AD7172-2 Noise tests.

[David Hess]

There are actually opamps that can drive large caps like 10uF or more. It doesn't mean they can drive them at high frequencies...
It's quite the opposite: they drive DC but they won't oscillate.

Quite a few old operational amplifiers will work that way, including the 741 and OP-07.  The large capacitance combined with the output resistance implements dominant mode compensation, with the ESR adding some phase lead so this works better with tantalum or electrolytic capacitors.

[David Hess]

Take a pen and calculate an impedance of the 10 uF capacitor for 10 Hz frequency, or 0.1 Hz Than compare output impedance of the OPA that tolerates capacitance and make a conclusion.
The fact is, there is no OPA that can be filttered just setting cap at the output no any voltage references. Output impedance ratio to cap's impedance is enormous, using R-C network is the only way.  Test results talk to itself, don't be stupid

The feedback provides a low impedance at low frequencies.  The large output capacitance provides low impedance at high frequencies where the open loop gain is lower and the output impedance of the closed loop system is increasing.

This configuration is especially useful for rail splitters, and has the advantage of driving difficult loads without instability.

[Kleinstein]

The capacitance at the OP-amp output can act as a higher frequency filter, but of cause not as a low frequency filter. Wether calling it filter or buffer is just looking at the same thing in frequency or time domain.

Take a pen and calculate an impedance of the 10 uF capacitor for 10 Hz frequency, or 0.1 Hz Than compare output impedance of the OPA that tolerates capacitance and make a conclusion.
The fact is, there is no OPA that can be filttered just setting cap at the output no any voltage references. Output impedance ratio to cap's impedance is enormous, using R-C network is the only way.  Test results talk to itself, don't be stupid

The results shown at the start are a bit misleading:
The case with no filter seems to also miss the high frequency filtering for the reference, that is standard and should not be skiped.
The case with the low frequency filter is not just about the filter, but about using the same source for the reference and input signal. With the same filtering for both paths reference noise is suppressed. This also works with way less filtering and does not proof the effectiveness of the filter. Test relativ short data window makes the test insensitive to the drift introduced by the filters.

The tests are done with a quite noisy reference. The alternative to excessive reference filtering that compromises stability is using a lower noise reference to start with.

[iMo]

The key message here by Kleinstein and others is the applying a large RC filters in the Vref affects the "long term stability" (as the large value capacitors introduce various drifts). Your measurements with your large RC will be less noisy when evaluating small/short set of results (as you did above), but may drift significantly when doing long term measurements.
It has no sense to achieve 1LSB peak to peak noise at 1sample/sec (doing 1minute long measurements) with the 1500uF filter capacitor, when it will do ie. 2674LSB p-p drift from day to day or week to week. Thus you may achieve much more reliable results with a noisier Vref chip with small less drifty capacitors in the filter and do averaging/smoothing of the results. Another variant is to use less noisy Vref chip per se.

[MasterT]

The key message here by Kleinstein and others is the applying a large RC filters in the Vref affects the "long term stability" (as the large value capacitors introduce various drifts). Your measurements with your large RC will be less noisy when evaluating small/short set of results (as you did above), but may drift significantly when doing long term measurements.
It has no sense to achieve 1LSB peak to peak noise at 1sample/sec (doing 1minute long measurements) with the 1500uF filter capacitor, when it will do ie. 2674LSB p-p drift from day to day or week to week. Thus you may achieve much more reliable results with a noisier Vref chip with small less drifty capacitors in the filter and do averaging/smoothing of the results. Another variant is to use less noisy Vref chip per se.

Long term stability competely relies on stability of the voltage reference.  Period. It has snothing to do with filter.

"It has no sense to achieve 1LSB peak to peak noise at 1sample/sec" - It does HAVE, this is an objective.
 I don't give a f. if FPGA-brain-dead retaired does not understand. I show results , it's pure Science - not a "likes-don't likes" vote for AI, opinion does not accounted.

[iMo]

A good test would be, for example:
1. a 1 day long measurement with the standard setup (or longer)
2. a 1 day long measurement with the large RC filtering (or longer)
3. all above done with separate Vref and Vinp sources, the same sampling period
4. make ADEV/MADEV of both results to see how it performs short/long term..

[MasterT]

A good test would be, for example:
1. a 1 day long measurement with the standard setup
2. a 1 day long measurement with the large RC filtering
3. all above done with separate Vref and Vinp sources, the same sampling period
4. make ADEV/MADEV of both results to see how it performs short/long term..

Ask AD to do it for you.
10 seconds at max. read above, or get lost

[Echo88]

They cant grasp your genius MasterT.
While were at it: can you give more hints to the stolen USSR secret sauce that made the 3458A the pinnacle of DVMs like mentioned in the 3458A-AFE-thread or will that lead to forceful defenestration and the end of this thread?

[MasterT]

They cant grasp your genius MasterT.

Who care. I know that I made over 200 inventions, that were stollen and than re-flashed to some idiots in the west.
 It can't be disscused openly, since there are a few NDA:
- how my algorith "identify AI" works;
- how human brain works;
- memory manipulation;
- parasitics economy functioning;
etc.

[Kleinstein]

Long term stability competely relies on stability of the voltage reference.  Period. It has nothing to do with filter.

