[3458A] Yet another 3458 repair thread.

[TiN]

Yea, probably best to go in something between A and B.
I have second A1 board to play with experiments, which have few holes in it, but nothing unrepairable.

[Grandchuck]

Plan D:  Keep it around for the future.  You could find yet another 3458A (with a good A3) at a very good price.  There is always room for one more.

In any case, just love your contributions ... have learned so much from your efforts.  Thanks and keep up the good work!

[TiN]

Guess what, I overdone on plan D...



I don't know where the 4th 3458A popped up from...



So far issues with unit 4:

* Stolen fan.
* RAM 1 LOW problem. Dallases are already ordered
* A2 board gives DAC FLATNESS CONVERGENCE ERROR: 198.
* Busted HI SENSE post on front terminal block
* Ugly and filthy outside and dusty inside.

Elantec on A2 ACV board was butchered up by somebody. Will need to get back to that later, for now I installed A5 and A2 board from 3458C unit.
It pass self-test and ACAL ALL with these donor boards.

That hand writing on A3 ADC U180 hybrid -1.30 suggests drift rate, I assume? 

Morally preparing myself for A3 replacement.

[texaspyro]

Oh dear,  I'm soon gonna need another 3458A...  I can't let TIN have more 3458A's than me (all 4 of mine work,  no board replacements needed).   

[acts238willy]

So...... TiN is the guy who's hogging up all of
those fleabay 3458s???

Glad I didn't pull the trigger on that first one.

Saw the missing fan and the A5 rusty board with no attach bolts
on that 2nd one...

[TiN]

So...... TiN is the guy who's hogging up all of
those fleabay 3458s???

Glad I didn't pull the trigger on that first one.

Saw the missing fan and the A5 rusty board with no attach bolts
on that 2nd one...

Far from it, some greedy guys have more...
that last 3458D with missing fan actually turned out very good so far. Data show stable ADC so far. A5 and fan very easy fix. A2 have common issue, flatness DAC 198, so I installed A2 from 3458C unit into it first.

[TiN]

Unit 4 fixed, issue with A2's Flatness DAC convergence: 198 was dodgy EL2039 wideband opamp. Replaced one with NOS chip and she worked as expected. 

Doctor, doctor.... we have severe GAS outbreak... 

Unit "3" is my reference, unit "2" is backup, top and bottom are units in this thread.
Fluke meter last calibrated in 2014. Calibrator ran thru ACAL 2 days ago versus my 10V LTZ1000 792X box and calibrated SL935 10Kohm/1ohm resistance standard.



Reason why I piled them all up like that - INL testing.

Data so far (to be updated)



Vertical span is +0.1/-0.1ppm. Typical 3458A per documentation is 0.1ppm, so good meter data should be well within the bouds. 

[Dr. Frank]

Unit 4 fixed, issue with A2's Flatness DAC convergence: 198 was dodgy EL2039 wideband opamp. Replaced one with NOS chip and she worked as expected. 

Doctor, doctor.... we have severe GAS outbreak... 

Unit "3" is my reference, unit "2" is backup, top and bottom are units in this thread.
Fluke meter last calibrated in 2014. Calibrator ran thru ACAL 2 days ago versus my 10V LTZ1000 792X box and calibrated SL935 10Kohm/1ohm resistance standard.
..
Reason why I piled them all up like that - INL testing.

Data so far (to be updated)
..

Vertical span is +0.1/-0.1ppm. Typical 3458A per documentation is 0.1ppm, so good meter data should be well within the bouds. 

Illya,
these are impressing measurements, showing that 3458As probably have a much more linear ADC than the 8508A.

I'd like to point out, that the linearity parameters (INL or DNL) are NOT specified for the 8508A at all, and for the 3458A, it's only a combined parameter 'Transfer Accuracy / Linearity'.

That's a subtle difference only, but Transfer is defined / measured differently, than Linearity.
In Transfer, you'd measure two different voltages, e.g. F.s: 10V vs. 1V, 2V, ... and then specify the ratio error..
Instead, the INL of the 3458A is typically 0.02ppm, and the DNL 0.01ppm (see hpj 4/89, p. 23)

If you would apply the error calculation on the ACAL process, using Transfer Accuracy of 0.05 ppm / reading + 0.05ppm / range, you would end up with a higher total 10:1 transfer error, i.e. 0.55ppm instead of 0.33ppm.

That distinction between Transfer Accuracy, and pure INL/DNL is important, I think, because the ACAL of the 3458A really lives from that superior ADC.

Frank

[e61_phil]

Both meters are specified with a table which shows the transfer specification.

These tables are not comparable 1:1 because Keysight describes another rule to combine the measurement.

For example a 1V to 10V transfer with the 3458A is described in a Keysight paper:

(0.05ppm + 0.05ppm * 1V/10V) + (0.05ppm + 0.05ppm * 10V/10V) = 0.65ppm


Fluke always uses the sum of squares:

sqrt ( (0.12ppm + 0.1ppm * 1V/20V)^2 + (0.12ppm + 0.1ppm * 10V/20V)^2 ) = 2.14ppm


In this case the 3458A is clearly the winner.

Edit:
I thought the 8508A will beat the 3458A in other ranges, but that is only the case for the 1kV range. (for 1kV HP allowed only 5% voltage variation for the transfer)

10mV -> 100mV
3458A :  6.5
8508A :  6.47
--------------------------
100mV -> 1V
3458A :  1.70
8508A :  2.14
--------------------------
1V -> 10V
3458A :  0.65
8508A :  2.14
--------------------------
10V -> 100V
3458A :  2.1
8508A :  2.47
--------------------------
100V -> 1000V
3458A :  (3.55)
8508A :  6.47
--------------------------

[Echo88]

Mentioned paper: https://www.keysight.com/upload/cmc_upload/All/Calculating_uncertainty_of_a_Ratio_measurement.pdf
Very interesting if one does voltage transfers.

