MAX5717 16-bits DAC reference voltage problems.

[MasterT]

Data sheet for DAC https://datasheets.maximintegrated.com/en/ds/MAX5717-MAX5719.pdf
page 15 shows buffering of the reference voltage by precision OPA MAX44246, loaded by 100nF cap.
Data sheet of the OPA https://datasheets.maximintegrated.com/en/ds/MAX44241-MAX44246.pdf
page 7 chart STABILITY vs. CAPACITIVE AND RESISTIVE LOAD IN PARALLEL  says that 4k (DAC input reference impedance) and 100 nF in parallel is unconditionally unstable circuits.

Is this a "copy paste" error in doc for DAC?
 
What options to buffer reference input, as it seems to me hell of requirements/ troubles load:
a). precision better than 64 uV;
b). impedance IS changing with a code, varying  about 4k to 10k - non linearly, with min resistance somewhere in the middle;
c); band width of the reference 1 MHz or better.

Precision OPA solves a, but my test shows that MCP6V27 ( what I have in stock) miserably failed  on b, probably due to relatively high output impedance, I see variation in voltage about 1 mV - 16 LSB for DAC with 1 bit INL - unacceptable.
Probably, the way to go is to buy MAX6126 - page 16 on the dac DS. But it is solving only 1/2 part, second part is DAC output buffer, and it's again has to be a/b/ and /c. 
I can try to "double buffer" in-ref and out by low output LME49721 ( 10 mOHms according to DS) in conjunction with MCP6V27
http://ww1.microchip.com/downloads/en/DeviceDoc/25007B.pdf  (page 2 - correcting offset, but question is it the only option?
 

[Kleinstein]

A compound driver, with a fast low impedance output driver with an AZ OP for offset correction is likely the best option. However the circuit from the MCP6V27 data-sheet looks really odd. I really doubt it could work - its more like someone mixed up the circuits. The extra driver could be something like a transistor circuit.

It likely needs something like an additional RC in parallel to the 100 nF cap.

[David Hess]

Any time you want precision with a variable resistance load, a separate buffer and error amplifier should be used.  Otherwise thermal feedback from the output transistors to the input transistors limits precision.  The output resistance of the error amplifier is only part of the problem.  For the same reason, dual and quad parts should not be used where this is a consideration.

Precision operational amplifiers are designed using a symmetrical layout to reduce thermal feedback but every little bit helps.

[MasterT]

It likely needs something like an additional RC in parallel to the 100 nF cap.

To correct stability issue in maxim-ic MAX5717 data sheet, there has to be a resistor in series with output, 10-100 OHm, than close loop from REFS pin back to OPA inverting input would keep precision in DC- low freq, and cap would lower impedance in mid-high freq. range. Chart STABILITY vs. CAPACITIVE LOAD AND SERIES ISOLATION RESISTANCE on p. 7 for OPA max44246

Same trick doesn't apply for output buffering, and now I'm thinking that double AZ + High power fast OPA is the right circuits, with one exception,  fast OPA can't be R2R. All of them have a voltage offset issue in the middle area, where n and p channels MOS switched. And AZ is too slow to do this 500k - 1MHz correction on the fly, though old BJT like ne5532 or even LM358 would be better than modern R2R IC.

[David Hess]

Precision OPA solves a, but my test shows that MCP6V27 ( what I have in stock) miserably failed  on b, probably due to relatively high output impedance, I see variation in voltage about 1 mV - 16 LSB for DAC with 1 bit INL - unacceptable.

Output resistance of the MCP6V27 is from 30 to 50 ohms depending on supply voltage according to figure 2-19.

I don't know why the variation is that high and it is specified to work with 10k loads.  High frequency performance would be a problem but not in DC applications.  Maybe it was oscillating?

[MasterT]

Precision OPA solves a, but my test shows that MCP6V27 ( what I have in stock) miserably failed  on b, probably due to relatively high output impedance, I see variation in voltage about 1 mV - 16 LSB for DAC with 1 bit INL - unacceptable.

Output resistance of the MCP6V27 is from 30 to 50 ohms depending on supply voltage according to figure 2-19.

I don't know why the variation is that high and it is specified to work with 10k loads.  High frequency performance would be a problem but not in DC applications.  Maybe it was oscillating?

Hmm, thanks for pointing out, I didn't realize I could calculate output impedance. I've seen figure 2-31/32, but it doesn't show impedance at DC. Actually, pocking on breadboard I discovered that  MCP6V27 is not so bad with output impedance, and it's definitely below 0.1 OHm for pure DC. The issue I observed (1-2 mV voltage dropping) was related to my attempts to filter out input with 15 MOHs resistor and 2 uF ceramic cap. I was pretty sure mega-ohms range is o'k, since DS states inputs 10^13 in common and diff mode. It was overestimation. Setting jumper instead of resistor makes output steady down to a few ppm. Measure changes in output voltage I noticed ~30 mV difference if I set a resistor on the input. Its like not 10^13 but rather 10^9 on inputs, and what more interesting any changes in output current  somehow get translated to the inputs. So now I know that filtering with big RC-constant  was not such a good  idea, better to keep R in kOHm range.

