K2001 ADC

[GigaJoe]

thermal something ... as i said it quite sensitive to temperature, i tossed paper towel around , average wandering down to 3 microvolts up-down. 

so i 'm still wondering about R826 (3K) ,   it simple to disconnect it and to see the result , the question is: if it affect something, how to find out .

and related:  does it mean opamp THD directly affect linearity ?

[Kleinstein]

The 3 K might effect linearity, but it is hard to tell how and how much. I looked at the effect of op-amp loading, but with the 2 OP-amp integrator. In my testest I could not see a difference (e.g. to about 0.1 ppm FS level).

It may be a bit more important with only 1 OP-amp in the integrator.
I would suspect something like an U² contribution, from the output current effecting the GBW of the OP-amp. So not the classic THD, but related.
The relation is no direct and it can depend on details of the run-up.
One could see an U² contribution as turn over error (change polarity of external reference, with good care about offsets). Not sure if this would be actually visible or still in the noise. The estimate I made for my system was 0.5 ppm FS - so barely visible with the K2001. With the larger integration capacitor and slower modulation in the K2001 the error may well be smaller than this.

[GigaJoe]

resistor R826 - removed.
0 - noise affection,  std dev 2.63 µV
guess it wrap up ...

20V range:  output:  -0.000,220 V ( graph: 1.2 - 2.6 µV )
Std dev: 2.63 µV ; LF noise: 1.11 µV ; HF noise: 2.24 µV


2 Volt range:   shorted ,  output:  -0.000,001,9 V ( graph: 1.4 - 2.4 µV )
Std Dev ≈ 0.162 µV ; σ_LF ≈ 0.017 µV ; σ_HF ≈ 0.161 µV


0.2V  range:  output:  -0.000,002,08 VDC, ( graph:  0.18 - 0.23  µV )
std dev  ≈ 0.080 µV  ; σ_LF ≈ 0.015 µV  ;  σ_HF ≈ 0.080 µV


0.16 x 10 = 1.6 µV , vs 2.6µV , so for 20 V almost the bottom
and 200mv  0.08 vs 0.016 ... a few times more noisy .....   i think i need to change something ....

[GigaJoe]

seems final .. did update the previous post with result ;   some cooldown ,  then will do a second unit ;  need to think about opamp replacement for composite integrator ( 2 opamp replacement)

[Kleinstein]

The 200 mV range noise is often dominated from the frontend noise (amplifier and resistors in the protection).
The observed 80 nV RMS for 10 PLC is actually quite good  (e.g. the HP3458 is similar).

The 2 V range has a very similar input path an amplifier. This means a good part of the 160 nV RMS noise is from the front end. This is normal and not that surprising. The rest of the noise ( roughly 3/4) should be from the ADC and a littel more noise from the FB resistor in the amplifier.

It is remarkible that the 2 V noise is more than 10 x lower than the 20 V range.
This means some ( especially the LF part) of the 20 V range noise is from the front end. For the 20 V range the path is different and looks like more noisy ( U322 and U341). For U322 it would be current noise that can be a problem - the LT1007 has quite some low frequency current noise that is not good for 25 K impedance. It is however hard to find a better alternative that is also very linear.

[GigaJoe]

yeah for 20V , i add some modification as well ... U341 replaced to OPA140 , in later version 2001 it BB OPA124; a lot caps for filters especially  ±8V and  ±6V , zeners was shunted by polymer caps, Q330 not so much for PSRR actually ;  so not ADC only ..., and true, i walk around LT1007 - there is no much to substitute , that significantly do an improvement ...

so i measure my 7V and:
7.05647777 ± 0.19µV ( ~95% confidence)
of course it not an exact # , but before i have ±2.6µV  ;  improvement over an order in confidence ...

interesting note:   RST command:   and output starting shaking like crazy, then stabilized approximately after 2000 readings , so yes , constants do recalculating and store in the memory

[Kleinstein]

The problem with replacing the LT1007 (U322) is that this amplifier also effects the linearity. In the 2 V and 0.2 V ranges the amplifier is used with gain and for good linearity very high loop gain is wanted. In addition the OP-amp drives quite some load (e.g. some 8 K) and thus can see thermal effects.
Already the LT1007 specs don't guarantee good linearity, but it seems to work.

There are a few options with lower noise (e.g. OPA210, OPA828 ?), but they have less loop gain.
An OPA189 or ADA4522 may work, but add AZ switching spikes and may not like the loading and show thermal effects.
The 20 V range is with some 15 K source resistance and thus sensivive to current noise.
Reducing the loading part should be possible with an extra buffer to provide most (like 90%) of the current .

