Binning LM399s

[rodpp]

Hello!

I made a board to check the performance of some LM399 and I would like to hear some suggestion/critics:



That board has 10 LM399H, sharing the same +15V from a bench power supply. The circuit is the simplest from the datasheet, but using the value 6K98 1% for the resistors:



Here are some data already collected:



The temperature in the graphic is calculated from a thermistor (3000ppm/ºC).
The big spikes on the middle coincides with the opening of my lab's window.

Here is the standard deviation:



Looking at these data, it seems that the position of the references in the board influences the performance.


What could be done to improve the performance?
Is this setup ok for only to choose the better ones?
Should I put the board inside a box, to avoid air flows influence?
How many hours is necessary to know the better ones?

Regards,
Rodrigo.

[Ian.M]

That's FUBARed.  They are so close together that mutual heating is inevitable.  Even putting that board in a polystyrene box isn't going to do much to improve the situation.

[rodpp]

Are mutal heating that bad for binning LM 399s?

That TiN LM399 board use 2 LM399 in contact one with another:

https://www.eevblog.com/forum/metrology/lm399-based-10-v-reference/msg703440/#msg703440

I could remove 5 LM399 from the board to have it more spaced. Or make another board with the LM399 in various rows.

[Ian.M]

The ones on the ends are in a significantly different thermal environment than the ones in the middle of the row. 

I'd suggest spacing them out so you can fit a block of polystyrene foam, with a cut out for the LM399 over each of them leaving a gap between the blocks. Make all the polystyrene blocks as identical as possible. Support the assembled board and blocks in the middle of a much larger polystyrene box so the whole board is equally exposed to the airflow that will result from convection within the outer box.  You'll probably also want a couple of temperature sensors, one in the middle and the other right at a corner so you can estimate the temperature differential across the test jig.

[rodpp]

Thank you for the sugestions. It will be a nice experiment.

Certainly there is a temperature gradient across the board surface, and it will interfere on the references output voltages. But, supposing a constant gradient, the performance should not be affected. If one reference is, for example, 2C above another one, but that difference remais constant, it should not interfere on the comparison results (standard deviaton).

[dietert1]

I would wrap the whole board into a towel and exclude any records of spikes. You want to run ovenized references in an environment without or with little air movement. For example all our HP 3456A voltmeters have a 3D printed cover over their reference modules, in addition to the cover the LM399 comes with. Then standard deviation on minutes to hours time scales drops to about 0.1 ppm or even less for a good LM399. Of course that isn't something easy to check/prove. I don't know what multiplexer and voltmeter you are using.

Regards, Dieter

[rodpp]

I will wrap the board into a towel, thanks for the suggestion.

To reach that sigma=0.1ppm I think that is necessary much aging hours. These are brand new, or at least NOS LM399s.

The setup I'm using is:

Power supply: Agilent E3631A
Multiplexer: Keithley 706 Scanner with 7064 Low Voltage Scanner Card
Voltmeter: Keithley 2002

[TimFox]

If you don't mind re-doing the Vectorboard, you could re-arrange the sockets into a regular polygon to reduce the inequality in their thermal environments.

[rodpp]

Yes, Ian.M suggested something similar too.

A circular pattern would be the best. Or put the LM399s forming a very sparse matrix.

I did that unique line of LM399s to avoid air wire, as all traces could be done only using solder. And didn't use sockets to avoid contact problems.

[rodpp]

Here are the data from a period of about 9 hours, without splikes:



The better reference with sigma=0.33ppm.

I wrapped the board into a towel, maybe that pattern of decreasing sigma with the position will change. I don't know if it is temperature related.

[Badwolf]

Hello

Unless you have something that produces a continuous air flow from the LM 399 number 1 to the 10, I don't think the temperature difference between all of them is that important.

Moreover we can see a gradient from 1 to 10, so I would think for a power supply / current problem.
Can you reverse your power supply and put it at the opposite side?
Even if the currents are weak, there can be interactions with this type of card

[dietert1]

So you are almost there. 0.33 ppm of 6.9 V is only 2.3 uV. So another thing to watch out would be thermal voltages at the scanner input connection and in the link from scanner to voltmeter. To check that i would wire a short that i can measure with the voltmeter through the scanner so see how stable the residual voltage is. Or if you have two spare channels on each end of your "1" .. "10" lineup, you could try and see whether the "1" end is worse.

