LM399 based 10 V reference

[oz2cpu]

i have been reading a million hrs in this thread :-) thanks to you all for the high level of skillsharing

one thing keep comming up, the opamt to use, i think i get the point,
the only thing that really matter is the drift pr temperature,
here are the 4 opamps all folks talk about, the best one is obsolete, darn it..
so.. what is the best, that i can get ? if cost is not an issue ?
anything else beat the mentioned types ?

[Kleinstein]

A few more modern classic OPs to consider are OPA277, OPA207, OPA2205 (dual) and ADA4077.  One could also consider an auto zero OP.

The accuracy of the LM399 is anyway somewhat limited, so even slightly lower grades are not that bad. It is not only the drift with temperature that matters. Also drift with time and humidity can be an issue.

[oz2cpu]

DANKE !
the mentionened opamps was a great help, wow some are really amazing,
another thing to consider, we dont see so many people use a temperature regulator on the opamp, and resistors,
it could be another case, round the LM399, like +40C regulated, and then isolated,
if done correctly, temperature ultra stable, the opamp and resistors used could be much less hiper-performing.

[oz2cpu]

I am impressed, darn stable at first impression

[oz2cpu]

to make 7V from the LM399 into 10V
i need two resistors on my opamp
in the ratio 3/7
prefered a two in one temperature compensated ratio resistor chip,
scrolling thru this thread, i dont see this magic ratio mentionened ?
and a solution that is best possible, no matter what cost.
all seems to use seperate resistors, 5k and 20k are mentionened quite a lot,
but i just dont see how they make this gain ratio ?

[Andreas]

in the ratio 3/7

Hmm,

what about 1.5 / 3.5 ?
could be made from a 8 Resistor chip like TDP1603

with best regards

Andreas

[oz2cpu]

thanks alot for the input andreas

i am unsure how you make the 3/7 or 1.5/3.5 ratio with 8 equal value resistors ?
please explain ?

the suggested 8 in one, 5ppm tracking
https://www.vishay.com/docs/60045/tdp.pdf

this is a 4 in one (i got many of them stocked)  4 x 1k type, also 5ppm tracking
https://www.vishay.com/docs/60005/orn.pdf

so i hope i can use them, i am a bit worried about build size in this feedback circuit, i prefer it to be smallest possible
less noise pickup

next question : capacitors for opamp feedback, and 7v filtering, they say 1uF, but i am very sure the type is important,
noise ? leak, temperate problems, and what not from capacitors must be investigated.
recommendations ?

[Kleinstein]

Using seprate resistor networks reduces the matching effectiveness. It can still be acceptable though not perfect.
In the 1.5 and 3.5 combination this would be 2 mA off current for the divider and thus a bit on the high side. The attchment shows the circuit, using 1.01 for the resistors of the 2nd chip. The sharing between the 2 chips is not perfect, but still good with only some 10 % unbalances part  (1/3 for the top and 1.25/3.5 for the lower part).

A reference source is essentially static and there is thus no really need to look for low DA in the capacitors. So polyester film should be good enough.
Directly at the ADR1300 reference the impedance of the reference is quite low and leakage is thus not as critical. So there even X7R ceramic or a tantalum electrolytic can be good enough. With X7R ceramic I would plan for more than 1 µF (e.g. 2 or 3 in parallel) because the capacitance under DC bias goes down. From the response curve for the ADR1300 the 1 µF already look like they are on the low side - enough to make it stable, but a little more would not hurt.

Edit:
Just to improve stability at the OPs one gets away with way less than 1 µF and 1-10 nF should be enough there. This can than be C0G ceramics.

[oz2cpu]

I try a mockup with the ORN resistors i got plenty off, like this..

i arranged the connections, so the ic is symetric powered,
 dont you think they TC match best this way ?

[Kleinstein]

With 2 chips one would get the best maching in the way I have drawn it. So array no 1 for the 2 parallel resistors in the upper part and one single and one in the parallel pair for the lower resistance. With quite some current in the divider one needs to make sure not to have to much copper trace and a suitable ground point.

The idea it to mix the resistors from both chips so they contribure about equal to both sides of the divider:  In my version the right chip provides 1/3 of the upper resisistor and about 35%  (1.25/3.5) to the lower resistor. So not ideal, but pretty close. So matching between the 2 chips is less critical (only about 1/10).

[oz2cpu]

here, i made it, as i drawn it, it seems to work, no temperature cycle yet, but i get 10V out :-)

i challenge your drawing: see mine again,

the two parallel resistors are in EACH ic,
and this way same power is lost in each ic = best TC match, correct ?