Ideally that is the case. However a poorly designed filter can add to the drift, e.g. from capacitor leakage current. In the current example this are a claimed 10 nA and 10 K resulting in 100 µV of difference that the filter can make. For the leakage current one can not expect it to be especially stable. So the filter as shown may cause drift issues.

[dietert1]

Let's remind that the best commercial voltage calibrators use PWM switched references with RC filters. As far as i know this works to about 0.1 ppm.
If one ovenizes the filter including its caps, this can also work at long time scales. The caps i tested were film caps, e.g. an AVX FFLI6L0657K cap (650 uF  1000V}. At time constants of 200 to 500 seconds one could see temperature changes as voltage changes: Capacitor at constant charge means voltage inversely proportional to capacitance affected by its TC. One also wants a method to initiate the filter without waiting days. There is no magic about the 10 sec time scale, except temperature will be near constant within 10 seconds, so it can be done without oven.

Regards, Dieter

[MasterT]

 100 uV accross 10k resistor indeed looks as a big error, but thinking a liitle bit about for a fix I came to solution "zero cost - no external hardware".
 All it needs to connect pin4 of the ADC to other side of 10 k resistor. Than in software periodicaly, let say 10 sec,
( 1 sample out of 100 ) I would reconfigure input MUX matrix to connect (adc-) -> Ref, (adc+) ->pin.4, and voila - real time capacitors leakage current monitoring. Accuracy could be better than 100 nV using LPF in software .
Two diff channels based on AIN0-AIN1 & AIN2-AIN3 would run as planned

[Kleinstein]

Some correction is indeed possible, but one has to be a bit careful. The extra measurements add to the noise - one starts with the noise one wants to filter out and thus needed the digital filtering (LPF) with quite some samples to keep the noise low. So only 1 out of 100 readings for the correction may be on the low side. To get a 10 fold noise reduction over a single reading it would need at least 100 readings to average. To add less than 1/10 to the noise it would be more than that.
With the digital filter one may have to be careful with the phase shift delay, not to amplify noise from the transition region of the analog filter. Not sure about that, but there could be some trouble from the delay, as the LP filter naturally delays things and may thus add current noise and past noise that don't fully compensate.

Another point is that the ADC input and switching of the mux can be quite some load to the filter and may thus need a buffer to not add extra drift.

[MasterT]

People say "A picture is worth a thousand words".
 Scale X - 7 min., Y - 2 uV per grid.
Red line: 1-st channel diff inputs shorted to track adc offset variation;
Blue line: 2-nd channel  measuring noise-free 2V;
Make your guess what reference for adc was in use.
Circuits updated with a 0.1u cap at pin.4

[MasterT]

Updates.
Quality of the internal V. Ref of the ad7172 is poor, so I move to external reference.
https://www.eevblog.com/forum/chat/lm431-(tl431)-0-1-10-hz-vp-p-noise/new/#new

 TL431B from OnSemi is my choice, beats up REF5025, REF6141, MAX6164 I tested so far with.  Noise tests 0.01-1Hz  - based on screen picture over 7 min. interval.

 Having ~48nV /Hz^-2 it costs only $ 0.5, about 1/50  compare to LTC6655 ($22CAD on mouser).

 

[MasterT]

Video shows dynamics.
ADC is sampling 3 channels, 1 sps /channel, 2 externally connected to filtered TL431B, one is maintenance tracking offset-reference fluctuation.
  Trends: update rate 1 seconds, data not filtered, as it is, zoom 2 uV/div , time scale 7 min overall.
  Big digits: update 3 Hz, filtered running average LPF 32 samples.
https://youtu.be/aFE5hyFjUE4

[Kleinstein]

One can clearly see quite some slow drift with the TL431 references, likely from thermal effects.
The slightly better references like ref5025 come with a much lower TC and also correction for 2nd order effects in the temperature. This correcton comes with some additional noise, but is needed to get a low TC over a large temperature range, so the reference can be used without temperature control.
Good 0.1 ... 10 Hz noise is pretty much useless if the TC is high.

It is still impressive how low the noise of this TL431 variant is. It is still not a very useful reference for a low sampling ADC. It could be a good choice for audio though.

[MasterT]

I should say that those up-down drifting caused by me hovering with android over test setup. TL431 installed into breadboard and there is 30 cm  du-pont wiring, so I'm not sure if it's thermal effect due IR radiation by my body, or RF interference injected from android device. Taking a picture takes less time, so photo is always less jerky.
 Even walking less than 1m away from test bench shifting voltage up and down.
 REF5025 is no difference. can't hold steady voltage if I move hands over it, in my another project I even cover IC by FR4 - with both side metalization, Cu known for good IR/ RF reflection.
 

[miro123]

I did some mid term /90bdays/ test with several TL431 - diffrerent manufacture and batches. Their mid term stability is quite poor. there is workaround over TC but not about humidity and long term stability. I did not made any further investigations. Ihave many cases with unexpected voltage fluctuations. e.g. DUTs are at lab condition continuously powered. lab temperature and humidity curves are similar, measurement are performed early in the morning /less background noise/. I sow frequent spikes in positive and negative direction. Also the ageing slope is quite steep