AD7172-2 Noise tests.

[MasterT]

 Subject of my research is oulined here:
https://www.analog.com/en/analog-dialogue/articles/why-does-voltage-reference-noise-matter.html

 Short summary:
1. Noise measurements of the ADC with input pins shorted - absurdity.
2. Simple LPF using ultra low leak polarized electrolytic capacitors is very efficient way to filter internal /external V-ref.

Tests results.
   Input always 2V, filtered low noise source obtained with help 10k & 1.500 uF RC network. Internal adc buffers active for both,
   IN & Ref-IN.
1. V-ref selected internal - 2.5V, using adc register 0x20 setiings. RMS value: 10-11 (LSB)
2. V-ref selected external - 2.5V, switch on, low time constant. RMS value: 5-6 (LSB)
3. V-ref selected external - 2.5V, switch off, big time constant. RMS value: 1-2 (LSB)

For reference, noise with input pins shorted ~1 LSB.

[Kleinstein]

The noise with a shorted input still makes sense, as one point for the noise testing. This however only applies with a short before the input buffers, not the ADC in unbuffered mode and than an short.
Reference side noise may well be more important with a voltage near the full scale. Still quite some measurements would be done with way less than full scale.

The filters used for the reference and input look like they could effect the accuracy. One point here is dirft in the input current and the other point is input current to the filter capacitors. The capacitors react to temperature and mechanical stress and because of the DA settling can take a long time (like a few hours).  The filter helps with the nois, but will reduce the accuracy.

[MasterT]

Ref. Input curent with buffers on = 100 nA. Since there is constant voltage & sampling clock frequency - no issue is expected with linearity.
Voltage drop accross 10k resistor = 1 mV
Initial accuracy of the internal voltage reference +-0.12% or +-3mV

As you can see accuracy is not sucrificied.
Capacitor leakage current is ~10 nA or 1/10 of the ref. inputs, the only concern its may rise with temp., than active LPF would be necessary.
Caps tested so far:
UKL0J102MPD
RNL1C152MDS1

[Kleinstein]

I see 3 problematic points with the 10 K resistors:
The input current of the ref. buffers has quite some drift:  +-0.25 nA/K typical from the datasheet. With 10 K and a 2,5 V ref. this tanslates to 1 ppm/K of drift. If the buffer is BJT based, there will also be 1/f current noise that can become a problem. The linked article on reference noise showed an low pass filter using the AD797. It works good for the higher frequencies, but at the lower frequencies it gets worse than it was before.

The temperature drift of the capacitor leakage current. 10 nA is quite low, but chances are it is also quite temperature dependent. If it is similar to semiconductor leakace the drift may reach 1 nA/K.

The final problem is the DA if the capacitor. With electrolytic capacitors this can real a few percent and is thus needs quite along time to settle after turn on. To get settling to the ppm level it would need some 1000-10000 RC time constants and with the low cross over frequency this gets really long, possibly more than 1 day. Here it is not so much the resistor itself that is the problem, the point of having a long time constant with an electrolytic capacitor.
It is a bit hard to tell appart the settling of the capacitors leakage current and DA. For the capacitor leakage (maybe DA) it is no do uncommon to take days for settling.

Filtering the low frequency reference noise is a tricky thing. Even film capacitors reactor to thermal disturband and stress / seismic events. A some point the effort for filtering gets larger than simply using a lower noise or 2nd reference chip.

This is anyway a topic seprate from the ADC itself, but more thing for the reference. For testing the ADC it should be good einough to have the same reference for the ADC and test signal. In this case there is not need the low frequency filtering - just the higher frequency filtering can be relevent.

[MasterT]

This is anyway a topic seprate from the ADC itself, but more thing for the reference. For testing the ADC it should be good einough to have the same reference for the ADC and test signal. In this case there is not need the low frequency filtering - just the higher frequency filtering can be relevent.