[Kleinstein]

The matching is not about getting the same power to both chips, but about getting a similar fraction in contributing to the upper and lower resistance.
So one wants to use the resistors from 1 chip in both the upper and low part and not have the upper part from 1 chip only.

In your version the help chip has 100% of the upper resistor and 1/3.5 for the lower resistor.  So about 70 % of the matching error is left over.  This is more like a worst case solution.

In my suggestion the richt chip contributes 33% and 35%, so relatively colse to equal. This gives around 7% of the matching error remaining or about a 10 fold reduction in the sensiticvity.

With chips from the some batch the matching between chips may be quite good anyway. So one can be lucky and still get a stable output - it ust needs a little more luck.

[oz2cpu]

thanks a million times, a very good answer,
since i got two brand new LM399 running in parallel, on the same mock up board,
i hereby make the other two resistor chip wirings exactly like you suggest,
and for the final challenge, temperature cycle it,
then it is going to be interesting if we are able to detect or reveal what perform the best :-)

here is your suggestion, redrawn a bit, for easier soldering,
but i am sure it is exactly the same as you explained, right ?

EDIT : your suggestion works, and i got 10V out from that one too

[dietert1]

Here i have some measurements on the 5x LM399 array with PWM gain stage that i made last year (see above). The device is still in prototype stage. I put it into a NH2 incubator controlled by an Arroyo TecSource at 22.1 °C. The 10 V reference output was monitored by an Advantest R6581T sitting in a temperature chamber at 23.25 °C.

First diagram shows a six months log of voltage measurements in units of uV, so the diagram vertical axis spans 1.2 ppm of 10 V. I added two temperature logs, the pink one inside the temperature chamber, the orange one ambient temperature (room temperature). I added 14.5 °C to room temperature values to make both logs fit onto the right y scale. That's about how much the R6581T heats above chamber air temperature.
There is a negative correlation between voltage and ambient temperature. Drift is estimated as -12 nV/day, about -0.4 ppm/year. A two parameter fit for temperature and time wasn't yet conclusive.

The red circle indicates a TC determination shown in the second diagram. The incubator temperature with the LM399 array & PWM got an increase of 1 °C. TC was determined as -0.3 ppm/K. It means the PWM can compete with precision resistor dividers.

As the incubator isn't a precision oven, it likely causes the negative correlation observed in the log.

Regards, Dieter

[try]

thanks alot for the input andreas

i am unsure how you make the 3/7 or 1.5/3.5 ratio with 8 equal value resistors ?
please explain ?

naked chip, value does not matter, it is all about ratios:
1 1 1 1 1 1 1 1

use two of them in parallel
1 1||1 1 1  1 1||1

gives
1 0,5 1 1 1  0,5 

switch them in series:

1-0,5   1-1-1-0,5

gives
1,5 and 3,5

total is 5
3,5/5=0,7

7V / 0,7 = 10V.

But honestly, what improvement do you expect over a simply solution with two resistors?
Are you able to estimate that in quantitative terms or even measure it?

Using that TDP1603 Andreas mentioned could be considered overkill (but it is an already much cheaper solution compared to what was promoted years ago here...).
You might even try out 8x 2,2kOhm resistors in metal film quality.

Real voltnuts obviously only use hermetically sealed custom-built Vishay VHP dividers at the cost of the GDP of Nicaragua.
Another possibility is to use custom-built wirewound precision resistors, wound by virgins at full-moon in leap years.

But if it makes you feel good throw in all the expensive stuff.


Regards
try

P.S.: You can try out an arrangement that takes into account heat distribution because the two paralleled ones are running with half the current: P= (I/2)^2*R = 0,25*I^2*R compared to I^2*R for one single resistor. But with larger resistor values it won't matter anyway.

[Kleinstein]

The resistor networks tend to be quite stable and the ORN ones are not that expensive (some $4). The specs only call for 5 ppm/K TC matching, but so far my measurements with 10 K / 50 K sets were more in the 1 ppm/K range for the relative TC.  Chances are good that the relative drift can also be quite good, though here only time will tell.

In a 10 V reference the 7 V to 10 V step is still one of the more critical parts. Chances are that when starting with a LM399 / ADR1399 the resistors for the gain may still be the larger TC part and also be a part of the long term drift. So it makes sense to use rather good resistors there. Compared to a $10-$20 reference the price for the 2 resistor arrays is not that bad, and they can still be in stock.

