LM399 based 10 V reference

[David Hess]

Yep, LM399H, date code 1836.. Some time back I put it inside my 34401A with a potential for replacing the original HP399 one (with higher noise), but I saw one shot per day. So I repeated the measurement again outside the meter (with 3.4mA current) and I've got 7 shots per day. Just charging the batteries and I will repeat the same with my lowest noise 1399..

So your LM399H is new production since the popcorn noise problem was "solved".

I suspect that the LM399 and similar buried zener references are more likely to display popcorn noise simply because they are much lower noise to start with, allowing low levels of popcorn noise to become apparent.

I have seen similar 4 µV range jumps in all 4 of my LM399 reference. While the size of the steps is similar the rate for the jumps varies. Some faster (like 1 jump per minute) to rather slow like about 1 jump per hour.

Popcorn noise is usually between two levels, but sometimes there will be several defects trapping charge leading to multiple levels.

The timing can vary widely over time, which is one of the reasons it took me so long to fully diagnose my 2230.  I had to wait days at a time to catch it.  Of course since it was only showing up in digital storage mode and it has a mostly fixed amplitude, at first I assumed it was a digital problem and I was checking address and data lines for a while until I worked back through the digitizer to the analog channel switch.

[iMo]

..
So your LM399H is new production since the popcorn noise problem was "solved".
..

Well, that is the question how ADI handles the chips. You may have 15y old wafers with X thousands of chips on them gathering dust on the shelf and with a higher demand you simply cut the dies off and sell them as the fresh ones.. Who knows..

[mawyatt]

..
So your LM399H is new production since the popcorn noise problem was "solved".
..

Well, that is the question how ADI handles the chips. You may have 15y old wafers with X thousands of chips on them gathering dust on the shelf and with a higher demand you simply cut the dies off and sell them as the fresh ones.. Who knows..

Usually wafers are stored in a dry nitrogen container, so shelve life isn't an issue.

Best

[Alex Nikitin]

FWIW, here is how the ADR1399 can jump. It is installed in my HP3245A which otherwise is quite stable. The 0.6ppm jumps might occur 2-5 times in a day and might not occur for 2-3 days in a row. It is very annoying. I plan to swap the ADR1399 for another sample, time permitting (as even testing one requires days and weeks). On the other hand, ADR1399 in my modified Keysight 34465AM doesn't jump at all (thankfully).

Cheers

Alex

[Kleinstein]

One can see popcorn noise with other devices too. I saw some with J111 and J113 JFETs and a TL031 OP-amp with small steps and more than 2 levels.
With some normal zener diodes I found an exampley with rather large steps (some 700 µV, see atachment), but not all of the zeners are that bad. Not sure if its the BZX55 that is especially bad or it was just the batch I have (I checked a few and they were all poor, not none as bad as the one shown).

[David Hess]

Popcorn noise used to be a big problem with some manufacturers.  AMD was a third party source for a lot of common analog ICs in the 1970s and as I recall, they had a lot of problems with popcorn noise, so much so that if this was an issue for you, it was better to just avoid them.

http://www.bitsavers.org/components/amd/_dataBooks/1974_AMD_Data_Book.pdf

[iMo]

This is with my 34401A as the null meter and with my WW pots divider box, set to aprox 70uV difference. 10h long run only, LM399#5 is not the lowest noise one, but with none visible jumps at least (within 10h)..
Y axis is 1uV/div, ambient is blue.

[David Hess]

There is a mathematical connection between flicker noise and popcorn noise, even if they are not produced by the same source.

As I said above, popcorn noise can be generated by more than one charge trapping defect, so it will then present with multiple levels.  As the number of charge trapping defects increases, the spectrum becomes indistinguishable from flicker noise, so it is possible for a part to display excess levels of flicker noise instead of popcorn noise.  This is one of the explanations for why flicker noise is generated at all.

I think Analog Devices has the most readily understandable description, where they consider both to be pink noise from different sources, but that is more recent then when I studied it:

https://www.analog.com/en/resources/technical-articles/managing-noise-in-the-signal-chain-part-1-annoying-semiconductor-noise-preventable-or-inescapable.html

[iMo]

McDonald in his book ("Noise and Fluctuations: an introduction", 1962) calls it "A random rectangular current" .

[David Hess]

McDonald in his book ("Noise and Fluctuations: an introduction", 1962) calls it "A random rectangular current" .

I guess he was not an audiophile.   It was named popcorn noise because it sounds like popcorn popping.

I thought flicker noise was named from the flicker in the least significant decimal digit after the white noise is accounted for, and this is a good way to identify it, but the origin is much older having to do with vacuum tubes.

[iMo]

I was thought the "flicker" noise naming comes from a similarity with optical effects of a "candle flame"..