 It's all about ADC noise tests with respect DC input voltage measurements, when in general case input signal is not correlated to reference voltage in any way. So situation is just opposite, low frequency noise is matter, high frequency has no meaning for SD ADC at all.


Indeed settling time of the filter is quite long, so I came up with simple circuits to disconnect capacitor during power outage..
Recorded voltage lost is 60 mV when circuits was w/o power for ~10 hours.

 

[iMo]

Did you try to run it in the differential input mode?
Like the input 2.5V source floating and the input AIN1 biased to REF+ (for example)?

[MasterT]

Did you try to run it in the differential input mode?
Like the input 2.5V source floating and the input AIN1 biased to REF+ (for example)?

ADC is running in diff. mode, max input range +-2.5V. The purpose of the voltage divider lowering input voltage down to 2V is to avoid saturation.
 I tested 2 cases, divider driven by Ref voltage taken right from pin 4, and driven by anothe max6164-4.096V divided /2.  Of cource, using max6164 shows higher noise level since non-filtered sub 10 mHz range taking power, but not much, instead of 1-2 LSB reads 2-4.
 In essense, if ADC is used "as it is" w/o reference filtering noise is ~10 LSB, see post #1

[Kleinstein]

Ideally an ADC would not react to high frequency reference noise. However the way the SD ADCs work is that they sample the reference at a certain frequence / times  and from this it is normal that a SD ADC reactors also to higher frequency (e.g. MHz range) reference noice. So some filtering on the reference side make absolutely sense. It is quite normal to have filtering, alone to handle the current spikes that most switched capacitor SD ADCs produce at the inputs. So quite some capacitance at the inputs is normal, though usually with less resistance.

For DC measurements filtering is usually not that practical as the times for the measurements are often quite long. For the input signal the settling makes really low frequency analog filtering not practical. Already suppressing the 50/60 Hz mains hum is a bit of a challange and slow settling.

[MasterT]

For DC measurements filtering is usually not that practical as the times for the measurements are often quite long. For the input signal the settling makes really low frequency analog filtering not practical. Already suppressing the 50/60 Hz mains hum is a bit of a challange and slow settling.

You 've missed the point again, settling time is for reference voltage only in question, and with cap always charged - no issue.
 Input filtering is not required, I used filter just to create noise-free source, or I can use 1.5V battery otherways.
More details on the test set-up:
sampling rate - 10 sps, two cahnnales diff. mode, 5 sps per channel. (I used second one to measure/ monitoring drop voltage across 10k resistors)
RMS calculated over 128 samples,  averaging time frame 24 seconds.
LSB ~= 150 nV

[maat]

Would you mind uploading the raw data? Or at least plot a histogram? That would make your contribution even more valuable.

[iMo]

Did you try to run it in the differential input mode?
Like the input 2.5V source floating and the input AIN1 biased to REF+ (for example)?

ADC is running in diff. mode, max input range +-2.5V. The purpose of the voltage divider lowering input voltage down to 2V is to avoid saturation..

I mean the diff "bipolar mode", where the AIN1 is biased to REF+ for example (the virtual ground).
Grounding the AIN1 as you have there works in the unipolar mode only, imho..

[Conrad Hoffman]

I'd be nervous about large capacitors anywhere in a reference circuit. As mentioned, the leakage will change with temperature. They will also pump current in and out with every little temperature change. Just seems like a separate research project would be needed to understand the many ways they can influence the circuit.

[MasterT]

Did you try to run it in the differential input mode?
Like the input 2.5V source floating and the input AIN1 biased to REF+ (for example)?

ADC is running in diff. mode, max input range +-2.5V. The purpose of the voltage divider lowering input voltage down to 2V is to avoid saturation..

I mean the diff "bipolar mode", where the AIN1 is biased to REF+ for example (the virtual ground).
Grounding the AIN1 as you have there works in the unipolar mode only, imho..

I still don't understand, ADC is "fully differential" - means you free to wire up AIN1 to anything, ground, Vref or Vpower. The only limits is 2.5 V at max. There is no difference in noise level vs common voltage, only linearity may be worse, as they say in DS.