[GigaJoe]

LT1007 ( both ) replaced by OPA189

20V range:  output shorted  :  -0.000,220 V ( graph: 1.6 - 3.0 µV )
Std dev: Std dev: 2.28 µV ;  LF noise: 0.91 µV ; HF noise: 1.98 µV


a little bit down ...  probably LF may be a bit lower after 24h warmup , current state after 2 hours  ....

I see more artefacts like:
+0.0002214
+0.0002221
+0.0002218
+0.0002218
+0.0002216
+0.0002209
+0.0002211
+0.0002214
+0.0002214
+0.0002192

will update this post tomorrow , to see noise after 24H

[Kleinstein]

The regular correction jumps (roughly every 55 seconds) are strange. I kind of have expected such steps for the refrence voltage or overall gain, as this is also seen with the DMM7510. However with a short there should be essentially (would be ppms of µV range offset) no effect of the gain. Having such jumps also for the offset could explain quite of the LF noise. The obvious part are the jumps, but it also means the some offset part is not corrected in between and thus amplifier or ADC noise from rather low frequencies (e.g. down to 0.01 Hz)  making it's way to the result.

A possible explaination for the rather high noise in the 20 V range compared to the 2 V range could be in the more tricky way to measure the ADC+amplifier gain. The 2 V range can directly get a reference voltage near 2 V (e.g. 7 V / 4 via the PWM divider). For the 20 V range I see no direct path to apply the 7 V ref directly to the input. For the gain part there is only the option to use the same resistors as gain of 1.5 and thus get some 7 V in amplifed x 1.5 to get 10.5 V to the ADC. To get the 0.5 gain one would have to subtract another 7 V measurement with gain of 1. This extra step would make the gain correction loop more complicated and possible longer to make up for added noise. The longer loop may also be used to zero averaging and thus more noise there. Alternating between a 7 V ref. and 10 V ref. could also cause the 0.5 Hz background seen before in the FFT data. The measurement before can have a small effect (e.g. dielectric absorbtion and thermal effects).

One could test this theory by looking at the signal going to the ADC with a 2nd meter or DSO over a few minutes. Ideally apply a small voltage (like 1 V to the input to see the difference between the input and a zero and the ref. voltages).

Another point to maybe look at is to plot the allan deviation for the 2 V range with a short. This may need a long set of data well after warm up. The question is if there is also the plateau up to some 100 seconds like in the 20 V range.

[GigaJoe]

i told you it the one big wonky calculator ....

i think it final ( The device itself is the only thing left to replace) :
Noise rms ( basically std dev, in our case , the same #):
20V - 2.15µV ( and still around 3 µV wandering , before was 9µV ,  3µV - still annoying )
2v - 0.23µV   ( previous result optimistic , i skip last digit by mistake in the script  )
0.2V - 0.16µV  ( prev .. same bad)

now it seems valid as the numbers it in the same circle

waiting parts for the second one ,and measure my 7V couple days ...

addon:
drop 1h sample:
here is output gibberish:  https://pastes.io/1OmTOPvV
interesting that noise calculated as 2.91 μV  , so the source voltage probably 1.96 μV ( but my exp. below 1μV )

[Kleinstein]

To compare to calculated noise, it could make sense to also record a measurement in non AZ mode (1 PLC and 10 PLC).
Especially the high frequency noise part or 1st point in the Allan dev. curve would be interesting. So no need for long records.

From my spreadsheet for ADC noise the K2001 ADC does not look great, but still quite a bit better than the AZ mode measurements.
The 1 PLC case and a little more averaging may not be that bad in comparison.

The difference to the AZ mode data would likely be mainly from 1/f noise that is poorly suppressed with too much averaging at the wrong place.

[GigaJoe]

so you saying,  I should do, let say for 6 hours, 4 measurements as combination of 1/10 PLC , AZ ON/OFF ?

[Kleinstein]

The tests with AZ on and 10 PLC are already done.
For the non AZ mode, there is no real need for a very long test. Some 20 or 30 minutes would be enough.

[Kleinstein]

The allan dev. curve looks wrong again. A log-log slope close to -1 is too steep to be realistic. White noise would be -0.5 and with 1/f noise one usually gets less slope and more of a plateau and often bend up later.
The curve could be the Allan variance which is square of the deviation.

[GigaJoe]

yep...
seems correct ; i added some presentation from NIST slideshow

[Kleinstein]

The test with the 7 V source is quite complex and contains quite some noise from sources besides the ADC.
There is noise from the external reference, noise of the low noise zener at the ADC, noise from the gain measurement steps, and for longer times noise from the LM399 reference.

