LM399 based 10 V reference

[dietert1]

It's a small raster board. I wanted something low noise for characterizing a JVR. May serve as an exercise for making a 5x LTZ1000 board if the LM399s are still to noisy. Or i can draw a board and make some more of them. If i remember right one LTZ1000 is about the price of 5x LM399s,

Regards, Dieter

PS: After another 24 hours the log contains 30 000x 100PLC measurements.

[Kleinstein]

The LM399 can be relatively noisy, though the noise level can vary quite a bit be tween units.  Not sure how typical the 2 units are that I have tested.

Attached is a reading of a LM399 relative to another LM399 at the ADC. So the curve is essentially the noise of 2 x LM399 as a ratio or difference. For comparison the blue curve is a JRV type reference measured in fast sequence (1 PLC with JRV, 1 PLC LM399A, 1 PLC zero, 1 PLC ADC ref=LM399) the points are average over 20 such cycles.  One can see quite a bit of popcon noise for the green curve (2xLM399) and much less for the blue one (LM399+JRV).  So the large jumps all come from one of the LM399 (the other one still shows a few similar ones, but very rare)  The JRV reference is quite a lower noise, but more drift (likely from temperature).

So to test the JRV (2N4391 + resistor+ bootstrapped drainvoltage), the LM399 is OK for temperature drift and may be good enough for the long term drift (the data from Alex suggest very low dirft, buit this was just 1 unit), but not good enough for short time ( < 1 hour) noise. Even 5 x LM399 would still be relatively high noise.

[dietert1]

Yes, appears we had the same idea: For the beginning some LM399 should be good enough to study the JVR. In your diagram the LM399 comparison shows about 2 uV p2p noise aside from the jumps. That would mean a LM399 could be 1.4 uV p2p and the average of 5x LM399 would be about 0.6 uV p2p.
Except i still have to make a good 7 -> 10 V gain stage for the LM399 board and some OPA189 preamplifier to get better resolution. In fact i wanted to make more preamplifiers with a multiplexer to get high resolution differences between the LM399s as well.

In this thread i just wanted to provide my numbers about initial drift of LM399.

Regards, Dieter

[iMo]

Well, how to discipline the jvr reference with help of the lm399?
Something like "399DJVR" ultimate reference?
The relationship between a GPS and an OCXO is almost the same as the relationship of the lm399 to the jvr..

[Kleinstein]

There is a similarity in the stability on different time scales. However I very much hope the drift seen with the JRV ref. is mainly temperature. So with an oven to stabilize the temperature, there is the hope that one does not need another ref. for the long term stability. This is a bit off topic here and more goes to the JRV thread.

The main point in the graph is that the noise of the LM399 can be quite different between units. The 2 LM399 involved both show jumps of some 4 µV, but one unit with about 1 jump per minute and the other unit with about 1 jump in 1-2 hours (no jump visible in the graph, but a few in other similar curves).  It is still a question which of the 2 refs is better:  The more jumpy one looks more noisy, but after averaging over a longer time this would average out, or if one could identify the levls would even takt the 2 levels separate. At least for this sample there were only the 2 levels visible, no obvious 3rd level. The one with only few jumps still has the 4 µV uncertainty, with sometimes very long between jumps, so not much averaging in a reasonable time.

There are a few DMM (e.g. Keithley 2010, 2182, 2001) that use 2 references: a LM399 for the long term stability and a likely lower noise and higher current zener for the shorter time directly at the ADC. The corss over is digital and may be mainly because of ADC gain, not referece noice. It is in software, with not many details known.

It could be rather difficult to do the corss over / combination in the pure analog domain - this would be similar to low pass filtering and this is not easy and could get away without the 2nd ref.. Digital filtering would need a 2 nd ref. and may be possible also for a longer time constant (e.g. hours). I would build it in a way to correct the long time drift of the drifty ref. by looking at the amplied difference. So the ADC / DAC resolution would not need to be very high, if the 2 refs are similar voltage. 

[qosch]

After reading in this thread a few evenings I now registered because I still have a few questions.
First I am interested what people refer to when talking of the "Branadic PCB". I've read plenty of very informative posts by Branadic but could not find schematics or a layout posted by him.
Also I have an idea that I am surprised I have not read about yet: putting two LM399 in series instead of in parallel. Maybe I am missing something obvious but to me it seems that this solution is simpler because it just needs one buffer instead of N+1 (N=number of LM399), the relative noise and offset of the output buffer is lower (because the signal is larger) and a DAC could be connected to generate 10V without any resistive divider (that adds more drift and noise). Of course the supply voltages would need to be higher and the current source becomes more critical because it is a source of correlated noise and drift.

[Andreas]

Hello,

if its in the context of the 10V PWM DAC the cirquit is here:
https://www.eevblog.com/forum/metrology/lm399-based-10-v-reference/msg2082496/#msg2082496

Of course you can series connect (to a limited number of zeners) the LM399 zeners.