[MasterT]

I'd be nervous about large capacitors anywhere in a reference circuit. As mentioned, the leakage will change with temperature. They will also pump current in and out with every little temperature change. Just seems like a separate research project would be needed to understand the many ways they can influence the circuit.

 Let me explain a little bit, seems AD article is  too long to read.

If I try to build a voltmeter based on AD7172-2, than I would look into table table 21 in DS and expect 0.25 uV RMS , right? Wrong.
All those tables were measured with likely inputs shorted, and nobody knows what reference was in use.

ADC IS a multiplicator, so Result = V-in x V-ref -> product by two values, and zeroing one (V-in) is dumb stupid things to do.

Next, if I apply extremely pure DC at the inputs what digits I should expect to flicker on my voltmeter indicator? 0.25 uV rms means 1.5uV p-p, s0 6-digits at the best. For human end user noise is matter as it looks on display, time term is about 1-10 sec, anything slower - is a drift and doesn't bother me much. If I need to measure a voltage I hope to get settled value in a few seconds, or even less than a sec.

I agree that electrolytic is not perfect components, and temperature sensitivity is a weak side, but temperature doesn't change in a seconds - so I underlined ones more - noise/ drift in the 0.1-10 HZ was considered in this research.

Why AD7172? Yes, the problem is not appear today, it's always were , but only with extremely low noise ADC (24-bits 1 LSB noise) this problem sharpen itself to the level, that if not taken care off, it's does compromised ADC noise performanse. See post #1, 10 LSB vs 1 LSB. 
Reference voltage filtering using 1-10 uF cap was sufficient for early SD ADC  > 1 uV rms noise, I'm talking about 1uV with G=1 ( disregard 10nV noise ADC with G=128 as a data manipulation).

[MasterT]

Would you mind uploading the raw data? Or at least plot a histogram? That would make your contribution even more valuable.

1. Internal reference "internaly" mux-ed via reg. 0x20 bits 4:5 = b10

   AD-7172. Wait, please...
   *0   6715222   6715205   6715211   6715209   6715224   6715226   6715231   6715228   6715218   6715226   6715224   6715219   6715235   6715214   6715214   6715225
   16   6715216   6715201   6715205   6715210   6715204   6715212   6715210   6715204   6715206   6715217   6715209   6715215   6715222   6715221   6715213   6715212
   32   6715209   6715203   6715203   6715205   6715204   6715213   6715207   6715205   6715207   6715222   6715208   6715222   6715213   6715201   6715203   6715215
   48   6715210   6715217   6715211   6715224   6715209   6715221   6715214   6715212   6715221   6715229   6715225   6715243   6715237   6715228   6715226   6715224
   64   6715222   6715208   6715214   6715214   6715210   6715205   6715218   6715222   6715229   6715233   6715228   6715217   6715229   6715223   6715230   6715226
   80   6715219   6715213   6715223   6715217   6715213   6715218   6715215   6715221   6715212   6715226   6715218   6715231   6715217   6715221   6715225   6715229
   96   6715214   6715219   6715218   6715222   6715235   6715236   6715226   6715225   6715227   6715230   6715240   6715224   6715225   6715235   6715225   6715208
   112   6715233   6715236   6715237   6715213   6715213   6715210   6715229   6715210   6715235   6715235   6715228   6715210   6715212   6715199   6715201   6715210   
   SNR: 9.909   aver uV: 2001291.059