Another consideration would be EMI interference: using shielded cables, adding blocking capacitors on both heater and reference output of each LM399. Andreas wrote up lots of tests, including the LM399:
https://www.eevblog.com/forum/metrology/emi-measurements-of-a-volt-nut/msg2684070/#msg2684070

Regards, Dieter

[bsdphk]

"The ones on the ends are in a significantly different thermal environment than the ones in the middle of the row.  "

Mount them in a circle ?

But still:  They shouldn't be in free air, cover them up with something which does not touch and does not leave enough free space above for the thermals to get too strong.

[Andreas]

Some general rules for LM399 (since your results are still much too "noisy")

- use a well stabilized power supply (also for the heater).
- best PSRR is given with heater voltages >= 15V
- use bypass capacitors 100nF directly at the zener (+ heater) pins
- shield pins from air flow (not only on the top but also on the bottom side of the PCB).
- do not tilt the LM399 this may give several PPMs deviation. (see pictures here)
  https://www.eevblog.com/forum/metrology/project-pimp-a-keithley-2000/msg1106829/#msg1106829
- for highest stability use LM399s with a zener voltage near 6875 mV @1 mA (gives lowest TC between heater off and on)
- keep EMI sources (LED lamps, SMPS, WIFI, Mobiles) away
- and of course you should give the LM399s minimum 14 days run in time before doing measurements.

with best regards

Andreas

[maat]

Typically, the first thing, that you should be looking out for, when binning LM399 is popcorn noise. This is rather a rather frequent problem. For my university project I was binning LM399s and I had to throw our 10 out of 25. (All but 4 came straight from Mouser).

Attached I have a log of one the Zeners:

To add to the confusion, I tested each diode in an amplifier circuit, which amplified the Zener voltage by (20/17+1) to about 15V. Since my 34470A does not have a 20 V range, I was using another known good Reference at 15 V (tested against a Fluke 5440B) and measuring the differential voltage. So keep in mind, that the (white) noise is roughly 2.2*sqrt(2) of a single Diode and the popcorn noise of the zener has to be divided by the gain (~2.2). The data was recorded by a 34470A in a stable enviroment (< 0.5 K) using 10 PLC and  Auto Zero = On on the 1 V range. The Zeners were burnt in for 100 h prior to testing.

So you will looking for popcorn noise on the order of 2 µV for your 6.95 V reference.

This is what a good example looked like:


Once you have weeded out the worst offenders, burn in the the rest for a few weeks, then start monitoring.

Another option is to use a low noise amplifier and have a look at the 1/f noise of each diode.

[Andreas]

How many hours is necessary to know the better ones?

many: there may be some "popcorn noise" which shows up only from time to time
see comparison of 2 LM399s here (Ch7 has large initial ageing whereas CH6 has up to 2 ppm popcorn noise)

https://www.eevblog.com/forum/metrology/lm399-based-10-v-reference/msg796829/#msg796829

with best regards

Andreas

[Andreas]

This is what a good example looked like:

shure that your instrument has no popcorn noise?
You should have at least 2 instruments to decide wether the pocorn comes from the zener or one of the instruments

It took very long time until I recognized that my K2000 has sometimes popcorn noise too.

with best regards

Andreas

[maat]

shure that your instrument has no popcorn noise?

Pretty sure (Of course I can never be sure). I measured the first sample against the Fluke 5440B for a week, both with a 3458A and the 34470A (and for the fun with a DMM6500 to see if it could do the job  ) The noise is fairly typical for the LM399. Any dips you see is most likely other people in the lab.

To further reduce the impact of the DMM it was running on the 1 V range. So unless the front end amp is popping its corn I doubt that I would even see any popcorn noise at all using this setup.

Anyway here is a weeks sample of which I have many more:


I also screened the op amps on the test board for popcorn noise and as I couldn't believe the number of diodes I was throwing out, I also measured the 1/f noise of both good and bad diodes. There is a fairly distinct difference. Measuring the 1/f noise actually gives the fastest results (within hours), but makes for less dramatic pictures in a thesis, so I took the time-domain approach.