For the Allan dev. curve I see different ranges:

The intial blue part may be dominated by ADC noise from not too low frequencies.
The rise in noise up the the peak at some 40 seconds looks like the extra LF noise due to the poor AZ implementation. It may include some 1/f noise of the K2001 zener ref, but I would not expect it to be relevant.
The 2nd blue part looks like the range beyound the zero averaging and thus LF noise suppression from AZ mode finally working.
The lower plateau is than likely limited by reference 1/f noise. This would likely be mainly the DMM internal LM399 and not so much the external ADR1399.

The initial point looks not that bad. It it not great compared to modern DMMs, but OK to call it 7.5 digits. At around 1 second the 1 µV level is reached.
The really anoying part is the peak at some 40 s. I am afriad this is mainly a software issue, that fails to suppress 1/f noise around the ADC.

[GigaJoe]

ok .. need more data to get ...   about week or so ...
with 1 PLC i missing data , a lot , need change to work with buffer
+7.0565416E+00NVDC,  +126011.551984secs, +69410rdng#, 00extchan
+7.0565392E+00NVDC,  +126011.910213secs, +69428rdng#, 00extchan

10 PLC AZ OFF:
probably that hump def. AZ

[GigaJoe]

set a script to use memory ...  1 NPLC  , AZ -off  ; a lot of repeated data:

7.05645130;7.05645130;7.05645370;7.05645130;7.05645130;7.05645370;7.05645610;7.05645610;7.05645370;7.05645610;7.05645130;7.05645130;7.05645130;7.05645130;7.05645370;7.05645370;7.05645610

but it sequence:
+7.0564513E+00NVDC,       +1.664722secs, +00050rdng#
+7.0564513E+00NVDC,       +1.698283secs, +00051rdng#
+7.0564513E+00NVDC,       +1.731648secs, +00052rdng#
+7.0564513E+00NVDC,       +1.764260secs, +00053rdng#
+7.0564537E+00NVDC,       +1.797670secs, +00054rdng#
+7.0564537E+00NVDC,       +1.831163secs, +00055rdng#

controller cant keep up   ; doesn't make sense for me to measure like this ...

[Kleinstein]

With 1 PLC the quantization error can become a bit more relevant. The steps get 10 x larger, while the noise is likely only about 3 x higher than with 10 PLC. This gives more often same values in a row. This is still not bad, just no full 7 digit resolution with 1 PLC.

If the datatransfer is limiting, one could skip the time stamp. The 1 PLC non AZ mode should be pretty fixed sampling rate. Even if there is a little error on the time scale this would no be so bad.

The distorted shape of the histogram for the non AZ mode comes from the drift part: initially slow and than a bit faster.

For the 10 PLC non AZ case the first point of the Allan dev curve is at nearly the same level (even a bit higher) as for the AZ mode. With normal (simple, like HP) AZ mode one would expect roughly 30% lower noise. This clearly points to an AZ mode that uses averaging - this helps for the short time, but lets the 1/f noise through up to some 40 seconds (the time used for zero averaging). Once beyond the averaging time there is the extra 40% (from usinf a 2nd measurement) noise for the the AZ mode, but also suppression of 1/f noise and thus the downward trend for the AZ mode after that.

[GigaJoe]

so ,practically , to display a data on the dmm display , ideally, it would be setting as moving average for around 20-30 sec, or 40-60 samples ( not 100 as at maximum ) that would be a lower noise ; and for maximum precision for very stable source, collect around 1200 samples , and do processing ...

[Kleinstein]

AFAIK the K2001 has the digital filter option the is doing a moving average. If really needed the averaging could be done later with saved data. I consider those high resolution meters mainly useful with saving the data. A human is hardly keeping up with processing the display data and is too error prone.
If in doubt one would like to look back at data.

How much averaging is wanted depends. For reading the display something like 1 update per second is usually good.
When using saved data there is no need to do much averaging in the meter, 10 or 20 PLC is aready slow enough to not get excessively large files.
If in doubt the faster data could help analysing issue in hindside.

The issue with the K2001 (and as it looks like quite some other Keithley meters), that causes quite some extra noise is how the AZ mode is done. It looks like they average the zero readings over some 30-50 seconds (could also be 60-100 readings) and only than subtract that average value from the readings. This works nice for a short time (e.g. up to 1 or 2 seconds on the allan deviation curve), but it is a desaster for the longer time scales. They really missed out on implementing the simple AZ mode (subract the last zero reading, or at most the average of the last 2) at least as a mode to choose. For some 20-200 seconds this could have reduced the noise some 4-10 times.
Averaging over quite some time (likely even several minutes) is good the reference / ADC gain measurement.