2. LPF in use, switch on

   AD-7172. Wait, please...
   *0   6715385   6715385   6715383   6715388   6715386   6715386   6715384   6715383   6715381   6715382   6715379   6715377   6715375   6715380   6715377   6715377
   16   6715378   6715377   6715376   6715374   6715376   6715376   6715378   6715376   6715375   6715374   6715375   6715375   6715374   6715375   6715374   6715377
   32   6715379   6715384   6715385   6715384   6715382   6715385   6715384   6715381   6715380   6715379   6715380   6715381   6715381   6715387   6715385   6715385
   48   6715385   6715385   6715386   6715388   6715388   6715387   6715385   6715385   6715385   6715387   6715387   6715385   6715383   6715383   6715383   6715379
   64   6715378   6715379   6715379   6715377   6715377   6715371   6715370   6715367   6715369   6715369   6715371   6715372   6715372   6715372   6715375   6715375
   80   6715379   6715385   6715389   6715391   6715390   6715388   6715387   6715388   6715389   6715389   6715391   6715389   6715388   6715387   6715385   6715385
   96   6715384   6715382   6715381   6715378   6715377   6715374   6715372   6715374   6715372   6715371   6715374   6715374   6715375   6715373   6715374   6715371
   112   6715367   6715366   6715367   6715369   6715370   6715370   6715371   6715373   6715372   6715373   6715374   6715374   6715377   6715382   6715384   6715383   
   SNR: 6.252   aver uV: 2001338.992

3. LPF in use, switch off

   AD-7172. Wait, please...
   *0   6715448   6715446   6715445   6715446   6715447   6715446   6715446   6715445   6715446   6715445   6715445   6715448   6715446   6715447   6715447   6715445
   16   6715448   6715446   6715446   6715446   6715445   6715445   6715446   6715446   6715447   6715447   6715447   6715446   6715446   6715446   6715446   6715448
   32   6715447   6715447   6715446   6715447   6715446   6715447   6715446   6715445   6715448   6715446   6715446   6715447   6715447   6715448   6715447   6715448
   48   6715448   6715449   6715449   6715451   6715447   6715448   6715449   6715448   6715449   6715449   6715447   6715446   6715447   6715447   6715448   6715447
   64   6715447   6715447   6715448   6715448   6715448   6715447   6715447   6715447   6715447   6715444   6715446   6715449   6715447   6715448   6715448   6715449
   80   6715449   6715446   6715449   6715449   6715448   6715448   6715447   6715449   6715448   6715448   6715448   6715448   6715446   6715447   6715447   6715447
   96   6715447   6715445   6715447   6715447   6715446   6715446   6715445   6715445   6715446   6715445   6715446   6715445   6715447   6715446   6715445   6715447
   112   6715447   6715445   6715444   6715447   6715447   6715446   6715445   6715447   6715446   6715447   6715447   6715446   6715444   6715446   6715445   6715445   
   SNR: 1.272   aver uV: 2001359.095
   
4. LPF in use, switch off, Inputs shorted -);

   AD-7172. Wait, please...
   *0   -328   -330   -330   -329   -331   -330   -330   -329   -330   -327   -331   -329   -328   -329   -330   -329
   16   -330   -330   -330   -330   -330   -331   -329   -331   -330   -330   -330   -330   -331   -331   -332   -328
   32   -330   -329   -330   -329   -331   -329   -329   -329   -328   -330   -329   -330   -330   -329   -330   -330
   48   -329   -330   -329   -331   -329   -330   -330   -330   -331   -331   -330   -330   -331   -328   -327   -329
   64   -329   -328   -329   -329   -329   -330   -330   -328   -329   -328   -330   -329   -330   -329   -329   -327
   80   -330   -329   -330   -330   -330   -330   -329   -330   -329   -329   -329   -330   -330   -329   -330   -330
   96   -329   -329   -329   -328   -329   -329   -329   -328   -330   -329   -328   -328   -330   -329   -328   -329
   112   -330   -330   -330   -329   -330   -328   -329   -330   -330   -330   -330   -331   -330   -329   -331   -330   
   SNR: 0.927   aver uV: -98.196
   
   
Math function:

Code: [Select]

void snr_ad7172(double &snr_r0, double &mV0, double &snr_r1, double &mV1)
{ 
  double accum0 = 0.0;
  double accum1 = 0.0;
  double tempd  = INP_BUFF;
  double  tempf  = 0.0;

  for( uint16_t i = 0; i < INP_BUFF; i++ ) {
    inp[0][i] = adc_datas[0][i];
    inp[1][i] = adc_datas[1][i];
    }
 
  int32_t tempr;
  for( uint16_t i = 0; i < INP_BUFF; i++ ) {
    accum0 += inp[0][i];
    accum1 += inp[1][i];
    }
  accum0 /= tempd;
  accum1 /= tempd;
  mV0 = adc24_mV( accum0 );
  mV1 = adc24_mV( accum1 );
 
  double snr0 = 0.0;
  double snr1 = 0.0;
  for( uint16_t i = 0; i < INP_BUFF; i++ ) {
    tempf = inp[0][i] - accum0;
    snr0 += (tempf * tempf);
    tempf = inp[1][i] - accum1;
    snr1 += (tempf * tempf);
    }
  snr0 = sqrt(snr0 /tempd);
  snr1 = sqrt(snr1 /tempd);
 
  snr_r0 = snr0;
  snr_r1 = snr1;
}

double adc24_mV(double raw)
{
  double mVolt = (volts_ref * raw) /8388608.0;
 
  return mVolt;
}



[iMo]

Did you try to run it in the differential input mode?
Like the input 2.5V source floating and the input AIN1 biased to REF+ (for example)?

ADC is running in diff. mode, max input range +-2.5V. The purpose of the voltage divider lowering input voltage down to 2V is to avoid saturation..

I mean the diff "bipolar mode", where the AIN1 is biased to REF+ for example (the virtual ground).
Grounding the AIN1 as you have there works in the unipolar mode only, imho..

I still don't understand, ADC is "fully differential" - means you free to wire up AIN1 to anything, ground, Vref or Vpower. The only limits is 2.5 V at max. There is no difference in noise level vs common voltage, only linearity may be worse, as they say in DS.

The 7172's DS states the input buffer has "true rail to rail" outputs - provided it is true you are free..

PS: btw. - while comparing 7177 and 7172 DSes it seems to me the 7172 is the 7177 but with 8 lowest bits stripped off the 32bit data train (plus some minor differences in the sampling frequencies). For example - the clean "24bit" noise free outputs at lowest sampling frequencies (Fig 11 and 14), with a single histogram bar ..

[Kleinstein]

It may be at first glance or for a novice use that the noise up to some 1-10 seconds is what matters. However for real use the relevant noise also includes longer times, like a few 100 seconds. This is to includes time scales needed to manually connect / swap cables and similar things.
For data-sheets the 0.1-10 Hz range is a defacto standard for the low frequency noise value. However the 0.1 Hz lower frequency limit is more set by the difficulty to get lower frequency analog filtering and not because lower frequency noise does not matter. For a DMM the lower frequencies do matter. It is more that the slightly higher frequencies (e.g. > 1 Hz) are less relevant.

Reference filtering is an old topic. The general conclusion is that it is not feasible to filter away the 1/f noise part well below 1 Hz. A fitler can help with the higher frequencies, but if not carefull it makes things worse below the cut off.
Instead of high effort with large low leakage capacitors it is usually better to get a better reference. The mentioned max6164 is not a suitable choice for the AD7172 ADC, as it is nearly as noisy as the internal reference and also with quite some long term drift. It would take a better reference, more like LTC6655 or max6325 to make full use of the ADC.

For a DMM both the noise with a shorted input and with a DC voltage matters. Having low noise with a shorted input / low voltage still helps as on average on has quite a bit less than full scale voltage. The data-sheets tend to point out the best points, but for the user the worst points may be more relevant, like just a bit over 10% of the range so that one can not yet switch to the next range.