[David Hess]

Hmm, thanks for pointing out, I didn't realize I could calculate output impedance. I've seen figure 2-31/32, but it doesn't show impedance at DC. Actually, pocking on breadboard I discovered that  MCP6V27 is not so bad with output impedance, and it's definitely below 0.1 OHm for pure DC.

It should be below 50 micro-ohms in a voltage follower application.

The issue I observed (1-2 mV voltage dropping) was related to my attempts to filter out input with 15 MOHs resistor and 2 uF ceramic cap. I was pretty sure mega-ohms range is o'k, since DS states inputs 10^13 in common and diff mode. It was overestimation. Setting jumper instead of resistor makes output steady down to a few ppm. Measure changes in output voltage I noticed ~30 mV difference if I set a resistor on the input. Its like not 10^13 but rather 10^9 on inputs, and what more interesting any changes in output current  somehow get translated to the inputs. So now I know that filtering with big RC-constant  was not such a good  idea, better to keep R in kOHm range.

Filtering a reference is really difficult to do even with a high input impedance buffer and chopper stabilized amplifiers have a lot of current noise from charge injection which is not reflected in their input bias current specification.  They also may have a large change in input bias current over temperature and common mode range.

Large value ceramic capacitors are *not* low leakage.  In practice two capacitors are used in series with their junction bootstrapped to the reference voltage to reduce leakage.  Check out figure 21 on page 14 of this application note:

https://www.intersil.com/content/dam/Intersil/documents/an17/an177.pdf

[MasterT]

Could be 50 mOHm, application is non-inverting voltage amplifier with 1.6x gain, I'm making adjustment 2.5V up to 4.0V.
The issue is that 50 milliOhms become >1 OHm if there is a resistor 15 MOHm on positive input. And its not reflected anywhere in DS, that voltage source output impedance seen by OPA on non-inverting input could drastically  affect output impedance of the precision OPA itself. I guess, that chopper capacitors  involved into reflection output current of the OPA back to the input offset voltage, but its just a speculation.

Leakage current of the ceramics cap is negligible, pity their capacitance is low. But for now I 'd avoid electrolytic at any cost, knowing its bad noise reputation.  16-bits is a funny things -);

[Kleinstein]

The AZ OPs can have a significant input offset current of a few 10s of pA. Especially the cheap ones are not that much checked.
The best choice for a filter cap is usually a mylar type film cap. One can get them up to a few µF at still reasonable price.

Some of the AZ OPs are actually relatively good when it comes to output impedance (open circuit), as the internally may have an output stage with local feedback. Still it is not a good idea to drive much current. Due to the very high gain the DC output resistance in the closed circuit should be very low (something like open look impedance divided by loop gain) and thus something in the µOhms (maybe even lower) range. It is just that the output impedance is more like an inductance and is thus going up with frequency and too much capacitive load can make it unstable with a kind of resonance in the impedance curve. It might help to also have a kind of RC snubber at the output of the OP to dampen that resonance.

[MasterT]

I should correct my previous post,  DataSheet for MSP6V27 does mention "bias current effects", on p.24,

4.3.4
SOURCE RESISTANCES
////
Make the resistances seen by the inputs small and
equal. This minimizes the output offset caused by the
input bias currents.

Some kind of small print at the back.  I didn't pay much attention for keeping both inputs seen same impedance, and small resistance. IMHO, this should be stated on a front page, due to the great impact on all parameters of performance. Simply mistakenly setting one resistor on input makes precise OPA is worst on a planet, with offset 30 mV instead of 2 uV, bias current nA instead of pA, and with all other parameters distorted in an unpredictable way.

So I moved on, use 1k at input, 100 OHm + 100 nF in series on the output and lock loop-back from sense reference pin of the DAC back to OPA. Like they recommended in DAC DS, erroneously omitting resistor. All works well, I can see using ADS1232 24-bits sigma-delta that voltage variation at the OPA output 100 mV-160mV, and stable down to +- 19-th bit at the reference pin. Unfortunately ADC is slow, so my test is good for 10-80 sample/second, can't verify if its same at higher sampling rate.

INL calculated by arduino Mega2560 ( linear LMS regression) shows I get +-1 LSB on 16 bit scale over full 4V range, and ENOB about 15.5 bits.

[Kleinstein]

With the Microchip AZ OPs you have to be careful abut how they use the bias current - in there interpretation the bias current is the average current of both inputs. However in AZ OPs the input currents of both inputs are about opposite sign. So the sum of the input currents and thus the Microchip Version of Bias current is small.  However the Offset current (difference in input currents) is not that small.

With normal, especially BJT based OPs it can help to have a balances input impedance, so that voltage drop will compensate and only the usually much smaller input offset is relevant.

However essentially all AZ OPs are different: the offset current are usually larger than the bias. So it does not help to have the input impedance balances. The extra resistor only makes things worse.