There are 2 constraints since the heater and the zener are not independant of each other.
- the maximum voltage rating between the pins
- you should never forward bias the substrate diodes.

But many DMMs have their best accuracy range around 10V and only some up to 20V.
So a higher reference voltage will be of limited use.

with best regards

Andreas

[Kleinstein]

Can use a series connection if the heater voltage is in the right range with 2 x lm399 this is usually OK.

However scaling from 14 to 10 V is about as difficult as scaling from 7 V to 10 V. It would be only with 3 in series that going from 21 V to 10 V may be a little easier in some ways or assuming an advantage from 2 identical resistor.

A DAC (e.g. PWM type) can be in the forward path (than 14 V is the better starting point), or in the fedback path (than 7 V is the better starting point). So it depends.

Noise wise the LM399 is normally quite a bit higher noise than a low cost precision OP like OP07. So there is not much difference between series and parallel connection.

[r6502]

Hello all,

I just designed a adjustable 10V reference as a replacement for the reference circuit inside of my "DIAL - A - SOURCE" "model "DAS-56A", based on a LM399H reference and some Vishay precision resistors. The DAS-56A is a calibrator is based on a 6 digit KV divider, that is supplied by a 10V reference voltage.

The original reference circuit of the DAS-56tA is based on 2 zener reference diodes, that have normally around 6.2V each. They are in series and the 10V ouput voltage is set by several precision resistors and a trim pot to 10V. This 10V is used as input voltage for the KV divider.

In my device one of the reference diodes has now only around 5V, so there is the need to do something. Instead of replacing the defect z-diode, i decided to build a new 10V reference circuit based on the LM399H. The resistors have been chosen based on availability at the distributors (actual: Mouser and Bürkin).
 
So I saw this thread, and looked a bit, not at all posts but at a lot and that I've drawn 1st the schematic and than the PCB, that should fit in the original shielded compartment for the reference voltage circuit. The dimensions of the original PCB are about 43.2mm x 40.6mm. I increased the size a bit to 45mm x 45mm.

I'm not 100% sure, what OP to use. In the schematic I pointed out to an good old LT1001 that will do the job I think.

There is also a thread for the DAS56A:
https://www.eevblog.com/forum/metrology/general-resistance-das56a/


Please take a look  and place your comments on my post. I atached Photos and the schematic ...



 

[Kleinstein]

Chances are the LM399 may be a little more noisy than the 2 old zeners. However the actual noise varies a lot between individual units.  Reference zener diodes like 1N827 are still available (though relatively expensive) - easier to get than a LM399 these days.

The LT1001 would be good enough, so would be the good old OP07.

The footprint for the resistors looks a little odd. The trim circuit can be build in a way that only 2 really critical resistors are needed - the other resistor (e.g. those to set the zener current) are far less critical. Different from the old zeners there is no need to accurately trim the reference current to a specific value.  So a single resistor should be good enough.

The vertical mounting of the resistors like on the old board is normally not a good idea for a precision circuit: The ends of the resistors see different temperatures and one can thus get some thermal EMF. In addition the distance from the solder joint to the resistor is quite short at one end.

Depending on the connections to the rest of the circuit, it may be a good idea to have a separate ground connection for the LM399 heater, as the current can vary.

The capacitor directly in parallel to the zener output is a little dubious. It is sometimes seen in other circuits, but a capacitor at the output does not really help much with a regulator circuit.  The LM399 has a shunt regulator circuit, not just a bare zener. Some filtering may be good, but more with a little series resistance. A capacitor with added series resistance can help a voltage regulator. For some odd reason  this point is still not very well investigated for the LM399 - ideally I would have expected a note on the capacitive load, like with the TL431 or similar.

[dietert1]

In my board i have a 51R resistor in series to C2 and the amplifier input. Plus some protection diodes for the LM399 against reverse voltage and against overvoltage. I left the 4 legs as long as possible and carefully formed some bends into them in order to reduce stress on the hermetic seals. Landing pattern isn't a 1/10 " but a 3/10" square.
Yes R1/R2 are not critical, i am using one good metal film (2.87K 0.1 % 15 ppm/K). While a solution with a potentiometer is simple and correct, i would prefer tuning with fixed resistors and omit R4, R5 and R6 but double up R7 instead.

Regards, Dieter

[iMo]

You may try to simulate the dynamic response of the 399 to the output capacitance. Below a model I put in LTspice (the TC characteristic is an example only as you need to know the actual parameters of the zener and other components on the chip, the schematics comes from the datasheet).

[mawyatt]

You may try to simulate the dynamic response of the 399 to the output capacitance. Below a model I put in LTspice (the TC characteristic is an example only as you need to know the actual parameters of the zener and other components on the chip, the schematics comes from the datasheet).

