Some old school instruments showing how it's done (HP 3325A and Fluke 8506a)

[joeqsmith]

Looking at the 1VDC signal from my standard over night (8-9 hours).   You can tell when I enter the room.   

I'll use my 33120A with no warmup to collect VAC next.  200mVPP 100KHz sine wave into a 50ohm terminator.   Not a great setup but should give us some idea how it will behave once things warm up.

[SilverSolder]

Is that using the 10V range, or 1V?

[joeqsmith]

Is that using the 10V range, or 1V?

We can also see from the data I collected, the step size is 1uV.    From the manual, we can see it has a resolution of 1uV in the 1V range.  So five counts of noise.  Eyeballing it, maybe one count of drift over that time.  Seems fairly stable and low noise for an almost a 40 year old meter.

I don't have any way to know absolute accuracy.  I used my old HP34401A as a reference so these two meters are fairly close.  My other old HP bench meter will throw up different numbers.   The voltage standard will have an error as well.   So don't read too much into the absolute numbers based on my home alignment.

[guenthert]

Is that using the 10V range, or 1V?

We can also see from the data I collected, the step size is 1uV.    From the manual, we can see it has a resolution of 1uV in the 1V range.  [..]

     Doesn't the 8505/6 famously have a 7.5 digit 10V range?  That would make 1uV resolution ambiguous.

[SilverSolder]

You can get 1uV on both the 10V and 1V ranges.

It may perform slightly better on the 10V range?  Worth a comparison.  (Need to enable averaging to get to 7.5 digits on 10V)

[bdunham7]

Just checked what 1V RMS from the 3325A looked like on the Fluke 8506a thermal RMS voltmeter.  No real reason to complain today!  (Other than using up my luck for the rest of the year...)

(Attachment Link)

OK, so here are eight digits of coincidental exactness.  And no, neither was calibrated or adjusted using the other.

[bdunham7]

     Doesn't the 8505/6 famously have a 7.5 digit 10V range?  That would make 1uV resolution ambiguous.

It appears to, when averaging is enabled, but as I observe it in action I think that last digit may have missing codes.  IOW, it can't display all possible values.  For example, as it drifts around, I see 10.00000 0, 10.00000 1, 10.000004, but never 10.00000 2 or 10.00003.  I haven't explored this at all, just a quick observation.

[SilverSolder]

Just checked what 1V RMS from the 3325A looked like on the Fluke 8506a thermal RMS voltmeter.  No real reason to complain today!  (Other than using up my luck for the rest of the year...)

(Attachment Link)

OK, so here are eight digits of coincidental exactness.  And no, neither was calibrated or adjusted using the other.

(Attachment Link)

It's not good to use up your luck for the rest of the year already in January! 

[SilverSolder]

     Doesn't the 8505/6 famously have a 7.5 digit 10V range?  That would make 1uV resolution ambiguous.

It appears to, when averaging is enabled, but as I observe it in action I think that last digit may have missing codes.  IOW, it can't display all possible values.  For example, as it drifts around, I see 10.00000 0, 10.00000 1, 10.000004, but never 10.00000 2 or 10.00003.  I haven't explored this at all, just a quick observation.

I get the full compliment of numbers in averaging mode.

It sounds like the A/D converter may need trimming, if it is missing codes (all the steps are adjustable...) 

That is serious project, though.  The manual explains it, but you need a ton of precise equipment and it is still an iterative process...  (I have never attempted it.)

I'm thinking that today, it ought to be possible to do something over GPIP to automate measuring the A/D performance, but haven't actually tried anything...


Regarding stability in general, I once did a test where 4 of these meters were left running overnight reading an EDC VS-330 voltage reference (which has been "massaged" with some better components here and there to improve stability) set to 10V.

Over the 8h period, the four meters and the VS-330 tracked each other as follows: 
DMM1 +/- 1.2ppm
DMM2 +/- 0.5ppm
DMM3 +/- 0.6ppm
DMM4 +/- 0.85ppm

This was at night, during the summer, where the temperature is very stable overnight in this location.