PS: I made a typo above - the name is D.K.C. MacDonald

[iMo]

..and my ADR1399#2 (no burn-in yet) with the null meter 34401A (set to initial 12uV diff) and WWpots divider - hopefully the depression there is not a jump.. Y axis 1uV/div, blue is ambient.
The lower noise is visible on the first glance - compared to the above LM399#5.

[Birb]

Hi, (some) random question(s)

When constructing parallelized LM399 arrays, is it good to separate the array into small subunits (e.g. 8 = 4*2) and have each unit be buffered by an extra op-amp before the average voltage is given to a final gain op amp?
Or is directly feeding the 8 LM399 voltages into one op amp sufficient (and have that op amp, with a transistor drive stage, give the current feedback to the LM399s)

Also, can a JFET be used as a self-start bias? (thus ensuring proper start-up with low offset op amps, while safely turning off when the circuit turns on)

Thanks for any feedback.

[Kleinstein]

Direct averaging should be OK. The LM399 output resistance is quite low compared to the resistors used for averaging.

It would still make sense to have access to the individual voltages, so get an idea on the spread and if some units have higher than normal noise or drift.

One can use JFETs for the start up. The easiest way is however usually start up from the driver side. So have the output driver in a way that it does not go all the way to zero.

Instead of multiple LM399 one could also consider the ADR1399. From the data the noise is roughly half to a third of an LM399. So 1 ADR1399 would kind of replace 4 to 10 x LM399 when it comes to noise.

[Birb]

Direct averaging should be OK. The LM399 output resistance is quite low compared to the resistors used for averaging.

It would still make sense to have access to the individual voltages, so get an idea on the spread and if some units have higher than normal noise or drift.

One can use JFETs for the start up. The easiest way is however usually start up from the driver side. So have the output driver in a way that it does not go all the way to zero.

Instead of multiple LM399 one could also consider the ADR1399. From the data the noise is roughly half to a third of an LM399. So 1 ADR1399 would kind of replace 4 to 10 x LM399 when it comes to noise.

By startup from the driver side, do you mean clamping the output to be ~1V or some small value?

Addendum:
interestingly i can obtain LM399s for relatively cheap (about 1/4 of the ADR1399) (probably genuine since there is an LT logo on the silicon die) hence it's preferable (though nothing stops you from taking 8 ADR1399s i suppose)

Below is a simple dual 399 reference that may be parallelized (4*2?? probably better than 8*1), probably buffered also (this has zero output protection) Is the JFET wired correctly?

Thanks.


 

[GigaJoe]

"probably genuine " - never heard about "fake" with thermostat on die .. ; mostly new pressed heating shield with NS or LT logo ( orig. heat shell older 80is has LT NS , 90x has to be PHILLIPINES on side ) and legs restored, it salvaged chips. usually legs bend \ broke at joint place. process of handling, unsoldering, restoring etc , and the history unknown ,  it can affect stability \ drift .. or can not ... usually all of them as working condition ...  bought there 6-7 Y. ago https://www.ebay.com/itm/126016030704  came with suspicious mark on shell ..  around 3 of 10 shows no drift after 9 month ( at least i can not detect it) .

[Birb]

date code is new, 2132-2152
Die has LT logo, with clear layout probably identical to some other die shots I've seen. Legs are completely new and unbent.
Leads are completely new, LT logo centered
Only slight oddity is that they seem to have 7.0XV nominal rather than 6.95, but it could very well be a batch thing.
However this does mean they will have to be aged.

[Kleinstein]

There is not much wrong with pre-aged reclaimed LM399.
If the weld on new leads, this could be a small issue as there could be another material involved and different therm EMF. Can still be OK.

The voltage of the LM399 does vary a bit. The units I have are also all a bit above 7 V. There is quite some tolerance.
For averaging it would be better to have them relatively close by. So one might consider avoiding more extreme units (e.g. <6.9 V or > 7.1 V) and the larger importance of the averaging resistors with those.

For the JFET start up, this would be better on the other side of the averaging resistors, one chip would be enough.
Start up from the output could be with a diode at the transistor base toward some 7 V (e.g. 10K:10 K divider from the supply).
The resistors R29 and R30 make little sense. The LM399 units and the FB divider is usually enough load. If R30 is used one should have a decouling cap at the transistor collector.
The amplifier may want a local feedback capacitor.

[GigaJoe]

assuming we have 5uV jump / noise , so need at least equal 25 units to stat. aver. to 1uV ..  if it cost $5 it close to to 1 LTZ thingi
is any advance averaging  25 ?