[bsdphk]

"- do not tilt the LM399 this may give several PPMs deviation. (see pictures here)"

Isn't that just the plastic-housing touching the can and the leads at different points depending on the direction of the force of gravity ?

[rodpp]

Wow! Thanks for all contribution! I'll address each one.

But before that, I would like to show the results after the @dietert1 towel suggestion. It improved the results significantly, probable reaching my setup capabilities.

Here are the best one hour of data collection after wrap the board into a towel:



The best reference now with sigma=0.06ppm, or 414nV.

Note that the temperature is not stabilized yet.

The specification of the Keithley 7064 Low-Voltage Scanner Card that I'm using says:

"Contact Pontential: < 1uV differential voltage, input to output with copper leads (<200nV typical < 1 minute actuation)."

Another observation is that there is no more a pattern where the sigma decreases with the position, the [performance x LM399 position in the board] is more aleatory now.

The graphic above is of a very limited data set, only 6 samples. But looking at all data collected, it is very clear the difference after the holy towel:

[rodpp]

Hello

Unless you have something that produces a continuous air flow from the LM 399 number 1 to the 10, I don't think the temperature difference between all of them is that important.

Moreover we can see a gradient from 1 to 10, so I would think for a power supply / current problem.
Can you reverse your power supply and put it at the opposite side?
Even if the currents are weak, there can be interactions with this type of card

Yes, it is very easy to put the power supply at the opposite side. It is one more good test to do.

So you are almost there. 0.33 ppm of 6.9 V is only 2.3 uV. So another thing to watch out would be thermal voltages at the scanner input connection and in the link from scanner to voltmeter. To check that i would wire a short that i can measure with the voltmeter through the scanner so see how stable the residual voltage is. Or if you have two spare channels on each end of your "1" .. "10" lineup, you could try and see whether the "1" end is worse.

Another consideration would be EMI interference: using shielded cables, adding blocking capacitors on both heater and reference output of each LM399. Andreas wrote up lots of tests, including the LM399:
https://www.eevblog.com/forum/metrology/emi-measurements-of-a-volt-nut/msg2684070/#msg2684070

Regards, Dieter

Yes, it is possible to check the scanner channels. The board I'm using has 20 channels and I'm using only 11 (10 LM399 and the thermistor). I connected the odd channels to the LM399s and the even channel to ground (except the last one, connected to the thermistor). Another test to include in the to do list...

And thanks for the link, I'll read it.

Some general rules for LM399 (since your results are still much too "noisy")

- use a well stabilized power supply (also for the heater).
- best PSRR is given with heater voltages >= 15V
- use bypass capacitors 100nF directly at the zener (+ heater) pins
- shield pins from air flow (not only on the top but also on the bottom side of the PCB).
- do not tilt the LM399 this may give several PPMs deviation. (see pictures here)
  https://www.eevblog.com/forum/metrology/project-pimp-a-keithley-2000/msg1106829/#msg1106829
- for highest stability use LM399s with a zener voltage near 6875 mV @1 mA (gives lowest TC between heater off and on)
- keep EMI sources (LED lamps, SMPS, WIFI, Mobiles) away
- and of course you should give the LM399s minimum 14 days run in time before doing measurements.

with best regards

Andreas

The power supply is direct from the Agilent E3631A to the board green borne. See below the last 24h of the power supply voltage measurement, it's from 14.993-14.998:



I'll check the noise with the oscilloscope.

I'm using Vcc = 15V.

Yes, I'll add the capacitors. One more in the to do list.

I'll observe all your considerations to improve this test.

Regarding the run time before doing measurements, today is the second day that that references are running. The LM399s I'm using are new ones, never used before. So, it seems that I must wait some days before any conclusion.

Typically, the first thing, that you should be looking out for, when binning LM399 is popcorn noise. This is rather a rather frequent problem. For my university project I was binning LM399s and I had to throw our 10 out of 25. (All but 4 came straight from Mouser).

