More than one reference all together? why?

[nour]

I want to build an LM399 Reference box ad I came across xDevs model design

I have noticed there are 2 LM399 connected together, why?
Would that help regarding the stability?
because I have seen the other LTZ1000 diagram and there was only one, so why not two like LM399?

[TiN]

Some designs use four LTZ1000 or even ten LTZ1000, such as Fluke 7001 system or Datron 4910. More references provide lower noise (due to averaging) and better prediction uncertainty (due to statistics, it's less likely that all references drift very differently).

Design with two LM399 is cheap (total BOM maybe 50-60$), but LTZ1000 5 times more expensive. Actually I use multiple LTZ1000 modules together as well. Some data. Also multiple references (doesn't matter eitherh LTZ or some cheaper LTC6655) can be compared to each other for better idea of stability. This way you don't have to worry about your DMM/voltmeter drift, as it's nulled out.

Beware with that design, values on schematics are not for exact 10V or 20V output and need correction.

[nour]

Thanks, I somehow knew that you will be the first one to reply  , anyway, the perfect person to do the first comment.

Ok, so basically due to averaging the noise will become less, that is very interesting, so I have two questions

First, how many LM399 should we but togeather to make the noise lower by an order of magnitude?
second, which is better, 2 LTZ1000A or 10 LM399AH regarding the noise and stability? or maybe those are two separate things, not quite sure about it really!

[nour]

Since you are here, I wanted to ask you about some details about your LM399 circuit!
what is the EXT_IN+ and EXT_IN- for?
on the inverting input of the LTC2057HV you have 4x 3.75K resistors connected in parallel and also 3x 466 connected in parallel and altogether connected on series, why? why not one for the first 4 and one for the other 3, actually why not only one resistor for all of them?

[nour]

Also, one more thing, what is the square in the red circle in the attachment image means?, I am using eagle and I haven't seen this before!

Thanks

[TiN]

Some voltnuts may drag me to the witch bonefire, but I'd say even 1 x LTZ1000 is more stable than 10 x LM399s. But LTZ is far more delicate circuit to begin, if you don't have high budget to spend. It's not enough to just buy parts and get it all together, testing the circuit stability (even LM399) require understanding and proper (read, expensive) equipment.

I've used multiple resistor positions, so I can use different resistors to trim for specific value, or match resistors with opposite temperature coefficients for better stability. You can use just one position if your requirements are relaxed.

EXT_IN+/- are there due to multiple purpose of the circuit. Alternative use was NOT to put LM399 modules there, but supply external voltage (for example 7.1366351 from one of LTZs) and use opamp + resistors on this board to generate different voltages.

Square on schematics means guard trace inside the square with the box trace. Google term : high-impedance guarding.

Here's interesting appnote from Analog Devices which explains a little bit about using multiple parts for lower noise. It's different purpose there with using multiple opamps, but idea is close. In two words - momentary noise of each reference are not matching together, so once outputs summed opposite "spikes" are compensating each other, giving lower peak-peak amplitude signal. However measurement of this noise levels is not easy, very slow process

[Kleinstein]

TIN's circuit used a lot of resistors in series / parallel connection and AFAIK also some for alternative use. So this is not a very good example, but more confusing.

With equal references the noise (voltage) goes down like the square root. So to lower the noise by a factor of 10 it would take 100 units in parallel. So it work wo use several in parallel, but it is not very effective in larger numbers. Another point to take into account is that not all units of the LM399 will have the same noise - so in might be worth testing for noise and leave out the noisy ones. Ideally one could adjust the weights (resistors used for averaging), so that even higher noise units could give a small positive contribution.

With stability, chances are that drift in the units will be similar, so the gain in stability tends to be even less. The use here is more to detect possible bad units. However this would need to look at the individual references and than decide by software which ones are good and which ones are likely bad. With 2 refs one could detect if they are drifting apart to much and especially the case of one shows a sudden jump / step - though no way to tell which one is bad. Same type refs tend to show a similar drift. Paralleling can not detect / reduce that common drift, but only the more random part that effects individual refs.

AFAIK it takes about 20 LM399 to get a noise level comparable to the LTZ1000. The number can very, depending on the current and maybe temperature used in the LTZ1000 circuit. Also noise of LM399 is not all the same. The drift values are essentially never reaching that level.

