10 V reference based on LM399, LT1001, and LT3042 in bootstrap configuration

[iMo]

The CA3140 was just an example - a vintage opamp with pA input current, no proposal to try with it..

But imagine you do not have any DMMs. You just want to create a 399 based 10V reference with lowest noise generally. So, the analog filtering could come handy. All you have is R and C and modern opamps.

Too large R creates a noise and a drift, too large C creates leakage and perhaps its dependence on T and perhaps noise as well, opamp with larger input currents, offsets, noises, and TC will not help either.

Thus my initial question above was what are actually the R/C/opamp limits provided we talk exactly the 399 (we all posses the 399 in quantities). I asked that because there is a decade of discussion about that filter and the 399, but the limits (towards lowest freqs) were not recognized/measured/validated, imho.

[David Hess]

the TC of the opamp's input bias should be less than say 1pA/C as that would result in 1uV/C drift which is 0.14ppm/C at the 10V output then (compared with 0.5ppm/C at 7V with the 399).

JFET and MOSFET input operational amplifiers almost always have an input bias current which doubles every 10C, because of either the JFETs or protection diodes.  Bipolar input parts are much more consistent as temperature rises despite higher input bias current.

It gets nearly impossible to filter the 1-120 mHz part.  Only the 2.5 Hz range is somewhat accessible for filtering and here it makes little difference if the cross over is at some 1 Hz or 0.1 Hz. Due to the 1/f noise type the band (some 0.1 Hz width) around 2.5 Hz only makes up a small part of the total noise anyway. So even this filtering part is limited and may not be worth it, if it comes with more noise below 0.1 Hz (e.g. from thermal fluctuations).

That is my experience; low frequency filtering to remove flicker noise is almost always useless because it so easily increases low frequency drift.

Instead of large effort in filtering there is the alternative to use a better reference like ADR1399 that also helps with the really low frequency noise.

Or use multiple references to contribute to the output.

The CA3140 would make no real sense, as it has way to much 1/f noise. It can be OK at 1 kHz but is probably way more noisy then the LM399 at 10 Hz and below.

The CA3140 was just an example - a vintage opamp with pA input current, no proposal to try with it..

Usually the noise from even the best reference dominates over the noise from the operational amplifier, but those old operational amplifiers built on high voltage CMOS processes were very noisy.  I wonder how much better the modern CA3140 is than the original one on RCA's high voltage CMOS process.

There are probably better parts now, but I used the CMOS LMC6081 when I needed low input bias current and precision, or sometimes the bipolar LM11.  The LMC6081 has such a low input bias current to start with, that variation at room temperature can be ignored.  The LM11 is no longer produced, but the bipolar LT1012/LT1097 is similar with better precision and is a good choice for low frequency precision filtering.

Thus my initial question above was what are actually the R/C/opamp limits provided we talk exactly the 399 (we all posses the 399 in quantities). I asked that because there is a decade of discussion about that filter and the 399, but the limits (towards lowest freqs) were not recognized/measured/validated, imho.

I never found an answer to that question either.  I suspect everybody who investigates it reaches the conclusion that I did; it is somewhere between difficult to impossible to filter low frequency noise without adding even more low frequency noise in the form of drift.  In my designs, I settled for a "best effort", and if I needed better performance, either combined multiple references or changed the design to take advantage of something like chopping if possible.

[Kleinstein]

The leakage is not a factor limiting how large one can make the capacitor. With a larger capacitor one can allway reduce the resistor. There is also the option with 2 capacitors to reduce the voltage seen by the one towards the signal to get a low voltage. This could also be a mix of a film capacitor and electrolytic for the auxiliary one. It is size and costs that limit the capacitance.

Film capacitors of some 10-22 µF are still relative afordable. With little need for RC much beyound some 100-200 ms there is no real need for increasing the resistor much beyond some 100 K.
If space is not an issue there are motor run capacitors as a sometimes relatively cheap version - rather bulky, but other wise good capacitors (polypropylene) up to some 50 µF.

But imagine you do not have any DMMs. You just want to create a 399 based 10V reference with lowest noise generally. So, the analog filtering could come handy. All you have is R and C and modern opamps.

The actual frequencies that matter depend on the application. In many cases the way one looks at the refernce, directly or inderirectly already filters out some of the frequencies. Lowest noise in general makes little sense as a target for a filter, as the filter can reduce the higher fequency noise, but at the same time will increase the lower frequency noise. So how much filter makes sense depends very much  on the use case.

[David Hess]

The leakage is not a factor limiting how large one can make the capacitor. With a larger capacitor one can allway reduce the resistor. There is also the option with 2 capacitors to reduce the voltage seen by the one towards the signal to get a low voltage. This could also be a mix of a film capacitor and electrolytic for the auxiliary one. It is size and costs that limit the capacitance.