I just used the Peak function on the meters for this, so I don't have any logs.

Bottom line, this old school stuff is all "good enough for Australia" when in good shape and given a quiet place to work!

[bdunham7]

I get the full compliment of numbers in averaging mode.
It sounds like the A/D converter may need trimming, if it is missing codes (all the steps are adjustable...) 

I actually did the A/D calibration procedure, although I went through it pretty quickly, not aiming for perfection at the time.  I don't have references that good anyway.  It took about an hour and I recall doing the first loop thrice, then on the rest of the procedure I just rechecked to make sure I was pretty close.  It seems pretty good throughout the ranges, but I'll keep an eye on it and see if those 'missing codes' really are missing.

That is serious project, though.  The manual explains it, but you need a ton of precise equipment and it is still an iterative process...  (I have never attempted it.)

It's just fiddly and takes time.  And if you do have accurate references, you just have to keep doing it until you just aren't improving anything anymore.  I've done similar adjustments on the 5100-series calibrators and they never seem quite happy.  This one seemed pretty amenable to adjustment.

Over the 8h period, the four meters and the VS-330 tracked each other as follows: 
DMM1 +/- 1.2ppm
DMM2 +/- 0.5ppm
DMM3 +/- 0.6ppm
DMM4 +/- 0.85ppm

This was at night, during the summer, where the temperature is very stable overnight in this location.

In my case, unfortunately, the room it is in is not stable because the furnace cycles and I'm getting variations of 5K on a daily basis.  The 731B and the 8505A both seem to have a negative tempco, and between them they move about 5ppm total over 24 hours.  That's still about as much bang-for-fifty-bucks-plus-some-work as one could ask for.

[joeqsmith]

Sadly we lost power today long enough that the UPS cut out.  The Arb was brought up cold but still it appears it was going to be fairly drift free.   The 2mV of error could be the Arb, meter, terminator, all of the above.   Close enough.   

You can get 1uV on both the 10V and 1V ranges.

It may perform slightly better on the 10V range?  Worth a comparison.  (Need to enable averaging to get to 7.5 digits on 10V)

After I noticed my manual was incorrect, I have not tried to enable the average again.   I just tried it out and it seems to work.   

[joeqsmith]

My HP meter in red, the Fluke in white.  Both meters have been on for over 24 hours now.   The Fluke is set to average.   The top cover has yet to be installed.   Using my old 731B for a source. 

The software currently captures the data once every 5 seconds, so you are looking at about 1 hour of data.  Again, the Fluke was aligned using the HP.  The voltage standard could very well be tighter.   Ignore the absolute readings.

[SilverSolder]

Interesting, are you seeing the "missing numbers" issue that @bdunham7 was talking about?
My GPIB setup is still down - when I get it back up and running, I'll try some logging to look for this issue specifically. 

As @bdunham7 said, it is hard to beat the metrology bang for the buck with these things.  But they can be hard to maintain if there are subtle problems unless you have another meter that is equally precise that you can compare it with.  One thing leads to another... !

The spec is nearly 2x wider on the 1V scale compared to 10V - probably due to the input attenuator?

[guenthert]

     Even if the ADC would be unable to report some codes, the averaging surely will.  I can only assume that what @bdunham7 sees is coincidence.

[joeqsmith]

Shown is about 4.5 hours.   Increasing the histogram's bin size to look at the missing codes,  the meter has not put out any 1s, 2s, 6s and 7's in the extended code.   We see something similar with the HP not showing 3s & 8's.

To test it out, I changed the voltage standard to delta E (uV resolution) and slowly adjusted the output voltage.  No missing codes.

[bdunham7]

     Even if the ADC would be unable to report some codes, the averaging surely will.  I can only assume that what @bdunham7 sees is coincidence.