[e61_phil]

Mentioned paper: https://www.keysight.com/upload/cmc_upload/All/Calculating_uncertainty_of_a_Ratio_measurement.pdf
Very interesting if one does voltage transfers.

Exactly that paper I had in mind. But I thought they always used the linear combination.

[TiN]

One has to be careful when comparing performance specifications and read all 5pt font notes. Obviously manufacturer often may want to show "best case" result or "typical" data, leaving up to application engineer to figure out his use conditions/specifications to apply. But 3458A documentation does mention INL 0.1ppm, while Fluke let you guess from transfer specs. In my calibration verification procedures, I follow HP's way and test against these strict levels for FAIL/PASS. I have seen that good working 3458A in shape able to deliver better than 24hour specs across whole DCV function using this worse case scenario, so for my own testing i see no benefit going the typical(?) metrology squared root uncertainty calculation.

Another thing about the above mentioned 1kV transfers. Heating/divider PCR error IS included in 3458A specification already. If you want best accuracy, you can indeed wait till readings settle out and then get better result for this range. Film Vishay resistor network for this is not too fancy, so obviously there was no goal for best perf in this range by design. From my brief testing, it takes about 15-20 minutes till 1kVDC signal settles to flat samples with sdev around 1ppm or better. And since 1kV range ACALed with 10V only, you don't have this parasitic effect during the transfer. I got my 752A now, so I will be able to investigate into this more during 1kV INL tests.

---

I'm still massaging some data sweeps, but I'm not sure you want to see more of INL chatter in this repair thread or not..

Btw, you can see the setup live if watching stack of instruments give you fuzzy feelings...

Here's complete sweep. NPLC50, DCV10, soak time 2 second per 0.1v step, single ACAL DCV before start of sweep. Sweep is from -10.9 to +10.9VDC, take 5 readings, store median of last 3.

Graph with direct poly fit INL versus the source (Fluke 5720A "Hulk", calibrated 2 days ago). Source considered as ideal here.



Meters A,B,C agree to each other. Can we say with good confidence that this curve dominated by calibrator's own INL here? Questions, questions.
Now question about 3458 meter D. It's INL data is horrible, aye? Well, here the confirmation of the theory - that bad drifty A3 also gives not only drift (2ppm/day in this case) but also bad INL performance. This would be naturally expected as reference voltages are drifting wrong ways and total ratios for charge in ADC slopes would be not so good either?
I'm interested to know, as INL sweep can be faster to do then waiting for a week logging proven stable reference to find out dodgy A3 board in meter..
Fluke shows somewhat wierd data, with very good INL on positive polarity, but somewhat ugly on negative?

Now lets see on very same data, but from different angle:



Here meter "A" used as the reference, assuming it's INL is magical 0.00 ppm. So we trade 5720A's INL error to 3458A INL error, and can see direct difference between meter "A" and rest of the pack. Please note different vertical scale compared to previous graph, +/-0.2ppm now (previous was +/-0.35ppm).

Because three good 3458 meters agree on INL points very closely, what if we do next math acrobatics:
"Reference point" = ( (INL of 3458 A) + (INL of 3458 B) + (INL of 3458 C) ) divide by 3.
Now this "reference across the hospital" used as zero and each meter compared to it, showing deviation from the averaged "corrected out" 5720 INL.
This idea implemented on next graph:



This data finally delivers long desired "typical" +0.05/-0.05 ppm across the whole sweep. If this is real or not, we will find out only next year, when my experiment extend to the next level. 

Perhaps same conclusion done by Jim Williams during his design for AN86 20-bit DAC, as we also see three 3458A's on the bench. It should be clear why need to use minimum three meters, better more to better remove errors of the source itself.
Please note different vertical scale compared to previous graph, +/-0.3ppm, increased to fit Fluke 8508 INL.

-------------

Also while at it, I was thinking how we can show "noise" or variance of the INL points? Previous graphs show nice and baby smooth curves, but they are not precisely true. Those datasets also hide some information from you , as each 0.1V step is actual median from 3-10 samples (depends on sweep config) so it's not a single reading from meter. Idea was to reduce short-term noise from the meter's ADC/REF, etc. But what if I want more real-life data, I cannot remove the inherent noise.

So instead of doing medians on step samples, I've ran sweep and stored every single sample into DSV file. Then same math plotting was applied with poly fit, but now twice. One is filtered out gaussian value of the step, show on graph with bright line. And peak-peak each sample data is now also plotted in opaque thick line. This gives some represenation of "peak-peak" window on each step point.

Current example, taken by _another_ 5720A and two old 3458A's in far far away country over remote Raspberry Pi log:



Red is differential INL between meters. And don't ask me what happen with meter "B" after +10V, I don't know 

What really amazes me - that we can even distinguish this kind of numbers at fractions of 0.1 ppm, in our crude homelab setups (even with more reasonable 5440B/4808 calibrators, no need 5720's for this data testing). Just to think about it, own LTZ1000A reference noise that is installed in the meter produce noise around 0.2ppm all by itself, but here we looking at tiny signal deviations 5 times less than this noise! I did not really expect it.

[e61_phil]

One has to be careful when comparing performance specifications and read all 5pt font notes. Obviously manufacturer often may want to show "best case" result or "typical" data, leaving up to application engineer to figure out his use conditions/specifications to apply. But 3458A documentation does mention INL 0.1ppm, while Fluke let you guess from transfer specs. In my calibration verification procedures, I follow HP's way and test against these strict levels for FAIL/PASS. I have seen that good working 3458A in shape able to deliver better than 24hour specs across whole DCV function using this worse case scenario, so for my own testing i see no benefit going the typical(?) metrology squared root uncertainty calculation.