[nour]

Thank you very much, I really appreciate your replies the are very valuable to me, I do understand your emphasis on the expensive equipment point, yes I know those parts on their own are not much to benefit from them unless I have the proper equipment to properly measure all necessary parameters which are very unfortunate for me!

[nour]

Is there a cost friendly way to reduce the noise of the LM399 without adding more units to the design?
like for example adding active filtering stages with very low noise floor without gain to the output?

[MK]

The frequency of the noise is very very low, and there are the intermittent steps and orientation changes to consider. Knocking out the mains hum etc is feasible, but filtering the VLF noise is impossible. There is noise with a period of days, weeks and months.

[David Hess]

Is there a cost friendly way to reduce the noise of the LM399 without adding more units to the design?
like for example adding active filtring stages with very low noise floor without gain to the output?

The most cost effective way to reduce the noise is to use an integrating converter and ratiometric design which rejects the noise.  Otherwise the best way to reduce 1/f noise is to use a quieter source.

[Conrad Hoffman]

Noise can be addressed as described above but I'd also make the case for several references that you can inter-compare, three or four. That way if some circuit flaw occurs you know which one is bad, and can even reset it after repair. I'm also a fan of having several different types of reference circuit, reducing the chances of all the references having similar long term drift.

[nour]

Is there a cost friendly way to reduce the noise of the LM399 without adding more units to the design?
like for example adding active filtering stages with very low noise floor without gain to the output?

The reason I said "without adding more units" because I have seen several teardowns for high-end benchtop multimeters, 6.5 to be specific that have only one LM399 reference, so I thought there is no need to have extra LM399 and maybe there are alternative ways to just reduce the noise with maybe just active filtering or something similar!

And one more thing, I think for DC measurements, the noise that would make a concern would be under 100 Hz noise, right, wrong? or maybe between 0.1 and 10 Hz?

The frequency of the noise is very very low, and there are the intermittent steps and orientation changes to consider. Knocking out the mains hum etc is feasible, but filtering the VLF noise is impossible. There is noise with a period of days, weeks and months.

I really didn't understand what do you mean by noise with period of day, weeks or months

[timb]

The frequency of the noise is very very low, and there are the intermittent steps and orientation changes to consider. Knocking out the mains hum etc is feasible, but filtering the VLF noise is impossible. There is noise with a period of days, weeks and months.

I really didn't understand what do you mean by noise with period of day, weeks or months

Think noise with a frequency in the uHz to pHz range.

[nour]

Think noise with a frequency in the uHz to pHz range.

Yea, I know it meant this, but I meant to ask about how is this happening, why it is happening, is it normal for all references to have this kind of weird noise(is it actually weird?)

[CalMachine]

Think noise with a frequency in the uHz to pHz range.

Yea, I know it meant this, but I meant to ask about how is this happening, why it is happening, is it normal for all references to have this kind of weird noise(is it actually weird?)

https://en.wikipedia.org/wiki/Pink_noise

[David Hess]

Think noise with a frequency in the uHz to pHz range.

Yea, I know it meant this, but I meant to ask about how is this happening, why it is happening, is it normal for all references to have this kind of weird noise(is it actually weird?)

In MOSFETs it is caused by charge trapping in the gate dielectric which shifts the threshold voltage.  In PN junctions it is caused by defects which alter the gain.  For the simplest case where only one charge trap or defect is present, pink noise becomes popcorn/telegraph/burst noise.  When many charge traps or defects are present with different time constants, then the aggregate popcorn/telegraph/burst noise becomes what we see as pink/flicker/1/f noise.

It can be filtered only with difficulty because of the long time constants required so it is better to avoid generating it or use a design which inherently rejects it.

[MK]

Hi,

Read this and you will see that there is a very low frequency noise component present in even the very best units available:


http://vmetrix.com/ZenerP.pdf

[Andreas]

The reason I said "without adding more units" because I have seen several teardowns for high-end benchtop multimeters, 6.5 to be specific that have only one LM399 reference, so I thought there is no need to have extra LM399 and maybe there are alternative ways to just reduce the noise with maybe just active filtering or something similar!

Hello,

its not active filtering but selection.
filtering is difficult at frequencies near / below 0.1 Hz.
You would need large (electrolytic) capacitors which have
(not constant over time) leakage currents that also create noise.

And you do not want a "stinker" LM399 with > 7uVpp in a precision instrument.
And also if using averaging: a LM399 with 7uVpp will spoil a group of references with 3-4uVpp.