Attached I have a log of one the Zeners:

--PICTURE--

To add to the confusion, I tested each diode in an amplifier circuit, which amplified the Zener voltage by (20/17+1) to about 15V. Since my 34470A does not have a 20 V range, I was using another known good Reference at 15 V (tested against a Fluke 5440B) and measuring the differential voltage. So keep in mind, that the (white) noise is roughly 2.2*sqrt(2) of a single Diode and the popcorn noise of the zener has to be divided by the gain (~2.2). The data was recorded by a 34470A in a stable enviroment (< 0.5 K) using 10 PLC and  Auto Zero = On on the 1 V range. The Zeners were burnt in for 100 h prior to testing.

So you will looking for popcorn noise on the order of 2 µV for your 6.95 V reference.

This is what a good example looked like:

---PICTURE----

Once you have weeded out the worst offenders, burn in the the rest for a few weeks, then start monitoring.

Another option is to use a low noise amplifier and have a look at the 1/f noise of each diode.

I'll pay attention in the noise. I'm using NPLC = 10, Auto zeron ON and filtering 10 samples (sequential, not moving average).

How many hours is necessary to know the better ones?

many: there may be some "popcorn noise" which shows up only from time to time
see comparison of 2 LM399s here (Ch7 has large initial ageing whereas CH6 has up to 2 ppm popcorn noise)

https://www.eevblog.com/forum/metrology/lm399-based-10-v-reference/msg796829/#msg796829

with best regards

Andreas

Interesting graphic, looking at the graphic scale (total of almost 2 years) the CH6 popcorn noise seems somehow sporadic.

[David Hess]

Interesting graphic, looking at the graphic scale (total of almost 2 years) the CH6 popcorn noise seems somehow sporadic.

Popcorn noise can be sporadic on a scale of hours to days.  It took me a month to track it down in a Tektronix made integrated circuit.

Since it is related to processing, I am surprised that any modern analog IC would suffer from it.

[KK6IL]

A few days ago, I set up a crude breadboard to burn in some LM399's, 3 board pulls from China, and 4 NOS parts I got as samples back in the dark ages, National date code 630 (1976?).
One of the NOS parts gave erratic readings on a 3456 set for 6 digits, 100 cycles, jumping from 7.10373 to 7.10394 from one reading to the next. Photos of the noise viewed on Tek 7834 with 7A22 plug-in, 100 uV/division, not the 1 mV reported by the readout. Lots of 50 uV short spikes - occasional long duration changes. I wouldn't have been  surprised if one of the China parts was bad, but wouldn't have expected it from a factory sample. Anyone ever seen a problem like this? Any possibility it will get better with aging?

I moved the part from the far right end of the pegboard to the far right, replacing the wires and 7.5K resistor (15V power supply) - no difference.

Yes, a crude setup - way beyond help from a towel. I just wanted to get a few weeks aging on the parts, and see how much the voltage changed.

John

[David Hess]

One of the NOS parts gave erratic readings on a 3456 set for 6 digits, 100 cycles, jumping from 7.10373 to 7.10394 from one reading to the next. Photos of the noise viewed on Tek 7834 with 7A22 plug-in, 100 uV/division, not the 1 mV reported by the readout. Lots of 50 uV short spikes - occasional long duration changes. I wouldn't have been  surprised if one of the China parts was bad, but wouldn't have expected it from a factory sample. Anyone ever seen a problem like this? Any possibility it will get better with aging?

It looks like popcorn noise to me but make sure it is not in the 7A22 or 7834.  The usual giveaway is that the amplitude shifts by a common amount, but for variable duration and timing.

The photograph below shows the popcorn noise that I tracked down to a channel switch in my Tektronix 2230.  It sure looks digital and that is what I initially thought but it is a very analog process.  It was absent for days at a time which made tracking it down very difficult.

I would not count on it getting better.

[Kleinstein]

Interesting graphic, looking at the graphic scale (total of almost 2 years) the CH6 popcorn noise seems somehow sporadic.

Popcorn noise can be sporadic on a scale of hours to days.  It took me a month to track it down in a Tektronix made integrated circuit.

Since it is related to processing, I am surprised that any modern analog IC would suffer from it.

The popcorn noise is usually attributed to single quantum states (e.g. electron occupation at some defects or impurities) that have a relatively large impact because they sit a some sensitive point. One would not note most of the defects with only a weak effect, but they can still contribute to 1/f noise.
While impurities can be reduced, the other prime candidate are defects at the oxide interface, that may not be avoidable. With smaller structures there may be less defects, but more of these can have a large effect.