Another thing about the above mentioned 1kV transfers. Heating/divider PCR error IS included in 3458A specification already. If you want best accuracy, you can indeed wait till readings settle out and then get better result for this range. Film Vishay resistor network for this is not too fancy, so obviously there was no goal for best perf in this range by design. From my brief testing, it takes about 15-20 minutes till 1kVDC signal settles to flat samples with sdev around 1ppm or better. And since 1kV range ACALed with 10V only, you don't have this parasitic effect during the transfer. I got my 752A now, so I will be able to investigate into this more during 1kV INL tests.

At first let us clarify: The 3458A is more linear than the 8508A! I don't want to argue about that. I saw similiar agreement of INL in our company between the 3458As there.

BUT , I don't expect in any Fluke spec that it will reflect only typical values or best case data. I'm pretty sure all values in their datasheets (99% values) are really written with a lot of confidence.
I read through the Fluke 5730A manual the last days and they say all specs are written including everything over the whole temperature range and the variety of different units. Normally one could expect half of the uncertainty. But that has to be proven. I think that is reliable if you look at the calibration sheet from 8508As which are calibrated with charaterized 5720As.
I don't want to say that HP specs are worse or something like this.

And what is the point of INL specification in the end? What really matters is the transfer you could do and that is specified (on this topic the 3458A specs lacks of some transfer specs for ohms and so on).

I'm not a friend of talking about typical specs. No one knows if this very unit on my desk behaves typical. And in case of the ultra low INL of a 3458A it is nearly impossible to verify, especially if you only have one 3458A.

RSS is not special for metrology it applies to everything which has a gausian distribution and is independend from each other.

The 3458A transfer specs doesn't include self heating. The transfer spec is limited to 5% variation. You cannot apply this spec to transfer 100V to 1kV.


I really enjoy your measurements. And I'm very curious about the 752A measurements.


Btw: I don't want to attack any of your great work. If it sounds sometimes like that, it is only due to my bad english

[Kleinstein]

The INL error from the bad ADC board looks interesting.  The jump at around 0 is kind of surprising though. Having so much trouble near zero could be really bad, as the range near zero is heavily used for things like auto zero.
The shown INL error should also show up with simple source reversal. The more square law part is still not easy to explain from wrong resistor ratios.

There might be a different test that would be more sensitive to wrong resistor ratios:  With a shorter integration time (e.g. 1 ms) the effect of the minor slopes during rundown would be much larger and might show up in the smaller range more DNL like error. So the test would be using the 10 V range but only do a small range scan of maybe some 0-50 mV (so use divider after PWM DAC) with averaging many (some 1000) samples at 1 ms.   An error might also show up as high noise at some readings.

For comparing to the Fluke 8508, are that data taken at the same time: AFAIK the Fluke is quite a bit slower so that this would cause a longer average waiting time if a fixed number of 3 readings is used.

Why is the median values used ? This is normally used of there are some possible outliers so that a robust statistics is needed that can better cope with these. The more normal noise case would favor the normal mean value.

The likely seen INL error from the 5720 looks like the kind of "expected" (the two sides are joined together) jump at around zero and a slightly surprising different gain for the positive an negative side. Due to the way the measurement is done I would more expect this kind of error from the 5720 than the 3458.

The large error for one of the remote 3458 indicates a kind of slight problem with that meter. This might be something like clamping starting a little too early or maybe a problem with settling, e.g. from AZ zero switching.

[e61_phil]

Is it possible to measure the INL of the bad 3458A the other way around? If it is drift related it should change the INL picture, shouldn't it?

[TiN]

e61_phil

I always welcome a discussion on these complex topics, not pretending to know everything, still learning here.

I don't expect in any Fluke spec that it will reflect only typical values or best case data.

Not argue with that. But it looks to me that "not so good" specifications like INL are conveniently "omitted" from the 8508A documentation, as statement like <1ppm it would not look good against 2 times cheaper and 10 year older competitor unit, that states 0.1ppm max.

And what is the point of INL specification in the end?

Great INL is what makes 8.5-digit meter a meter, not an indicator. If you trying to do compensation for tempco/power coefficients and similar analysis, without knowing actual INL how one can get accurate results and be confident that deviations are not come from inherent ADC error contributions. This can be handy to investigate front-end performance, because there are many ranges in meter, and error characteristics on them are different.

Also you can get idea of transfer spec from HP's 3458HFL spec sheet which is just selected 3458A (unless I'm proven wrong with actual HFL unit guts photos):



Now back to the 1kV range. 3458A calibration manual clearly states:

Normal 100 V and 1000 V range measurements use a 100:1 resistor network to attenuate the input. To correct errors introduced by this network, we apply zero volts to the input. Then, we apply 10 V and measure the actual value. Finally, we measure 0.1 V, with the zero error removed, and compute the gain adjustment constant.

Input voltages greater than 100 V (1000 V range) create a self-heating error in the resistor network, as shown in Figure 3. This easily identified error is simply specified as part of the instrument's published error.

Based on this I don't know where you get idea that self-heating is not included in the published specifications for 1kV range?

-------

Kleinstein

Interesting ideas. You mean to run little sweep -50..50mV on 10V range with external divider? I have Keithley 262 that can use 1:100 ratios for that, and program Hulk to output -5/+5V sweep. Also another thing to look into is 3458's auto zero function, as all previous sweeps are with AZERO ON, but it'd be interesting to see also data w/o AZER.

For comparing to the Fluke 8508, are that data taken at the same time: AFAIK the Fluke is quite a bit slower so that this would cause a longer average waiting time if a fixed number of 3 readings is used.

8508 is at least 10 times slower than 3458, however all five meters are triggered at same time (with tiny overhead from RPi). Then we read back data as it's ready from the meters.