The self heating should not be that crictical, as in a voltage reference the power is essentially constant in time. The slightly complicated circuit comes from distributing the resistors of the 2 sets of 4 so that they contribute about similar and get away with an easy layout / soldering.

[Kleinstein]

For one of my LM399 refs I got the popcorn noise on a very long time scale, but still about the same amplitude / step size.  This makes the curve for the short time look well, but a for a real world use the long times that are usually important. I may be good to look at the noise also for a little longer and maybe in the time domain.

What is the vertical scale, is it volts (rel. to 10 V or 7 V ?) or PPM of voltage.
Another point is that the reference in the meter may contribute, both at the fast and slow end.

[Kleinstein]

In the initial phase all references are expected to drift a little more. So there is a good chance the drift will go down after some 100 hours of operation.
The Ref102 data show some drift and popcon noise with about the 0.3 ppm step size. This is comparable to the step size typical for a LM399. So some of the noise may as well come from the reference in meter and also some drift could in principle come from the meter as well.

[Kleinstein]

It is not just the time after warm up, but usually the time after soldering or production that matters for aging. The turn off / turn on cycle can introduce some new drift cycles too, as part of the hysteresis. At least for the LM399 the hysteresis part is usually small and no long warm up needed for the reference. For the 10 V level of the LM399 there can also be some drift of the resistors.

[Andreas]

Hello,

it is a unfair comparison in my opinion:
The plastic package of the REF102CU will also be sensitive to relative humidity.
(up to around 0.5ppm / %rH and with a time constant of several days).
So you will have also seasonal changes depending on region.

see eg. here for a REF102AP
https://www.eevblog.com/forum/metrology/best-out-of-the-box-10v-reference/msg1242608/#msg1242608

https://www.eevblog.com/forum/metrology/tales-from-the-mouth-of-a-voltage-reference-ageing-box/msg3618917/#msg3618917

The statistical average will not help much for humidity.

with best regards

Andreas

[Kleinstein]

The higher frequency noise, like above some 10 Hz can be filtered with not that much effort. So that is usually not the noise to worry about. The weak point of the LM399 is more the popcorn type noise, that can extend to rather low frequencies / happen at a long time scale. I have one unit with jumps about once per hour - at first this looks good, but can also be trouble as the noise extends down to very low frequencies.

[Andreas]

The weak point of the LM399 is more the popcorn type noise,

Sorry no:
Popcorn noise is not related to LM399.
From theory popcorn noise is the result of impurities in the silicon chip.
You can see popcorn noise on all kinds of voltage references if you look long enough.
Here a example of a AD587 with a LNA 0.1-10 Hz on scope.
Or a MAX6350 here as reference of a ADC which had popcorn noise at certain temperatures only.

Of course I would do a screening for popcorn noise for a stable statistical reference and sort out the stinkers.

with best regards

Andreas

[Andreas]

6500 is noisy toy, but very stable.
34410A is very low noise equipment... but poor stability.

Hello,

How many NPLC for each instrument. And how is the measurement rate?
~450 ms for one measurement looks a bit strange to me.

my K2000 at 10 NPLC and 50 Hz needs around 600 ms. (is the 6500 faster?)
The 34401A needs 400 ms at 10 NPLC.

with best regards

Andreas

[Kleinstein]

With timing from the RP and some extra delay for data transfer and processing the 440 ms make absolute sense.

The data for the DMM6500 look indeed a bit noisy.
The 34410 drifty part / low frequency noise is also a bit unusual. I have expected better performance for both meters.
For comparison it may be interesting to compare to the noise of both meters with a short at the input.

Much of the noise from the references could be on a shorter time scale (e.g. minutes) and may be a bit hard to see at this scale. So it may help to look at a shorter part, like 0.5 or 1 hour, to see if there are correlations on the shorter time scale.
The slight problem here is that there is noise from the DUT, the meters references and also the ADCs are of comparable size. So it can be a bit tricky to tell those noise sources appart. Things are hard enough when the meters are lower noise.

It may be better to test the references by having 2 x  LM399 based 10 V reference (or the 7 V level) and directly take the difference there and than only look at the difference with one of the meters in the 100 mV range. This would reduce the noise sources to 2 x LM399 and only a little noise from the DMM.

[guenthert]

Some time ago Ill try to explain about "time delay between points" to Mr.Kleinstein, but my English skills is poor, my explain looks bad.

No worries, your English is much better than the Russian of most readers here