LM399 based 10 V reference

[Harfner]

but where did you hide the LM399.
In one of the black boxes near the statistical divider?

With best regards

Andreas

The LM399 is most probably hidden in the white cube above the black boxes. The third picture does show the main board from below and there you can see a white cover as well. That is a protection against air currents. 

[chickenHeadKnob]

Quote: "Just another DIY poor man's selfcal DVM & voltage/resistance calibrator ADS1256+AD5971B"

Part number for DAC is AD5791B, and is found here:http://www.analog.com/en/products/digital-to-analog-converters/da-converters/ad5791.html

For those of us that like to play along at home. 

[julian1]

Quote: "Just another DIY poor man's selfcal DVM & voltage/resistance calibrator ADS1256+AD5971B"

Part number for DAC is AD5791B, and is found here:http://www.analog.com/en/products/digital-to-analog-converters/da-converters/ad5791.html

For those of us that like to play along at home. 

The inexpensive one on mouser is AUD $84 and the priciest variant is $323. Do you know what the grade differences are?

[Echo88]

The pricy AD5791SRU-EP operates from -55°C to +125°C, while the standard A/B-grade-AD5791 work from -40°C to +125°C.

Page 17: http://www.analog.com/media/en/technical-documentation/data-sheets/AD5791-EP.pdf
Page 28: http://www.analog.com/media/en/technical-documentation/data-sheets/AD5791.pdf

[branadic]

Just another DIY poor man's selfcal DVM & voltage/resistance calibrator

Pretty nice  Mickle T.

I could today measure my a few years running National Semiconductor LM399H using our new Keysight 3458A with semiautomatic readings (Keysight Interactive IO). With 100 NPLC and within 486s I could read:

min:  10.00173439V
max: 10.00173691V
mean: 10.0017357892213V
std: 517.228098173513nV
peak-peak: 2.52µV

Hopefully in near future fully automatic reading is possible. Also didn't know that there is an extra digit using GPIB interface.

[Edwin G. Pettis]

You do realize that anything past the 6th decimal place in those voltage readings are essentially garbage, no accuracy to speak of?  You realize those calculated amounts you posted are also mostly garbage numbers, assuming that the figures used to calculate them are good to the 6th decimal place, that puts the limit of the calculations well short of the 13 decimal places you display and the std. calculated value isn't really any good past the 2nd or 3rd decimal so why even bother putting them there?  Really, sloppy work as Bob Pease would say.

[MisterDiodes]

You do realize that anything past the 6th decimal place in those voltage readings are essentially garbage, no accuracy to speak of?  You realize those calculated amounts you posted are also mostly garbage numbers, assuming that the figures used to calculate them are good to the 6th decimal place, that puts the limit of the calculations well short of the 13 decimal places you display and the std. calculated value isn't really any good past the 2nd or 3rd decimal so why even bother putting them there?  Really, sloppy work as Bob Pease would say.

Yes, I can just hear Bob Pease about now winding up for another lecture on this subject   I was on the receiving end of a couple of those and learned my lesson.

Even if this device were nulled to a JJ-Array -  is only going to be around +-0.3ppm uncertainty on a good day, or +-3uV on a 10V scale.  It is a good practice to try to keep the digits of precision reasonable and within the limits of measurement resolution, no more; even for intermediate calculations.  Otherwise you'll tend to generate un-beleivable results.

Be careful of that extra magical digit on GPIB on HP / Agilent / Keysights - that is -generally- false noise data and not noted for adding -true- accuracy or resolution to your measure.  You don't get something for free there.  Even Keysight will tell you that.

The 399's can certainly be pretty good performers even though they can be fairly noisy.  Nice devices for 6 digit meters and devices that need power cycling - otherwise their relative high noise can be a problem.

[The Soulman]

The 399's can certainly be pretty good performers even though they can be fairly noisy.  Nice devices for 6 digit meters and devices that need power cycling - otherwise their relative high noise can be a problem.

Ok that seems to be general consensus, noob question: why not L-C filter the heck out of them at the expense of load regulation?

[David Hess]

The 399's can certainly be pretty good performers even though they can be fairly noisy.  Nice devices for 6 digit meters and devices that need power cycling - otherwise their relative high noise can be a problem.