Code: [Select]

fluke.write("OUT %.7f" % set_value)
    time.sleep(3)                   # Soak little time for settle
    for ix in range (0,10):         # Take samples cycle
        dmm.write("TARM SGL")       # Trigger 3458A
        dmm2.write("TARM SGL")      # Trigger 3458B
        dmm3.write("TARM SGL")      # Trigger 3458C
        dmm4.write("TARM SGL")      # Trigger 3458D
        dmm5.write("*TRG;GET;RDG?") # Trigger 8508D
        val = float(dmm.read())     # Read data from meter
        val2 = float(dmm2.read())   # Read data from meter
        val3 = float(dmm3.read())   # Read data from meter
        val4 = float(dmm4.read())   # Read data from meter
        val5 = float(dmm5.read())   # Read data from meter

Next trigger for next sample will not be issued until all meters provided the data. But I don't see how it's important in this case, as even when triggers and read back is sequential it would be visible on the INL plot by small temporal shift, and not the INL error data change. We talking about tiny fraction error over 2-3 minute instability of the used source.

57xx do have separate gain constants for different polarity quadrants. Output from DAC reversed in polarity at buffer output by a relay before it goes further in calibrator signal path, so separate corrections are used depends on signal polarity. 5440B have relay reversal at reference before the DAC, so the crossover error is visibly larger on that unit. Errors as reported by "Hulk", acquired by comparison with it's internal 6V/13V dual LTFLU VREF and zero short:

Code: [Select]

------------------------------------------------------------------------
                        DC VOLTAGE OUTPUT SHIFTS
------------------------------------------------------------------------
RANGE Point     Zero Shift       Full Scale Shift       Spec (+/-)  Shift (% 24hr spec)
220 mV +FS    -0.00001 mV     0.00001 mV   0.03   ppm   7.27 ppm    0.38
       -FS     0.00001 mV    -0.00001 mV  -0.06   ppm   7.27 ppm   -0.77
  2.2V +FS   0.0000002  V   0.0000002  V   0.08   ppm   3.86 ppm    2.13
       -FS   0.0000004  V   0.0000004  V   0.17   ppm   3.86 ppm    4.41
   11V +FS   -0.000001  V   -0.000001  V  -0.06   ppm   2.77 ppm   -2.09
       -FS    0.000000  V    0.000000  V   0.03   ppm   2.77 ppm    1.08
   22V +FS   -0.000001  V   -0.000001  V  -0.05   ppm   2.73 ppm   -1.95
       -FS    0.000001  V    0.000001  V   0.03   ppm   2.73 ppm    1.00
  220V +FS    -0.00001  V     0.00018  V   0.84   ppm   3.73 ppm   22.44
       -FS     0.00001  V    -0.00018  V  -0.81   ppm   3.73 ppm  -21.69
 1100V +FS     -0.0000  V      0.0019  V   1.71   ppm   5.45 ppm   31.37
       -FS      0.0001  V     -0.0019  V  -1.68   ppm   5.45 ppm  -30.86
       
I don't think it's a big deal, because we use 3458A data as reference or corrected outsource data (from triple 3458A idea, shown on my post above). By this method such zero error from source is removed.

Last night was also testing some python code, and ran quick NPLC5 sweep while at it. Outcome quite interesting:



Now we remove noisy misused 8508A with it's RESL6 data, and just focus on 3458A in this fast mode. The effect on INL going from NPLC50 (previous post) to NPLC5 is actually not that big, as one may expect.

This is another aspect that I do not enjoy about 8508A. As far I tried, there is no way to configure 8508A for max resolution with custom NPLC conversion time. Best accuracy/lowest noise configuration is RESL8, FAST_OFF which translates to about NPLC1024 judging from spec conversion time. I understand reasons why Datron/Fluke limited this and removed ability of user to set NPLC, but it's rather inconvenient if we trying to work out the unit’s ADC performance.



I did some sweeps on other end of the spectrum, using NPLC1000 on two 3458s, but actually result was rather bad compared to faster NPLC5...NPLC50 sweeps. I'd expect that because ambient temperature/short term drifts variations and such. When we go long conversion times/lot of samples we are measuring less of ADC's INL but more of stability across time + INL.

The large error for one of the remote 3458 indicates a kind of slight problem with that meter.

Yea, those meters known to throw some RAM/ROM checksum errors sometimes. NVRAMs are on last breath in those, so perhaps some little bit got flipped.

Is it possible to measure the INL of the bad 3458A the other way around?

In reverse polarity? I doubt that would make any point, as sweep is covering both polarity already.

--------

Another big point is ACAL. Phil expressed his concerns about it before, even saying it is a "drawback" of the meter, however I respectfully disagree on that. ACAL is just another tool that engineer can use to improve the confidence and achieve better data results between the traceable metrology-proven calibrations. ACAL does not replace the need of the calibration with known uncertainty equipment, but rather makes possible to do educated accuracy improvements at time of the measurement, different from calibration day.

As side effect this also means lack of manual range adjustments in 3458A, and deleting those range adjustment corrections during ACAL procedure in 57xx MFC. The point is that 3458A internally does very precise transfers, which are quite expensive to do externally (as you can learn from 8508A calibration procedures/certificate). And if you replace those transfers with manual adjustments, it's rather an open question if typical lab can meet tight 3458A specifications and account for all external errors of doing such a transfer set. After all owner of 3458A want to use it to measure the DUT, not fiddle with meter for days with manual calibration before you can use it.

For volt-nuts ACAL is pure gold, which also enables people like me to test DUTs versus 24-hour specifications of the 3458A, instead of relying on annual spec. Sure if volt-nut lives and works at NMI, he could in theory shorten the external calibration cycle to maybe 30 days and rotate meter between references (here we talk about accurate known sources for all functions/ranges, not silly 10V/10Kohm ACAL only, aye?) to obtain real traceable 30-day specifications. But for mortals like me, that is not viable solution at all and tracking meter with ACAL versus annual and traceable 10V/10k/1ohm standards calibration makes use of all other ranges possible too, not just 10V/10k. Here is just use ACAL as a tool to confirm meter's drift and confidence in results on _all_ ranges between the traceable external calibrations.