Without knowing exactly how the whole thing works, we can't be sure.  In any case, I shifted back to monitoring a 10V reference this morning and temperatures have been more stable, so I'm seeing numbers from 9.99999 6 to 10.00000 4, with the vast majority being 9.99999 9, 10.00000 0 and 10.00000 1.  I haven't been staring at it the whole morning, but check it out whenever I come into the room and I have yet to see a 9.99999 7, 10.00000 2 or 10.00000 3. 

It is almost scary to see it appear to be accurate and stable to 100ppb.  I moved the reference from on top of the meter to beside it and that seems to have stabilized it greatly.

[SilverSolder]

Shown is about 4.5 hours.   Increasing the histogram's bin size to look at the missing codes,  the meter has not put out any 1s, 2s, 6s and 7's in the extended code.   We see something similar with the HP not showing 3s & 8's.

To test it out, I changed the voltage standard to delta E (uV resolution) and slowly adjusted the output voltage.  No missing codes.

That is a little bizarre...   Some kind of rounding or precision issue with the calculation of the running average?

Floating point operations are not perfect, of course. 
https://docs.oracle.com/cd/E19957-01/806-3568/ncg_goldberg.html

[SilverSolder]

     Even if the ADC would be unable to report some codes, the averaging surely will.  I can only assume that what @bdunham7 sees is coincidence.

Without knowing exactly how the whole thing works, we can't be sure.  In any case, I shifted back to monitoring a 10V reference this morning and temperatures have been more stable, so I'm seeing numbers from 9.99999 6 to 10.00000 4, with the vast majority being 9.99999 9, 10.00000 0 and 10.00000 1.  I haven't been staring at it the whole morning, but check it out whenever I come into the room and I have yet to see a 9.99999 7, 10.00000 2 or 10.00000 3. 

It is almost scary to see it appear to be accurate and stable to 100ppb.  I moved the reference from on top of the meter to beside it and that seems to have stabilized it greatly.

Yes, it is amazing when two independent entities get that close, and stay that close.  Like two icebergs floating a cigarette paper's thickness apart, yet never touching!

[bdunham7]

That is a little bizarre...   Some kind of rounding or precision issue with the calculation of the running average?

Floating point operations are not perfect, of course. 
https://docs.oracle.com/cd/E19957-01/806-3568/ncg_goldberg.html

I suppose the next thing is for someone else with one of these to test and see if they can get it to read exactly any of the voltages that I'm seeing as 'missing'.  I was unable to replicate it at a few other much lower voltages where I have 1uV adjustment capability. 

B/t/w, thanks for the post inspiring me to dig this thing out and finish the project.  I already completely dissembled it twice and replaced some power supply caps, two fuses and two resistors on the F/R input board.  Apparently someone used the rear connector to overload the Xref circuit--which I haven't tested yet.  Does anyone have a line on that strange connector they used on the back side?  Its an Amphenol low mating force, probably mil-spec, but I can't seem to find a part number or a mating connector.  Now I finally did a basic cal and tested it, all it needed beyond the DC cal was an adjustment to the AC zero and it all looks pretty good, including the current ranges.  Ohms may be a tad off, but now I'll have to buy or make a batch of resistance standards....

[SilverSolder]

I suppose the next thing is for someone else with one of these to test and see if they can get it to read exactly any of the voltages that I'm seeing as 'missing'.

I'll have a play with this, once I get GPIB back up and running after my re-arrangement exercise.

B/t/w, thanks for the post inspiring me to dig this thing out and finish the project.  I already completely dissembled it twice and replaced some power supply caps, two fuses and two resistors on the F/R input board.  Apparently someone used the rear connector to overload the Xref circuit--which I haven't tested yet. 

It can get a little lonely playing with these, they are not super popular.

When you are looking at PPMs, drifty components that wouldn't be noticed in any other circuit suddenly begin to matter.  So they are like owning a classic car, where you have to spend some time taking care of it to get the best out of it - but when it is on song - what a ride!