[Kleinstein]

Averaging 25 units may not be that practical. There is still a slight advantage in averaging a few units, as one can detect when there are bad ones and can get a warning when the differences get much larger. With multiple LM399 it may be feasible to use just the zener and an external oven at a somewhat lower temperature.

For the LTZ1000 there is not just the cost for the chip. It also need some pretty demanding resistors, that add to the costs.

[Andreas]

When constructing parallelized LM399 arrays, is it good to separate the array into small subunits (e.g. 8 = 4*2) and have each unit be buffered by an extra op-amp before the average voltage is given to a final gain op amp?
Or is directly feeding the 8 LM399 voltages into one op amp sufficient (and have that op amp, with a transistor drive stage, give the current feedback to the LM399s)

My suggestion,
use 8 independant units with own output buffer and isolated power supply.

The heater currents are rather large (and change with environment temperature) and may influence the output voltage depending on routing of traces.
Star grounding is much easier with completely independant units.
Each unit has floating supply with only one connection between heater GND and zener GND.
Other possibility would be to use negative voltage for heater supply to get the GND currents away from precision GND.

If one unit begins to drift due to some reason you can also exchange it more easily.

Note: zener current is too high with currently used resistors.
The output voltage pulled down by 200 Ohms may not reach 10V with only 15V supply.
I would place a small capacitor between Pin1 + Pin2 of the Op-Amp U6A to make it more stable for capacitive loads.

And: LM399 with 6860 mV have the lowest T.C. according to experience of some users here in the forum.

with best regards

Andreas

[iMo]

And: LM399 with 6860 mV have the lowest T.C. according to experience of some users here in the forum..

Such a 399 to get is like to win a jackpot (and only when buying used on ebay).
All newer are 7+ Volt. My HP399 (in dmm from 1999) is 7.005V.
The 1399 and ADR1001 are below 7V.

[Birb]

When constructing parallelized LM399 arrays, is it good to separate the array into small subunits (e.g. 8 = 4*2) and have each unit be buffered by an extra op-amp before the average voltage is given to a final gain op amp?
Or is directly feeding the 8 LM399 voltages into one op amp sufficient (and have that op amp, with a transistor drive stage, give the current feedback to the LM399s)

My suggestion,
use 8 independant units with own output buffer and isolated power supply.

The heater currents are rather large (and change with environment temperature) and may influence the output voltage depending on routing of traces.
Star grounding is much easier with completely independant units.
Each unit has floating supply with only one connection between heater GND and zener GND.
Other possibility would be to use negative voltage for heater supply to get the GND currents away from precision GND.

If one unit begins to drift due to some reason you can also exchange it more easily.

Note: zener current is too high with currently used resistors.
The output voltage pulled down by 200 Ohms may not reach 10V with only 15V supply.
I would place a small capacitor between Pin1 + Pin2 of the Op-Amp U6A to make it more stable for capacitive loads.

And: LM399 with 6860 mV have the lowest T.C. according to experience of some users here in the forum.

with best regards

Andreas

Unfortunately cost is a limiting factor, 4 dual zener references (each requiring one op amp and resistor network) is still reasonable.
The zener current is chosen to be ~2.5mA roughly. Does lowering this to 1mA have any benefit?
That said, should the four references be separated onto 4 PCBs and an auxillary PCB handle the averaging + power supply?

Additional note: Would enclosing the 4 reference boards / monolithic reference in an insulated (and shielded?) environment yield better results? The resistor networks should have good TCR tracking and enclosing it in an oven further improves temperature's effect
However this might contribute to accelerated ageing, so idk.

Thanks.

[Andreas]

The zener current is chosen to be ~2.5mA roughly. Does lowering this to 1mA have any benefit?

On LM399 the zener current is always ~0.25 mA. The rest of the current is shunted away from the zener.
There is no benefit going away from the 1.0 mA recommended by the data sheet.

Additional note: Would enclosing the 4 reference boards / monolithic reference in an insulated (and shielded?) environment yield better results?

A wind shield removes low frequency noise from air drafts.
A metallic case distributes/equalizes the environment heat during gradients.
Against EMI you need additional measures to shielding as most of the EMI comes through power supply and output wiring antennas.

with best regards

Andreas

[GigaJoe]

"use 8 independent units with own output buffer "

that , I'm scratching my head now ... i did 3 X  zener - buffer - averaging opamp .... now im thinking ... what benefits of buffer ?
in overall in stationary circuit a current across zener indifferent, over balancing resistor it will distribute accordingly,  more there less there .
assuming one ( few) drifted to awuful 50 microvolts .. that would be fraction of microamp flowing change . and with let say 0.2 Ohm dynamic resistance ...  V = 0.1 µA × 0.2 Ω   = 0.0000001 × 0.2 = 20 nV .   So what the point of buffer rather an additional noise and complexity ?