It does not mean however that we cannot also get better results from fancier parts 3458A. Using better components like VHP resistors in shunt/ohm current section might be helpful, but that's topic for yet another article. All I want to say here - I don't buy that marketing push of "8508A is filled with VHP resistors and selected parts, hence it's more stable and does not need ACAL to meet its spec". 3458A don't need ACAL to meet specs either, on that part, but artifact calibration is helpful and integral part of the meter operation and meter was designed around it, it is not just few extra relays to make required switching.

-------

Ok, enough rambling for now, time to hunt for some 3458A parts to get these "C" and "D" units facelifted and patched up.

[e61_phil]

starting painfull quote in this forum...

Great INL is what makes 8.5-digit meter a meter, not an indicator. If you trying to do compensation for tempco/power coefficients and similar analysis, without knowing actual INL how one can get accurate results and be confident that deviations are not come from inherent ADC error contributions. This can be handy to investigate front-end performance, because there are many ranges in meter, and error characteristics on them are different.

In my opinion you always have to consider for all sources of error in such measurements. And this will be reflected by the transfer specs. But for good transfer specs you will need a good INL of course. Therefore, this goes into hairsplitting. I think Fluke have a very good idea of the guaranteed INL of their meters. I would expect they need that to get reliable transfer specs. The fact that they don't show it in the datasheet might mean something.

Here is a INL measurement from Fluke included: http://support.fluke.com/Calibration-Sales/download/asset/2114953_a_w.pdf (Figure 8 )

Also you can get idea of transfer spec from HP's 3458HFL spec sheet which is just selected 3458A (unless I'm proven wrong with actual HFL unit guts photos):

For me that doesn't mean much. I could also look onto your results. For my 3458A that means nothing. The option 2 reference is also "just selected". For traceable data you need some specification. With a standard 3458A there is no other way than using 24h specs for resistance transfer I would expect. Nothing else is guaranteed.

Now back to the 1kV range. 3458A calibration manual clearly states:

Normal 100 V and 1000 V range measurements use a 100:1 resistor network to attenuate the input. To correct errors introduced by this network, we apply zero volts to the input. Then, we apply 10 V and measure the actual value. Finally, we measure 0.1 V, with the zero error removed, and compute the gain adjustment constant.

Input voltages greater than 100 V (1000 V range) create a self-heating error in the resistor network, as shown in Figure 3. This easily identified error is simply specified as part of the instrument's published error.

Based on this I don't know where you get idea that self-heating is not included in the published specifications for 1kV range?

I was talking about transfer specs. And at that point the datasheet says: "Measurements on the 1000V range are within 5% of the initial measurement value". If you leave these 5% you have to apply the also specified 12ppm x (vin/1000)^2. Which means you cannot use the 1000V transfer spec to transfer 100V to 1000V.

This is another aspect that I do not enjoy about 8508A. As far I tried, there is no way to configure 8508A for max resolution with custom NPLC conversion time. Best accuracy/lowest noise configuration is RESL8, FAST_OFF which translates to about NPLC1024 judging from spec conversion time. I understand reasons why Datron/Fluke limited this and removed ability of user to set NPLC, but it's rather inconvenient if we trying to work out the unit’s ADC performance.

I fully agree here. That is really annoying. For some fast measurements the 34401A works much better than the 8508A.

Is it possible to measure the INL of the bad 3458A the other way around?

In reverse polarity? I doubt that would make any point, as sweep is covering both polarity already.

What I meant was the direction of the INL measurement. If your measurement is actually sweeping from -11V to +11V than one could try to sweep from +11V to -11V. My broken 3458A is drifting almost monotonically upwards in readings. Therefore, the direction should make a difference if the bad INL is drift related.

Another big point is ACAL. Phil expressed his concerns about it before, even saying it is a "drawback" of the meter, however I respectfully disagree on that. ACAL is just another tool that engineer can use to improve the confidence and achieve better data results between the traceable metrology-proven calibrations. ACAL does not replace the need of the calibration with known uncertainty equipment, but rather makes possible to do educated accuracy improvements at time of the measurement, different from calibration day.

Yes, that is a big point and perhaps worth a complete separate discussion .
Just a few comments on that point..

As side effect this also means lack of manual range adjustments in 3458A, and deleting those range adjustment corrections during ACAL procedure in 57xx MFC. The point is that 3458A internally does very precise transfers, which are quite expensive to do externally (as you can learn from 8508A calibration procedures/certificate). And if you replace those transfers with manual adjustments, it's rather an open question if typical lab can meet tight 3458A specifications and account for all external errors of doing such a transfer set. After all owner of 3458A want to use it to measure the DUT, not fiddle with meter for days with manual calibration before you can use it.

I don't understand this point. You have to verify every range for a proper calibration. For low uncertainty you will need a good TUR. No matter if you only want to measure (calibrate) the range or adjust the range. So if your cal lab can calibrate the range (by applying a voltage for the given range with low uncertainties) than it is also able to adjust the range. There is no process which take days if you want to adjust it but is done in minutes for calibration. The only difference is in saving cal constants.
If your cal lab isn't able to adjust the meter in every range it will also not able to calibrate it.

The cost and effort to calibrate the 8508A with a characterized 5720A comes from the low uncertainties and not the lack of ACAL in the 8508A.

For volt-nuts ACAL is pure gold,

fully agree here

I don't buy that marketing push of "8508A is filled with VHP resistors and selected parts, hence it's more stable and does not need ACAL to meet its spec". 3458A don't need ACAL to meet specs either

I doubt that (just my feeling). That should be easy to verify. One could operate a 3458A without ACAL for a year. Or more pratical (because the meter will need the ACAL ) log every CAL constant after each ACAL for a year. That will show the stability of the meter without ACAL.

[Kleinstein]

It is not a surprise that the INL looks very similar at 5 PLC - the 50 PLC numbers should be just the average of 10 PLC internal conversions. Things might change a little at the very low end, so 1 PLC or maybe less (like 1 ms or even less).  Averaging can help to reduce the noise and also residual mains hum for something like 1 ms integration.  Here the details of the rundown come into play. At 10 PLC much of the resolution is from the coarse reference settings. The shorter the time the more critical is the exact resistor ratio and thus errors here might show up as an INL or DNL error: Mainly as short range variations over something like a few mV to maybe 10 mV.  So one would expect something like a more or less periodic pattern at low PLC setting.