One usually can accept some noise from the reference and still use an ADC with lower noise, as quite often the signal source also has some noise on itself. So for highest resolution one would need to average over some time anyway. This only makes sense if the meter is also stable for more than a few seconds. The faster time scale noise can be averaged out digitally (even in the head), but having added drift and very low frequency noise is bad as settling effects from the signal source (or thermal EMF at connectors) can look similar.
For the higher resolution DMM it is usually accepted to have RMS noise of about 1 digit and this a little movement there.

[iMo]

FYI - While reading the REF5025 DS (used in my 24bit SAR ADC) I've found similar articles:

For additional information about how to minimize noise and maximize performance in mixed-signal applications such as data converters, refer to the series of Analog Applications Journal articles entitled, How a Voltage Reference Affects ADC Performance. This three-part series is available for download from the TI website under three literature numbers: SLYT331, SLYT339, and SLYT355 for Part I, Part II, and Part III, respectively.

[MasterT]

1.

It may be at first glance or for a novice use that the noise up to some 1-10 seconds is what matters. However for real use the relevant noise also includes longer times, like a few 100 seconds. This is to includes time scales needed to manually connect / swap cables and similar things.

2

Reference filtering is an old topic. The general conclusion is that it is not feasible to filter away the 1/f noise part well below 1 Hz. A fitler can help with the higher frequencies, but if not carefull it makes things worse below the cut off.
Instead of high effort with large low leakage capacitors it is usually better to get a better reference. The mentioned max6164 is not a suitable choice for the AD7172 ADC, as it is nearly as noisy as the internal reference and also with quite some long term drift. It would take a better reference, more like LTC6655 or max6325 to make full use of the ADC.

1. Reasobable, to have time frame slightly longer than 10 sec, but cables swapping is not acceptable for me. If I need to track 2-4 sources I'd like to have 4-inputs simultaneously sampling device. I still don't have bench DMM and realy don't like them as all I seen have only one input. Running 2-4 signal in parallel brings time lenght demands down <10 sec again.

2. " is not feasible"   Can't get a difference 10 LSB & 1 LSB in your head ? Retaired or what ?

It;s not about what voltage reference part number is better and what is worse. RC network is always x1000 times lower in noise , x1000.000 times over wideband.   There is 2 MHz clock, so overall noise-accumulating  bandwith up to 10-th MHz. Even unprecedently low noise like 1nV OPA configured as buffer G=1 active LPF  still would output  microVolts of noise.

[Kleinstein]

Some filtering of the reference path absolutely makes sense is standard practice and also shown in the datasheets. This port of the filtering should of cause still be there. The problem is trying to also filter away the low frequency (e.g. 0.1 - 10 Hz) noise. This usually comes with a large resistor and capacitor and this usually leads to worse performance below the filter frequency. The solution with the large capacitor comes with drift and settling issues. The version from the AD article comes with quite some extra 1/f noise.

With a DMM and similar more DC measurements the drift and lower frequencies really matter. It is more like the 0.1 - 10 Hz range that one can work around by a little more averaging.
Multiple channels and simultaneous sampling can in some cases helps, but not always.
I would even consider the part above 1 Hz, where the filter actually works as less important.

Trading in the DC and low frequency ( < 0.1 Hz) performance for reduced noise in the 0.1-10 Hz range is usually a bad choice with a DMM.
A better reference helps in the whole frequency range and often also comes with better drift specs.  It may not be as good as the filter at 10 Hz but the cross over where a better reference can be better than the filter may be at some 1 Hz.  The quality of the reference does make a difference - the whole point of the filter is to reduce reference noise and this is especially an issue if one starts with a rather noisy type like the ADC internal one or the max6164. Already a relatively cheap max6070 is expected to be 3 x lower noise - not just in the 0.1-10 Hz band but also below.

The test with 128 samples is just in a way to test the range where the filter helps and ignore the low frequencies where it makes things worse.

Using the same reference for the input signal and the ADC reference normall eliminats much of the reference noise.  It only works good with the same level of filtering on both sides - a more moderate fitler would work just as well. Comparing the case with no filter at the refernce side to the case of the same type of filter for the ref. and input is thus not fair. With even more filtering at the reference it would get worse again.