Does anyone have a line on that strange connector they used on the back side?  Its an Amphenol low mating force, probably mil-spec, but I can't seem to find a part number or a mating connector.  [...]

The rear connector is an ELCO 8016- 20 Pin socket - you can find both sockets and connectors on eBay.....

[joeqsmith]

Shown is about 4.5 hours.   Increasing the histogram's bin size to look at the missing codes,  the meter has not put out any 1s, 2s, 6s and 7's in the extended code.   We see something similar with the HP not showing 3s & 8's.

To test it out, I changed the voltage standard to delta E (uV resolution) and slowly adjusted the output voltage.  No missing codes.

That is a little bizarre...   Some kind of rounding or precision issue with the calculation of the running average?

Floating point operations are not perfect, of course. 
https://docs.oracle.com/cd/E19957-01/806-3568/ncg_goldberg.html

Again it does show the least significant digit hitting all the buckets.  But I guess the question is does it do it at 10V and there are specific codes that are in question.  Of course just eyeballing things isn't going to get you too far. 

Voltage source set at 10V.   The easiest thing to do would be to add some error by adjusting the trimmers but I would rather not touch it again as the unit is close now and buttoned up.   Instead, because my meter reads the standard a little hot, I just used a decade box to shunt it small amounts.  The histogram is plotting the least significant digit.  Note that the 2,3 and 7 buckets are lower but they certainly do occur  with my meter.   

... I have yet to see a 9.99999 7, 10.00000 2 or 10.00000 3. 

Shown is with the LSD set to 2. 

[bdunham7]

The rear connector is an ELCO 8016- 20 Pin socket - you can find both sockets and connectors on eBay.....

You made my day!  I was absolutely convinced it was Amphenol for some reason.  That cross-pin or tuning fork design is pretty interesting--the male and female ends using the same pins. Now I'll be able to make a 4W Ohms harness and test the X-REF circuits to make sure they're not blown up by whatever burned the 39K resistors on the input board.

[bdunham7]

Again it does show the least significant digit hitting all the buckets.  But I guess the question is does it do it at 10V and there are specific codes that are in question.  Of course just eyeballing things isn't going to get you too far. 

Voltage source set at 10V.   The easiest thing to do would be to add some error by adjusting the trimmers but I would rather not touch it again as the unit is close now and buttoned up.   Instead, because my meter reads the standard a little hot, I just used a decade box to shunt it small amounts.  The histogram is plotting the least significant digit.  Note that the 2,3 and 7 buckets are lower but they certainly do occur  with my meter.   

Shown is with the LSD set to 2.

Assuming the 8506A and 8505A use the same software, it looks like mine is just sloppy A/D converter linearity adjustment.  Interesting that your counts are lower, but not zero (I've eyeballed more than 250 samples by now...).  I suppose that the distribution of LSDs could be used as a test of linearity, or at least the accuracy of the alignment of the ladders which should correlate to linearity.

[SilverSolder]

The rear connector is an ELCO 8016- 20 Pin socket - you can find both sockets and connectors on eBay.....

You made my day!  I was absolutely convinced it was Amphenol for some reason.  That cross-pin or tuning fork design is pretty interesting--the male and female ends using the same pins. Now I'll be able to make a 4W Ohms harness and test the X-REF circuits to make sure they're not blown up by whatever burned the 39K resistors on the input board.

It took me something like 2 years to figure out what that connector is...   it is one of the best kept secrets on the Web! 

[SilverSolder]

[...] I suppose that the distribution of LSDs could be used as a test of linearity, or at least the accuracy of the alignment of the ladders which should correlate to linearity.

That's the kind of approach I was thinking of, perhaps tuning the A/D converter by feeding it a triangle wave sweep with added white noise, and looking for patterns in the output that shouldn't be there?

That is something for a dark winter's night for sure...

Absolutely have to have GPIB (or serial) working for anything like that!