Without AZ one might get some extra drift from the bad ADC.  There is a different resistor ratio involved, but the error might be comparable. So something like 0.5 ppm/day which would be 0.5 ppm of an equivalent reference Level of some 15 V ( = 12 V *50K/40K). So 7 µV per day. Still not clear how obvious this is compared to normal temperature induced drift. Comparing the gain drift might be still easier, as there is slightly less thermal drift included.

For the ACAL function, I see the argument of TiN. Having ACAL can really help in a calibration if one does not have the equipment to do the full calibration of the other ranges (e.g. 100 mV or 1000 V).  One has to do a kind of test, if ACAL is actually working right. However one could argue that one might get away with a slightly smaller TUR here. Especially in the early days of the 3458 it was difficult to get the accuracy at the 100 mV range. The number of point's to check maybe about the same for the 3458 and 8508 - however for the 8508 it might take a little longer at each point.

One possible advantage of ACAL is, that the transfer from range to range mainly depends on the quality of the ADC and not much on resistors. So the range to range transfer would be quite accurate even after longer time. For something like a 10 to 1 V transfer one can do this directly in the 10 V range with no extra problem. But for something like 3 V to 0.3 V  it can help if one can use the 10 V and 1 V range. This at least makes things faster.

For the ohms and current ranges I see the ACAL more like a way of reducing costs in having only one really stable resistor. I would really prefer to also have at least a 2 nd very stable shunt (e.g some 100 Ohms). This would allow averaging between the 2 scale factors and also a consistency check over something like 2 or 3 ACAL steps. A 2 nd reference element and a good self test for the ACAL steps could add quite some confidence over long time - though one would be at the guy with 2 clocks problem.

[e61_phil]

For the ACAL function, I see the argument of TiN. Having ACAL can really help in a calibration if one does not have the equipment to do the full calibration of the other ranges (e.g. 100 mV or 1000 V).  One has to do a kind of test, if ACAL is actually working right. However one could argue that one might get away with a slightly smaller TUR here. Especially in the early days of the 3458 it was difficult to get the accuracy at the 100 mV range. The number of point's to check maybe about the same for the 3458 and 8508 - however for the 8508 it might take a little longer at each point.

If you don't have the equipment to calibrate 100mV or 1000V than you cannot calibrate the meter on that ranges. I think it is that simple.
You can hope that these ranges are correct now, but you can also hope your meter without ACAL hasn't shifted. Maybe your chances with ACAL are better. But nevertheless, these ranges aren't calibrated. I also don't see the point for lower TUR.

Perhaps we should start a new thread on that topic. I find it very interesting.


The 8508A may need a bit more time, but the more cost intensive thing are the lower uncertainties, I would expect. It is half of the (normal) 3458A spec.

3458A vs 8508A 1year
100mV: 12ppm vs 5.5ppm factor 2.2
1V: 8.3ppm vs 3.4pmm factor 2.4
10V: 8.05ppm vs 3.4ppm factor 2.4
100V: 10.2ppm vs 4.9ppm factor 2.1
1000V: 22.1ppm vs 5.5ppm factor 4


There are also 2ppm extra in the 3458A datasheet one should add for NIST traceability. I don't know what to do with these 2ppm. The Fluke spec already includes everything.

[TiN]

Got few more data sweeps. This one bit longer. Here sick ADC is visible very clearly.



------

e61_phil, perhaps we have different angle on reading the specifications. You points are mostly about traceability, while for my experiments I was more focused on actual meter performance. In the end we arrive at same point of needing known and characterized reference sources for each function/range, but the purpose is bit different.

This leaves me with perhaps more interesting question, than ACAL function or application itself. Here is very practical problem:
Let's say I got four of my 3458A's fixed up and working (leave drifty A3 board out of the scope for now). They were never sent to Keysight for official calibration, but let's say for simplicity (lol) that two of the units were adjusted (that is run external ACAL with 10K/10V) by international voltage reference transfers with LTZ boxes/chips.

Then I got my Fluke/xDevs SL935 resistance standard 1 Ohm/10K with two calibrations on it (DCC bridge, uncertainty 0.33ppm). Because it is ovenized, we can ignore it's <0.01ppm/K tempco.
I also have few 10V LTZ boxes with value derived from comparison to Wavetek 7000 10V , that was in turn assigned value from Fluke 734A in USA (calibration <0.6ppm by traceable accredited lab).
Also have SR104, which was calibrated few years back, but have value agree with SL935 transfer within 2ppm from expected (transfer done with ratio mode of 8508/01). These are my primary standards in homelab.

I arrived at same "problem" as you outlined - to say that 3458A DUT is calibrate, one need to verify all of it's ranges and functions. So I built Fluke 5720A from scrap for that purpose, obviously with no metrology history. It's been running for about half a year already, and I was using it's ACAL procedure versus above mentioned Fluke SL935 1ohm/10k and LTZ 10V to evaluate  stability. Let's leave ACV/ACI functions out of the question for now, because they have separate set of complications.

Now we get to a point. Both 3458A and 5720A are ACAL instruments. It is clear that I do NOT have traceable calibration on either and cannot issue a legal certificate with absolute uncertainty of either.
However executing performance verification between the units (using 3458A to measure all 5720 points) gives me data within 3458A 24 hour spec. Also both 3458A's agree to each other very well, and even with borrowed Fluke 8508A that was calibrated in 2014 (by unknown body to me) is better than 1 year spec (but fail for 24 hour spec, no wonders).