[MasterT]

The 7172's DS states the input buffer has "true rail to rail" outputs - provided it is true you are free..

I see. Yes, buffers CM includes Gnd & Vdd.  Ether internaly generated -Vcc or buffers is "pre-chargers" and complete voltage sensing cycles includes direct switching - > by pass.
Buffers at the reference inputs is  important in this topic, since non-linearity is un-avoidable if R in series with Ref.

There a few excelences in this ADC (I'm not affiliated in any way and not advertising any product):
- byffers;
- noise-free channels switching, there is even "delay" option to sample glitch-free. I tried same things with mcp3562 and was disapointed with high noise if two diff channels in use

 The only things I don't like, they wire Ready/ Dout on the same pin, creating complication with common SPI

FYI - While reading the REF5025 DS (used in my 24bit SAR ADC) I've found similar articles:

For additional information about how to minimize noise and maximize performance in mixed-signal applications such as data converters, refer to the series of Analog Applications Journal articles entitled, How a Voltage Reference Affects ADC Performance. This three-part series is available for download from the TI website under three literature numbers: SLYT331, SLYT339, and SLYT355 for Part I, Part II, and Part III, respectively.

  Sooner or later it comes to understanding that voltage reference filtering is a must. The problem here, as I already mention, common OPA outputs uVolts - over MHz bandwidth, and to get down to 150nV is AZ OPA configureed with RC network like in a picture or "some tricks" in the ADC at the reference inputs, pre-chargers /AZ etc.

[David Hess]

I agree with Kleinstein.  Filtering low frequency noise, especially flicker noise, is a fool's errand and without heroic efforts will introduce more low frequency noise, at least in the form of drift.  It is better to use a lower noise reference, or multiple references in parallel, to reduce low frequency noise from the source.

I settle for the 0.1 to 10 Hz noise measurement because it correlates well with lower frequency measurements and it is much easier to get good results.

Ideally an ADC would not react to high frequency reference noise. However the way the SD ADCs work is that they sample the reference at a certain frequence / times  and from this it is normal that a SD ADC reactors also to higher frequency (e.g. MHz range) reference noice. So some filtering on the reference side make absolutely sense.

I have assumed, and it has been my experience, that high frequency noise close to the Nyquest frequency will alias to low frequencies compromising low frequency noise, but with the high sampling frequency of a delta-sigma converter, this high frequency noise is easy to remove with moderate filtering and without adding to the low frequency noise.

[David Hess]

Sooner or later it comes to understanding that voltage reference filtering is a must. The problem here, as I already mention, common OPA outputs uVolts - over MHz bandwidth, and to get down to 150nV is AZ OPA configureed with RC network like in a picture or "some tricks" in the ADC at the reference inputs, pre-chargers /AZ etc.

That is the circuit I like to use, and not just for reference filtering, because it provides the lowest AC output impedance.

There are some modern precision operational amplifiers which support driving the output capacitor directly, which might be as good or better.

[MasterT]

Sooner or later it comes to understanding that voltage reference filtering is a must. The problem here, as I already mention, common OPA outputs uVolts - over MHz bandwidth, and to get down to 150nV is AZ OPA configureed with RC network like in a picture or "some tricks" in the ADC at the reference inputs, pre-chargers /AZ etc.

That is the circuit I like to use, and not just for reference filtering, because it provides the lowest AC output impedance.

There are some modern precision operational amplifiers which support driving the output capacitor directly, which might be as good or better.

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
 

[KT88]

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.
The reference driver and reference cap also don't work as a filter - the cap is a reservoir for charge. The sample action takes some charge out of the ref-cap and the buffer recharges the cap before the next sample is converted. The cap size depends on the allowed error that can be introduced - a rule of thumb is to allow for a voltage drop of 1/2LSB - but that depends, of course...
A similar kind of amplifier can be found in more recent LDOs that are specified to operate with larger ceramic caps.