Film capacitors of some 10-22 µF are still relative afordable. With little need for RC much beyound some 100-200 ms there is no real need for increasing the resistor much beyond some 100 K.
If space is not an issue there are motor run capacitors as a sometimes relatively cheap version - rather bulky, but other wise good capacitors (polypropylene) up to some 50 µF.

I had trouble finding suitable capacitors because datasheets do not cover leakage to a sufficient level, so I ran my own leakage tests on different series of capacitors in an environmental chamber at elevated temperature.  I ended up using a 0.47 microfarad 50 volt polypropylene capacitor from Panasonic which was a compromise on physical size.

But imagine you do not have any DMMs. You just want to create a 399 based 10V reference with lowest noise generally. So, the analog filtering could come handy. All you have is R and C and modern opamps.

The actual frequencies that matter depend on the application. In many cases the way one looks at the refernce, directly or inderirectly already filters out some of the frequencies. Lowest noise in general makes little sense as a target for a filter, as the filter can reduce the higher fequency noise, but at the same time will increase the lower frequency noise. So how much filter makes sense depends very much  on the use case.

For practical reasons the 0.1 to 10 Hz test is about the easiest, but showing the benefits of chopper stabilization will require going down to 0.01 Hz.

[Kleinstein]

The 0.1 - 10 Hz frequency band is easy to test and for this reason the defacto standard in data sheets.  The frequencies that actually matter are often lower, but here testing gets more involed as AC coupling is no longer easy.

For a filter at a voltage reference one should get away with a polyest capacitor too, as one usually has to wait for warm-up and settling of the reference anyway. The main downside of polyester capacitors is the larger DA and thus somewhat slower settling (may need a few minutes) - this should not be an issue here. The MKS type is reasonable size and price even at 5-10 µF.

For filtering a LM399 ref. at my ADC I have 5 K, 6.8 µF and an additional capacitance multiplication of about 3 fold - so a 10 ms time constant. I might opt. for a little larger resisistor, but not much. The initial idea was to also suppress noise at around 25 Hz, but I no longer need this and got a different way around it.

[David Hess]

There are probably better parts now, but I used the CMOS LMC6081 when I needed low input bias current and precision, or sometimes the bipolar LM11.  The LMC6081 has such a low input bias current to start with, that variation at room temperature can be ignored.  The LM11 is no longer produced, but the bipolar LT1012/LT1097 is similar with better precision and is a good choice for low frequency precision filtering.

I will add that since the AC output impedance of the filter is low, there might be a suitable low input bias current and low noise chopper stabilized operational amplifier.  This would have the advantage of removing the flicker noise contribution of the operational amplifier, however noise from the reference dominates even with a noisier operational amplifier.

[iMo]

This is with ADA4523 chopper. With DC  offset it does aprox -1.1ppm/100kOhm Rfilt (it does fit the 100pA inp current as per DS).

[Kleinstein]

The ADA4523 is not really suited for a high impedance source. The ADA4523 is more like a corner case for very low sources impedance and in my opinion to high in power (and thus possible thermal issues) to be really usefull even then.

I would more consider LTC2057, MCP6V51 or maybe OPA186 or AD8639 - so also looking more for a low current noise and not too much input bias.
As the voltage is fixed one could as well use a 5 V supply type with many more to choose from.

[iMo]

Out of curiosity - the 2057..

[David Hess]

So in the LT1012 case, we see the flicker noise of the operational amplifier, and in the chopper stabilized cases, we see the flicker noise from the reference.

The LT1012 results do not look right to me.  With that noise density graph, which seems high, wouldn't that be 980nV RMS instead of 98nV RMS?  The datasheet says 500nV RMS.

Why is the flicker noise for the ADA4523 simulation so high?

I would try to verify the simulation results with a test circuit before I relied on them in any case.

[iMo]

The LTSpice calculates the total RMS value from F1 to F2 as set in the "noise" directive. Therefore from 0.1-10Hz as above.
The data coming off the noise analysis are subject to the correctness of the noise parameters set in the opapm model, of course.

PS: the sim shows following rms with the LT1012 (0.1-10Hz, opamp with 17nV/rtHz and 20fA/rtHz at 10Hz) with the inp filter as above:

10Meg   98nV
1Meg    100nV
100k     104nV

[Kleinstein]

Why is the flicker noise for the ADA4523 simulation so high?

For much of the simulation the source impedance is set by the capacitor. So the 1/f noise part is not classical ficker, but a white current noise with a capacitive input impedance.
The higher current noise is a problem with some of the low voltage noise chopper amplifiers.
The specs for the current noise look false for quite a few and I would not really trust the datasheets in this respect. Even with the classical OP-amps they don't give the full picture, as there are 2 inputs and partial correlation. So there is more than just 1 current noise part.

[David Hess]

The datasheet RMS noise specification comes from the noise test circuit that Linear Technology published.  That number should be higher than the SPICE result because of the shape factor of the real world filter, but it will not be 5 times higher.