We know both instruments have different ACAL approach, but for me data results show a good confidence than my meters are likely in spec. And ACAL does deliver actual calibration uncertainty in this case (relative to standards, not NIST/SI(!), even if we take a leap and extend calibration period to multiple years to include alien 8508A) as the internal ACAL transfer errors are already included in the absolute 24-hour specification of the 3458A/5720A. You can see why I was happy to see recently aquired 5725A to produce 10A current (again ACAL'ed by 5720A versus external 1 ohm/10kohm/10V) that 8508A reported within 40ppm.

Why this is important? I do not produce measurement equipment or provide legal calibration services to customers, so I fail yet to see need of traceability for my case here. Yes, in theory I could ship 5720A to Fluke for truckload of $$$$ (about 5K just for cal + shipping obviously), but even if the unit travels to USA and back undamaged, I would not be able to trust the results (shipping stress/temperature extremes, you call it) either. So paying for piece of paper that I cannot even use have little sense for me here. That's why ACAL is more than just a check for me, as I'm better off with own testiing for the calibrator and 3458As versus characterized standards. Eventually I will get standards with even better uncertainty, that will help me to establish absolute volt/ohm in lab to better uncertainty, so in the end I will just rely on statistical probability by using multiple references (>8 units for 1V/10V, >3 units per value for resistance).

I hope this helps to see my view?

---

I doubt that (just my feeling). That should be easy to verify. One could operate a 3458A without ACAL for a year. Or more pratical (because the meter will need the ACAL ) log every CAL constant after each ACAL for a year. That will show the stability of the meter without ACAL.

Well, i'm not sure "easy to verify" can apply here. But I'd love to see your take on this. I do log every CAL constant after each calibration (external measurement to 5700/5720A).
Here's data from my HP3458 "B" unit, where you can find calibration constants dump:

PDF-report, January 5 2017, calibrated vs characterized 732B/SR104, measured by ACAL'd 5700A (with same standards + traceable SRL1. This 5700A was also tested versus 3458A/02 calibrated by Keysight, Loveland.

In february 2017 this unit was shipped to me, and sits in my lab powered on 24/7 since. It does have pimped up A9 reference (15K VAR resistor to drop LTZ oven temperature).

PDF-report, March 6, 2018, calibrated vs my LTZ/SL935, measured by ACAL'd 5720A Hulk unit (same standards)

All measurement points are sum of uncertianty (keysight style) and relative to reference standards.

---

I've also got some 182M INL plots in mix. Those are interesting to look at too.

[ap]

Well, as a voltnut why would you not want (when you do precision measurements) to know what uncertainty every measurement is assigned with. And if e61_phil says that "If you don't have the equipment to calibrate 100mV or 1000V than you cannot calibrate the meter on that ranges. I think it is that simple.", it is just correct. And even if your gut feeling is that your meters are very precise (and they most likely are), what does "very" mean. Thats why in research and metrology every measurement is assigned an uncertainty and a confidence level. That does not mean that the value is guaranteed to be within the limits (actualy, with 95% confidence level, 5% can be outside; and they sometimes are, from certified cal labs, as I have had to experience unfortunatelly).

With all the equipment you have, why not go the extra little step (for you) and do a formal traceable calibration of at least some of your gear. Get your 10V and 10k calibrated on a yearly basis, and derive DCV, R, DCI from it in all ranges. Use a Fluke 752A and e.g. a set of SR1010 resistance transfer standards (costs for it will disappear in the noise of what you already spent on your equipment) and percision shunts and transfer all ranges from 10V and 10k. Do the uncertainty math and establish an uncertainty list of your standards/gear. And thats it. Every ISO9000 certified company could do that internally, and without being certified formally as a cal lab, pass ISO cert for its test gear and factory calibrations. Been there, done that.


With respect to the 3458A uncertainties e61_phil mentioned, to be fair to it, one should rather use the opt 002 or even the better the 2ppm/a reference to compare to the 8508A. Then things look better (4.05ppm+reference verus 3.4ppm at 10V and opt 002). And when you do a highest precision 10V measurement using transfer against a 10V standard, the 3458A usually beats the 8508A because of its linearity. It is almost a shame though that the additional uncertainty of the 10V/10K standards listed in the manual are so high (2ppm/3ppm). But I would guess that one can get much lower when ordering the more expensive cal services from Keysight. May depend on site too.

[TiN]

ap

Get your 10V and 10k calibrated on a yearly basis, and derive DCV, R, DCI from it in all ranges.

Yes, in process of that, but it takes long time to acquire and proof all the equipment. As we all know, equipment for ebay is always 100% faulty regardless what seller says, unless it's tested otherwise.

and do a formal traceable calibration of at least some of your gear.

Shipping of the 3458A alone to US is some 400-500$, and local labs don't have uncertainty good enough. Also language barrier in a bit here .
For 2019 there are plans to perform calibrations for my standards to intrinsic volt/ohm. Only missing part left will be AC Voltage, cost of which is not disappearing in the noise, but actually can topple any of the DC/Ohm cal set easily.

I do not try to sound like ACAL is the solution for everything, it is not, and will never be. For me however it is important to see numbers derived from ACAL results, and number acquired by slow delicate manual external transfers to be below the DUT's manufacturer 24-hour specifications.

It is typical in metrology to have measurement results backup by multiple methods. That's why we have Kibble balance and Si spheres for kilogram, and they compliment each other for better confidence in measurement, not compete or exclude. So it's the process that thrills me, not necessarily the numbers on the paper with a stamp.

Since I'm not in calibration business, it will take long time (only few hours a day for this hobby) to perform a full 3458A 24-hour level specification calibration, even if I'd have all equipment necessary on table (also calibrated and verified!). When I don't need absolute best uncertainty, ACAL is good enuf tool to bring meter to best capability between those tedious calibrations.

one should rather use the opt 002 or even the better the 2ppm/a reference to compare to the 8508A.

Agree on that part. Considering that 8508 don't grow on trees, one could easily buy two new 3458A with option HFL reference and still have some money for beer, instead of one 8508

[ap]

Well, quick search only, sounds like Microprecision has a lab in Taiwan and can calibrate the 10V to 0.6ppm, 10k only to 2ppm. That saves you shipping to the US. And better than Keysight uncertainties on paper.