Interesting!! Here's link to some earlier LM399 LTSpice simulations we did back in 5/1/2020 to get a "feel" for how the LM399 behaved.

https://groups.io/g/LTspice/topic/73959871#120404

https://groups.io/g/LTspice/files/z_groups.io/Files%20sorted%20by%20message%20number/msg_120195/Voltage%20Ref%20V1.asc

https://groups.io/g/LTspice/album?id=246490

Best,

Edit: Forgot to add the schematic from June 2020

Here are 3 references in the schematic to allow comparisons, one is simple Zener with 2N3904 as a diode powered from VCC (+15V), other is LM399 model powered from VCC, and one with LM399 model & Op-Amp to produce +10.000V Output. Also a couple test circuits with the LM399 and other references (unpopulated LTZ1000).

[Kleinstein]

Thanks for the spice circuit of the LM399.  To make it work with my installation I changed the transistors to 2N3906 and used a different zener. In addition R6 should be larger, more like the original 2 K. Otherwise much of the circuit is inactive or it needs a way higher current.

The AC impedance than looks quite similar to the data-sheet. Not very surprisingly the 100 nF cap is not really helping the regulator, more like adding a resonance at some 50 kHz.  50 Ohms and 100 nF in series on the other side look good.  For EMI suppresion less than 10 nF would be sufficient.

[dietert1]

The 50 ohm resistor and the capacitor should not be in the supply path from R1/R2. C2 moves towards the input of the amplifier.

Regards, Dieter

[iMo]

@Kleinstein: The model v5 with my own zener diode.. The current via zener set to aprox 250uA (as it is in reality afaik)..
The first line in the zener's model are the temperature coefficients you may play with..
PS: could you show me your model?

[Kleinstein]

Attached is my model variation with the setup to simulate the dynamic impedance with a parallel capacitance. The main changes are the zener ( just a type that comes as standard), R6 = 2 K as in the data-sheet and R2 reduced to 10 K to get an AC impedance closer to the DS curve. The the PNPs are 2N3906 ( instead of the less common 3905).
The zener details have quite some effect on the temperature curve, but this is not the relevant part here.


With an AC current source of strength 1 the resulting ouput voltage directly corresponds to the impedance. 0 dB corresponds to 1 Ohm and 20 dB corresponds to 10 Ohms.

[r6502]

Hello All,

When I designed the schematic, my focus was, to get a configuration, that will be adjustable to 10V, regardless of the tolerances of the LM399 zener voltage and the tolerance of the available resistors. The zener voltage of the LM399 is specified to 6,95V +/- 2%. I made a excel sheet, where I calculated several resistor values, that will be available on stock at a distributor. Than I came to the conclusion, that with the values, represented in my schematic, I will be able to adjust the reference to 10.000V. I made an annotation of the schematic, that's why the parts may have different IDs compared of my last posing.

In the beginning I had thought, to bye a new set of 1N827 zener diodes, but when I looked at the schematic of the original reference circuit, there are several resistors marked with "SAF" (Set At Factory) to be able to settle the tolerances of the of the Zener diodes and to set the output voltage to really 10.0000V. Due to the fact, that the resistors with low TK are relatively expensive and I do not have a stock of those, especial many different values,  I decided to replace the original reference with one based on LM399.

The fixed resistors I used for my schematic are Vishay 102 series resistors with TK of 2ppm°C. The trimpot is also a Vishay (series 1202) part with a low TK of 10ppm and based on metal foil. The cosponsoring TK of the resulting resistor 2x5k parallel to the trim pot will be lower then the TK of the trim Pot but higher the the TK of the two fixed resistors.

The simulation with the C parallel to to the LM399 zener is very interesting. I did not really thought about this parallel C, it is just like something, that I always place a C in parallel to a zener diode or similar circuits. I'm using Altium - is it possible, to use the Model you posted  also in Altium, or do I have to install LT spice for this?

@Kleinstein:
I made a photo of the used resistors from Vishay - see below. Here is also a link to this series of foil resistors:
http://www.vishaypg.com/foil-resistors/list/product-63001/. The spacing between the pads is 3,81mm ( 0,15inch) . The heater of the LM399 gets its own tacks of course - have a look at the PCB screen shot.

So have a nice week Guido

[Kleinstein]

The simulation of the LM399 internal circuit is rather basic, using kind of normal transistor models and also the zener model is just a random pic of what is available. For the AC simulation this should be reasonable OK, for the temperature dependence this could be off quite a bit. As no special features are used a different spice variant should be OK for such a simulation too. 

The main point from the AC simulation is that just a 100 nF cap does not help at low frequencies, more like makes things worse at a round 50 kHz and only help above some 100 kHz.

For filtering the solution from Dietert looks good and reasonable.
The amplifier part could also have the usual capacitor in the feedback path (e.g. about parallel to R3).