Getting low noise even with the poor AZ mode would require to really keep the 1/f noise down.

[GigaJoe]

AZ triggering, quite clear picture:

--


cover fan fo 1 minute, then open it , spike 100uV P-P:
no idea why,

[GigaJoe]

Some updates... 4 days sampling my 7V reference.

Good news: the 7V is performing extremely well.

And the not-so-good (mediocre, to be precise) about 2001:

Here are the results: it ran for approximately 4 days, then I calculated a baseline, let's say an average, accounting for drift and daily day/night fluctuation. I then took the raw sanitized numbers, averaged them over 50 seconds (roughly 100 samples), and created a plot showing the difference from baseline. This makes it clear how the device is performing. Red dots = excluded samples.
https://i.ibb.co/xqtf7QnV/7-V-long-full-madev-fixed-window.png

In defense of my 7V reference — I used the 34411A for sampling, as it has one of the lowest tempcos at 5 LSD (OFF to ON state; the fan likely helps), with only a 3 µV difference between the average at the beginning and the end of the 4 days. I consider that precision, noise, etc., cannot guarantee a drift measurement accurate to 3 µV.

So from the plot I can see drift, and my script calculates it as 0.1901 ppm/24h — but it's slowly trending downward. I think it needs at least a month to stabilize. (an assumption , device sitting on the shelf for many years, due to extremely strange behavior )
As for those sudden drops and jumps — I can't attribute them to LM399 behavior, even if that effect exists, it would still be buried in the noise after averaging. So those jumps are something else. So far, those are the 2 disappointing factors: slow drift upward, and jumps ..

PS: i'm assembled some 5V source, ad586 + OP07+transistor ... still thinking how this can improve ADC , but again ... it microvolts there and there .... and it def. can .....

[Kleinstein]

The saw tooth like signal is mainly there at the start. It gets less with time.
This part is very likely the meter internal correction of the gain, from drift of the zener diode at the ADC. The measurements of the 7 V reference are used to recalculate the ADC and amplifier gain and this is than corrented at some time. It looks like this is about every 50 seconds.
Chances are there is done together with the AZ averaging / correction (would fit the position of the noise peak with a shorted input). The amplitude seen is however a bit large to be offest drift.

The first hour of data look a bit off from a meter that is still drifting internally and it not yet warmed up. So maybe recalculate the allan dev. for data without the first 1 or 2 hours. This K2001 seams to have quite some drift before the gain corrections.

One can likely improve on this gain drift, by changing the zener current (e.g. change R854). These low noise zeners usually have a current dependent drift and with the right current one may get close to zero drift there. As is the meter is likely still warming up initially. The signal creeps up, which corresponds to the ADC ref. to drift down. More zener current should reduce the drift and thus the saw tooth signal.
So maybe try 1.2 K or 1.5 K instead of 1.62 K. The difference should be visible especially in the first half hour after turn on.

[GigaJoe]

I actually had an idea to unsolder it and find out a current equal to precise TC == 0 ... 
another wild option: 2DW232 ... but it has around 6ma at 5.4 as a zener,  and extremely sharp curve, a bit left or right and it off  .  with zener +diode it around 6.2 V , and current to TC=0 was around 20mA in may set, even 6.2V  can be not enough to make 10V  and all this chain reaction  happens ....

so yeah ... Zener current for jumpy correction are really good point ;  lets do it


Upd1:
3k9 - in parallel , 1K14 resultant , I think it too much ..
saw amplitude became higher https://i.ibb.co/93BGzKBt/Clipboard-04-28-2026-02.webp

factory: 2.37mA 6.4199 V
current: 3.37mA 6.4393 V


i think i need to start from 1.5ma  , and monitor behavior and increase the current ... 
this is kinda decade resistor and logging together ...
will see overnight the behavior  ...
btw , it looks much cleaner as the previous one
https://i.ibb.co/rnz31s3/Clipboard-04-28-2026-01.webp
In passing i hook 460uF to fan , it would be really laughing moment if it affect the noise ..

thinking:  No ... this is sucks ... it a fundamental flaw in the design ...  the zener voltage converting  to 10V that a feed source for LM399 , in this chain changing voltage in zener due to thermal, changing lm399 reference, that suppose to be constant , and calculation get skewed ... In the current design , the easiest way would be to cut negative 10 volt from ADC, and make stable -10V.
 .... or precisely adjust zener to minimize thermal runaway.