimmer pot on each one.

Now, back to your question on matching,  40ppm is 0.004%.  You need one superb set for the first decade. The second can be a little worse. The third worse again.  So it comes down to how many resistors do you need to get 11 resistors within 0.004%. The chances are good with 100 resistors. With 150, you probably can do it easily.  With 50 0.1%, you may be able to find 11 within 0.01%. 

These numbers are for 1% resistors, right? I did check Dave's 1% resistor data (400 resistors) and it had 4 sets of 40 ppm. So I assumed (ASS-U-ME?) that 0.1% needed 10 times less resistors to get about the same numbers of sets.

No that is for 0.1%. It is likely that the 0.1% resistors will have a wider spread within the 0.1% limits then Dave had with his 1% limits. But it is one of those things that you will not know till you get the resistors. You could find 90% are within 0.01% if you are lucky.

But you are probably right. I was probably being far too conservative.

I wouldn't be overly concerned about using all 10K resistors in the KVD which means you just buy lots of the one value.  When the 720a was designed, they didn't have the pico-amp fet input opamps that are available now to make a output buffer amplifier. So to minimize the output resistance, they used lower values in the lower decades.  They were often relying on very sensitive mechanical galvanometers that can't show lower currents then a few microamps.

That's a good idea. I'll just need 10K and 25K (24.3K + 1K pot) resistors.

A 1K pot is way to big. You will regret it.  If you have resistors that match between decades to at least 0.05%, you don't want the pot doing much more then that. I would have the 24.3K (25ppm) and a 820 ohm (1% 100ppm) in series and put a 10K pot across the 820 ohm resistor. I prefer this circuit as only a fraction of the current is going through the pot wiper.

[amspire]

I was pondering the idea of constructing a KVD on the cheap a while ago. My line of thinking was, instead of purchasing expensive resistors, to buy a spool of low tcr wire like evanohm and wind it around a glass rod in order to obtain a whole decade. Then I could tap that at exactly determined intervals to obtain decade divisions.

The best way to tap leads seemed to be to weld them instead of soldering them. Perhaps by discharging a cap through the junction. But I have never tried this so I don't know if it would work in high precision resistors.

Another point I considered was the winding method. Normal helical wining would make the resistor inductive. There are low inoctance windings like bifilar and Ayrton - Perry windings but these are hard to tap precisely because the exact point may not be exposed. So my idea is to wind the wire around two glass rods simultaneously in a figure of eight fashion which would cancel out inductance and at the same time expose the wire enough for easy tapping.

You could try that, but it is not at all easy.  Lets say you had 10 meters of wire per section - 100 meters overall. to get a match of 0.001%, you have to tap with a 0.1mm accuracy. If you use a finer wire so it is 100 meters per section, then you are talking about 1.1km of wire for the whole network. All that 1.1 km has to be wound perfectly with a minimum stress on the wire (stresses of resistance wire is one of the biggest causes of long term drift in a wirewound resistor). And that is just one divider!

In a KVD, the tap has to carry half the current, so the tap has to be done in a way that will not degrade with corrosion of the wire.

Richard