[Dr. Frank]

..

Get your 10V and 10k calibrated on a yearly basis, and derive DCV, R, DCI from it in all ranges.

Use a Fluke 752A and e.g. a set of SR1010 resistance transfer standards (costs for it will disappear in the noise of what you already spent on your equipment) and percision shunts and transfer all ranges from 10V and 10k.

Do the uncertainty math and establish an uncertainty list of your standards/gear.

Every ISO9000 certified company could do that internally, and without being certified formally as a cal lab, pass ISO cert for its test gear and factory calibrations. Been there, done that.

Adrian, Illya, Philipp,

now I'm really confused about (volt-nuts grade)  Traceability and Calibration in regard to AutoCal instruments, like the 3458A, 57x0A, 752A, SR1010, and similar.

To adjust a 3458A, you basically need a certified / traceable 10V and 10k reference. (I disregard the additional ACV calibration, at the moment)

A working 3458A will then adjust all its ranges and modes  to the 24h specification by its ACAL ALL, theoretically in a traceable manner. A similar process (theoretically and practically) is going on inside the 5700A, Illya has built.

To really calibrate either the 3458A, or the 5700A, you need a verification, that all their modes and ranges really fulfill the 24h specification. So another ratio transfer instrument is needed.

Now it's getting circular:

How to achieve this, if the (transfer) instrument used for verification itself does not intrinsically have (accepted) transfer traceability? Especially not with the required T.U.R. of say 1:5.

Or in other words, what is the special feature of these 57x0A instruments used by any Cal Lab to calibrate the 3458A, that they are accepted for verification / calibration?

That's not described anywhere, or do I miss something?

So if there is no special requirement, is Illya's  5700A basically good enough to perform a full verification / calibration of his 3458As, presuming, that his 10V / 10k standards are good enough, i.e. (volt-nuts grade) traceable to some level of uncertainty?

Why should one need in addition to a 5700A another 752A for DCV verification, or a set of SR1010s for Ohm ranges?
The 752A gives traceable results 'by design', maybe more obvious than a 57x0A, same goes for the SR1010.

But these could also NOT be verified by better standards.. so in the end a verification against these standards would also be reduced to a complementary cross check, only.
So the 3458A would be checked by the 752A, for example, and the 752A itself reversely by the 3458A.

Is this already a sufficient criterion, to get traceable calibration?
(Again: I do not know of any calibration procedure / description, defining additional measures on the used ratio standards)

Additional questions for discussion:

The 57x0A is meanwhile accepted as an instrument to deliver internal traceability by AutoCal, fully accepted in U.S., but partly only in EU. (See Fluke paper: https://eu.flukecal.com/de/literature/articles-and-education/electrical-calibration/application-notes/artifact-calibration-eval)
Did I understand that correctly?

Has there ever been done a similar test for the 3458A, so that its internal traceability is accepted  as well by the metrology community?

If so, is transfer traceability / calibration in the end only a technique of having a verification by '2nd opinion', a sanity check only, with statistical analysis?

Frank

[TiN]

From my understanding, traceability have nothing to do with calibration (comparison of the UUT to known reference) itself, in only means that there is unbroken chain of calibrations with assigned transfer uncertainties all the way up to SI Volt/Ohm/Kelvin/etc.

Using ACAL instrument by definition defy this principle, unless ACAL also provides uncertainty of the transfer as well. And 5700 does that with exception of AC functions, which is well explained in PTB/NMI/SP's public study document of the Fluke's ACAL. AC function traceability broken because internal AC/AC thermal sensor flatness/accuracy is not verified or measured, so one need AVMS like 4920 or 5790A or 792A to perform the ACV/ACI traceable verification.

So yes, shipping and doing external full verification to cal lab standards would give traceability, but I doubt that it will improve uncertainty of the applied specification much, because standards I use to ACAL the units are in comparable level already.
If either of my ACAL instruments have internal fault, making internal transfer to specific function incorrect/errornous it would be spotted by multiple of external verifications against different instruments during the my own performance verification.

So the 3458A would be checked by the 752A, for example, and the 752A itself reversely by the 3458A.
Is this already a sufficient criterion, to get traceable calibration?

No, because there is no trace to SI unit chain. If there is imaginary fault that makes 3458A ACAL for example -10ppm off, and at same time imaginary fault in 752A that makes -10ppm off error, then we would have hidden error -10ppm to the Volt ratio. Traceability verification to cal lab will catch this error. 

Has there ever been done a similar test for the 3458A, so that its internal traceability is accepted  as well by the metrology community?

Not that I remember reading about. For me it looks that labs are legally bound to use traditional calibrations up to NIST/PTB/SI realization, so for business point ACAL is useless, and in best case just a diagnistics/check tool at best. Something hints me that most labs do not even bother running ACAL/Calchecks are preserve data history of the 57xx unit.

All this actually gives quite practical angle that can in theory save some $$$. If I want my 3458A/02 calibrated and adjusted to best uncertainty possible, I could do the adjustment to best data I possibly can using ACALed 5720A (as I can do adjustment cycle even every hour if desire so, unlike the calibration lab service), and then just ship the unit to the callab with cheapest calibration level possible just to obtain traceability and get both the legal traceability and paper with a stamp, and instrument with low uncertainty. Eh? 

Another bothering thing about all this traceability thing. We keep hearing "unbroken chain of calibrations to SI" but to my knowledge, there is no place/site/service available where I can take my nice official calibration report, such as this resistance certificate, and trace back the measurement to the SI directly. Report says that MI 6010B DCC and LN4210 were used as references with their ID numbers 140132 and 140124. So next in chain up should be two calibration certificates for these 140132 and 140124. And then certificate of whatever was used to compare those standards to the QHE. But what is the mechanism to obtain those reports?