In the  7 to 10 V scaling circuit not all the resistors are equally critical, one could get away with less of the good BMF resistors. Other resistors usually have a different footprint. Especially R1 and R2 are not that critical with the LM399. The adjustment range can be quite a bit smaller, so a lower value for R5, R6 and than R5,R6 are less critical so that a single more normal resistor (e.g. good thin film like PTF56 series) would be sufficient. If a 2nd optional resistor is preferred for coarse adjustment is wanted this would be something like an optional one (likely 100-400 K range) in parallel to R7 and / or R3.

[iMo]

.. Interesting!! Here's link to some earlier LM399 LTSpice simulations we did back in 5/1/2020 to get a "feel" for how the LM399 behaved.

FYI - here are the LM399 and MAC199 die shots with an analysis what is there on the chip. A good simulation model would be nice to have!
https://www.richis-lab.de/REF02.htm
https://www.richis-lab.de/REF02a.htm

[mawyatt]

.. Interesting!! Here's link to some earlier LM399 LTSpice simulations we did back in 5/1/2020 to get a "feel" for how the LM399 behaved.

FYI - here are the LM399 and MAC199 die shots with an analysis what is there on the chip. A good simulation model would be nice to have!
https://www.richis-lab.de/REF02.htm
https://www.richis-lab.de/REF02a.htm

Thanks, we had seen this work by Noopy and it helped develop the spice models we created. It's interesting to see the differences in these chips and trying to understand how they behave.

Best,

[dietert1]

Meanwhile i have been running the 5x LM399 board with a PWM gain stage 7 => 10 V for more than 2000 hours and initial drift stopped. From the data at 1500 h and 2250 h less than 1 ppm/year is expected. Recently i used some epoxy glue to fix the five thermal protectors, didn't make a big difference.
This is a ratio measurement where my DIY reference runs against the built in reference of a HP 3456A. Soon i hope to finish a setup with a low thermal scanner and difference measurements (better resolution).

Regards, Dieter

[branadic]

Would be nice to share not only results, but schematic and the like, a good base for a valuable discussion. 
What about (switching) noise, t.c. of your solution?

-branadic-

[dietert1]

Since i learned a lot from the previous PWM discussion in this thread, here are some schematics i made when working on this. Hope this info serves the experts, since i won't have time to contribute much more.
The first schematic represents the circuit i am still using (log above). You can see a second order filter and ripple cancellation. Other details are a resynch Flip-Flop, separate power and supply grounds, a 7 V guard for the filter circuitry and circuitry for operating the filter caps near zero voltage. This is based on biasing large electrolytic caps. This circuit is very slow and useless for a programmable calibrator.
I built ripple cancellation as a mod that would allow me to reduce PWM frequency to 1 KHz, yet in fact i did not modify the MSP430, so it still runs at 10 KHz. The second schematic shows a study of a low resistance output stage that would serve another factor 10 or more of precision. It is much better than any integrated switch i know of: Switching within some nsec and less than 1 Ohm residual resistance. One day in the future that could replace the main PWM output stage. Ripple cancellation can remain with the IC switch.
Don't have a more detailed schematic of the reference board, but an image. The LM399 thermal protection caps are glued to the yellow 100 nF film caps.
The log i showed is with the reference board and the PWM on a table, wrapped in a towel, no metal enclosure, no additional oven. Since we have a 2 or 3 °C day/night temperature variation the log demonstrates that the TC of the whole circuit has less than 0.5 ppm/°C. This isn't a determination, just an upper limit. Each blue data point in the log represents 500 measurements with NPLC=100 and it looks like those are quiet on a fractional ppm scale. The 1 ppm steps are definitely present.

Regards, Dieter

PS: My PWM prototype has an offset of 11 ppm (max of log). The PWM ratio remained unchanged after implementation. In my opinion a more perfect PWM divider would be more accurate, i mean without tuning or "calibrating" it. Accuracy depends on the symmetry of the PWM switches: resistance and/or timing difference. At 10 KHz a 1 nsec timing asymmetry makes 10 ppm. IC muxes have delays of about 50 nsec that can easily contribute an asymmetry of 1 nsec.

[dietert1]

Today i took an image of the PWM board. At the top left corner it has a n-channel JFET to start up the references, J2 as shown here https://www.eevblog.com/forum/metrology/building-your-own-voltage-reference-the-jvr/msg1343014/#msg1343014. This was missing in my schematic.
In the lower right corner there is a 3.3 V regulator for the MSP430 module. The mezzanine on that module is the sync Flip-Flop. The second PWM switch i added later for ripple compensation is hidden under the processor module. The large filter capacitors are what i found (spares from speaker xover revision). Maybe 100 V film capacitors are completely sufficient for low leakage and cleaning the board would be more important.

Regards, Dieter