To check the simulation and model accuracy, I would simulate Linear Technology's published noise measurement circuit which is shown below from the LT1008 datasheet.  The LT1008 is an LT1012 with external frequency compensation, but the test circuit works for practically any operational amplifier.  Current noise is measured by adding resistance in series with the inputs and subtracting the voltage noise measurement.

I always liked that Linear Technology published their test circuits for measuring parameters.  Many times I verified the values given in their datasheets when performance was limited by the operational amplifiers themselves.

[Kleinstein]

Current noise is measured by adding resistance in series with the inputs and subtracting the voltage noise measurement.

I always liked that Linear Technology published their test circuits for measuring parameters.

Having the test circuit is really good - for the current noise test with the LT1007 and quite a few others they have resistors on both inputs and assume correlated current noise. If the current noise is correlated depends on the internal circuit. In case of not correlated current noise the test circuit shown in the LT1007 datasheet (also OP27) the actual current noise is 1.4 times higher.

With some AZ amplifier the current noise specs are even worse - some seem to give the shot noise corresponding to the bias current, which is pretty useless and only a lower bond.

[iMo]

..
..To check the simulation and model accuracy, I would simulate Linear Technology's published noise measurement circuit which is shown below from the LT1008 datasheet..

For example..
PS: with V3 in series with the R1 I get the same..

..The datasheet RMS noise specification comes from the noise test circuit that Linear Technology published.  That number should be higher than the SPICE result because of the shape factor of the real world filter, but it will not be 5 times higher..

BTW, the 0.1-10Hz noise in the LT1012's DS is 0.5uV PEAK_PEAK, we have the "Total RMS" in our LTSpice simulations..

[David Hess]

With some AZ amplifier the current noise specs are even worse - some seem to give the shot noise corresponding to the bias current, which is pretty useless and only a lower bond.

I have seen that mistake with linear operational amplifiers that have input bias current cancellation.

Many years ago I made measurements of some improved 741 type operational amplifiers and the input current noise followed the input bias current exactly like it should.  This is very handy when input current noise specifications or input bias current specifications are missing, but of course it does not work when input bias current cancellation is present.

BTW, the 0.1-10Hz noise in the LT1012's DS is 0.5uV PEAK_PEAK, we have the "Total RMS" in our LTSpice simulations..

Ah, ok, I confused the two so your results are right on.  I think the printing on datasheets is getting smaller every year.

[ejd.pol]

Hello,

Here is another update on this DIY voltage reference. Accommodating all changes (with many thanks for the kind suggestions!)
required a new build. I did not design a real PCB yet, as the new version may need some more changes,
and some parts of the new design were just not settled enough.

The part that got the most attention was the resistor divider. I had some decent (not top, but decent) precision resistors
that I had bought on some previous occasion. It was a mixed bag of two dozen or so single values.
Given that the LM199 does not give an accurate output voltage (stable yes, accurate not),
the actual values of these resistors did not matter very much. I chose them for their tempco and stability.
Just as an example: the set contained one resistor of 62.5 Ohm and one resistor of 62.6 Ohm.

As the LM199 gave a measured output of just over 7 Volt, I wanted to make a 3:7 resistor divider.
So I wrote a program that calculated all kinds of compositions of two or three resistors from the set.
After some selection stages, I ended up with two sets composed of three resistors connected in series,
that had a very good ratio, and whose sum was close enough to 10 K (R5 through R10 in the schematic).

After completing the build on another prototype board (with another helper board to emulate the Kelvin connections),
I measured the output being 20 mV over 10 V. With some trial and error, I added a 120K resistor in parallel
to the 7K compound resistor, and that brought the error down to just 1.3 mV over 10V.
That was the limit of what I could do with my DMM at home, and that is what is shown in the schematics.

I completed the buid with a repurposed metal box (it used to be a video switch) that miraculously had the right holes
on the front side for an on/off switch and two output connectors. I added two 9V batteries in series,
and thought about how to make a stable supply voltage somewhere between 12 and 15 Volt.
As mentioned previously, I had an LT1129 module that seemed to fit nicely, and so it was chosen.
The output voltage of the module was originally 10.2 Volts, so I increased it to 13 Volts.
The resulting build can be seen in the attached pictures.

The schematics show the build as measured on a Keithley DMM6500 and an Agilent 34401A.
The two meters do show some voltage differences, but within their specified accuracy.
Using these meters, again with some trial and error, I added a 6 MOhm resistor in parallel
to the already present 7K and 120K resistors (three 2M resistors in series). This increased accuracy
by another factor of 6, bringing down the output voltage to a comfortable 0.2 mV above 10 Volts.

So the end result is that I now have an accurate reference that I can use at home to calibrate my meters.
The resulting design does not use a trim pot, but relies only on a hard trimmed resistor divider.
The bootstrap configuration gives a very stable output, which can easily be loaded with 10s of mA.
Currently it still lives on a prototype board, and for now that is fine for me.
I do wonder how much the reference will drift over longer periods of time.
It is turned off most of the time, so drift will (hopefully) be slow.

Again many thanks for all helpful comments and suggestions!