Attached noise measurements of LM399#2 (7.2uVpp) and LM399#1 (3.6 uVpp)
y-axis 1 uV/div
x-axis 1 sec/div



And you also want to sort out references with "popcorn" noise
(jumps of several uV or even ppm´s for several seconds or even minutes)
Attached a example of a MAX6350C reference over several minutes with jumps up to 4-5 ppm.

Hi,

Read this and you will see that there is a very low frequency noise component present in even the very best units available:

http://vmetrix.com/ZenerP.pdf

I would not call the seasonal changes as noise.
Since you can correlate it to the environment.
So it is more a drift. (temperature / humidity / pressure dependant).

Noise is something statistical that cannot be predicted.

with best regards

Andreas

[nour]

So basically, you are saying I should buy some of them and go through the process of categorizing them according to their noise and choose the best one 

maybe I will just buy a lower noise reference like LTZ1000A and use it, right?

Hope you correct me if I am wrong about it!

[Kleinstein]

Filtering on the LM399 can be useful, but this is usually only for the higher frequency range, like higher than 10 or 100 Hz. For a DMM the noise frequencies of interest are more in the 0.1 mHz - 1 Hz range. Normal analog filtering is not practical in this range.

In some applications it is possible to use a kind of filtering, by using different references for different frequency ranges. So one could use a lower noise, but higher drift reference in combination with a low drift, but higher noise reference (to cover the really long times scales). This is kind of using a digital filtering scheme which can work well for very low frequencies too. The Keithley 2001 7.5 digit DMM is using such a 2 reference scheme, though I am not sure the implementation is made as well as it could be.

Choosing the best out of several LM399 is a viable way. So is averaging a few (like 2 or 3) of them, with checking there individual quality and if needed leave out the really bad ones or adjust the averaging weight. With a suitable low weight even a noise reference could help a little to reduce the overall noise. There is little sense in using more than about 5 of the LM399, as then you start to get to the price range of the LTZ1000 with even lower noise.

[Andreas]

Hello,

also a LTZ1000 can have popcorn noise.
https://www.eevblog.com/forum/metrology/ultra-precision-reference-ltz1000/msg833804/#msg833804

So you have to check this also with a LTZ.

With best regards

Andreas

[MK]

And you also want to sort out references with "popcorn" noise
(jumps of several uV or even ppm´s for several seconds or even minutes)
Attached a example of a MAX6350C reference over several minutes with jumps up to 4-5 ppm.

Hi,

Read this and you will see that there is a very low frequency noise component present in even the very best units available:

http://vmetrix.com/ZenerP.pdf

I would not call the seasonal changes as noise.
Since you can correlate it to the environment.
So it is more a drift. (temperature / humidity / pressure dependant).

Noise is something statistical that cannot be predicted.

with best regards

Andreas

seasonal shifts were only one paragraph in that paper, Fig 1 is the most telling diagram.

[David Hess]

The reason I said "without adding more units" because I have seen several teardowns for high-end benchtop multimeters, 6.5 to be specific that have only one LM399 reference, so I thought there is no need to have extra LM399 and maybe there are alternative ways to just reduce the noise with maybe just active filtering or something similar!

Good multimeters benefit from using analog to digital converters which integrate the input and by necessity the reference.  This produces a sin(x)/x response which low pass filters the input and the reference and tracks the desired resolution by reducing the noise from both.

Even with this though, pink or 1/f noise can be seen in many cases in the display where the last digit "flickers" between values in excess of what would be expected from quantization error and other noise sources hence the name "flicker noise".  Unfortunately because flicker noise increases at lower frequencies, the longer integration times which produce a lower cutoff frequency also integrate more of the flicker noise which increases at low frequencies.

Burst noise sounds like popcorn or morse from a telegraph hence these other names for the same thing.  It is caused by contamination of the semiconductor fabrication processes and has been a solved problem for decades now which is fortunate because it can be very difficult to screen for.  I suspect low noise zener references suffer from it more simply because it is more noticeable with their inherently lower noise levels unlike higher noise references where the burst noise is covered up.

Below is an example of popcorn noise which took me weeks to track down in the vertical amplifier chain of a Tektronix 2230 DSO.  The fact that it looks so digital made me initially search for a problem in the digital logic but it was really in a linear transconductance amplifier.

[Muxr]

Consequently I am working on a non metrology related project which requires an ample amount of Zener noise. I just actually bought like 20 different zeners to characterize the noise with, and pick the one with "best noise"  .