Ok that seems to be general consensus, noob question: why not L-C filter the heck out of them at the expense of load regulation?

The problem is that it is difficult and nontrivial to filter with a low enough corner frequency to remove low frequency noise without introducing more low frequency noise in the form of drift do to the high impedance required.

But I am not sure what specification MisterDiodes is looking at.  Burried zener references like the LM399 are lower noise than bandgap references but all references are pretty noisy.

[Edwin G. Pettis]

I am quite sure MisterDiodes was referring to the LTZ compared to the LM399, the LM399 being somewhat more noisier and yes, while you can accomplish some filtering of a reference's output, it must be done very carefully or it can end up degrading the output's performance.

[Kleinstein]

One can filter the higher frequency noise components like > 100 Hz and this sometimes makes sense (some ADCs convert some of that noise to the relevant low frequency range). However there is not much a classical low pass filter can do about the low frequency (e.g. < 1 Hz) range and this noise is often the dominant part due to a strong 1/f component. The obvious way would be a second LM399 or similar to average, or a reference with a higher current (more current to a LM399 does not help).

There is some sense in using 1 or 2 more digits for intermediate results, but not that many. You gain rather little (minute amount of rounding errors), but it is also essentially no extra effort for a computer. For human readings, those extra readings are more distracting than good.

[branadic]

You do realize that anything past the 6th decimal place in those voltage readings are essentially garbage, no accuracy to speak of?  You realize those calculated amounts you posted are also mostly garbage numbers, assuming that the figures used to calculate them are good to the 6th decimal place, that puts the limit of the calculations well short of the 13 decimal places you display and the std. calculated value isn't really any good past the 2nd or 3rd decimal so why even bother putting them there? Really, sloppy work as Bob Pease would say.

Noone said anything about the accuracy of the readings right? They are within the given 4ppm of the gear. But resolution carries information too.

I'm really happy that you are working accurate and that you are in the position to criticize. I'm sure Bob Pease would be very proud of you. 

[Andreas]

Hmm,

here in the volt nut section we don´t throw away any gear nor any digits 

with best regards

Andreas

P.S: at least the long term stability seems to be very high. Allan deviation goes down after 1 minute.
We should have a longer measurement (perhaps with the connectors covered by a piece of cloth).

[Gyro]

I am quite sure MisterDiodes was referring to the LTZ compared to the LM399, the LM399 being somewhat more noisier and yes, while you can accomplish some filtering of a reference's output, it must be done very carefully or it can end up degrading the output's performance.

Something I've wondered about for a while, is it possible that too low an input impedance filter (ie. low R in an RC filter) could degrade a reference Zener - I guess I'm thinking LTZ rather than LM399 where the Zener is buffered. Given that the breakdown is at least partially avalanche, would having a capacitor more or less directly across the Zener, feed extra energy into the junction as it breaks down noisily, possibly shifting its characteristics over time?

The nearest analogy I can think of is a GM tube, where parallel capacitance across the tube needs to be minimised to avoid shortening its life. A simple Neon bulb oscillator (series R, parallel C) would be another.

I don't think I've ever seen anything written on this in relation to Zeners. Sorry if it's a little OT.

[Kleinstein]

The reference side of the LM399 / LM329 is not just a zener-diode, but a buffered circuit, that internally set the zener current and provides a low impedance output.
This is why there is a limited effect of having a large cap directly in parallel to the LM399 and not much sense in running it with much more than 1 mA. The internal circuit also makes the LM399 less sensitive to changing current than a normal reference zener.

Capacitance directly parallel to a zener is usually not a problem, but also of limited effect due to the low impedance.

Even with the LM399 filtering is difficult, though it can be done to a limited degree. The 0.1 Hz -10 Hz range often used in DS is just much more convenient to measure. But usually it is more the 1 mHz-100mHz range that is really causing trouble.

[Gyro]

Sure, as you say, the 399 is buffered and the Zener current controlled, and it would also be silly to have a very low R in an RC filter.

The question I was asking really relates to the LTZ or discrete Zeners (and therefore getting seriously OT). I was just curious about the effect that low impedance parallel capacitance might have on the physics of Zener / avalanche breakdown (at the microstructure level) and whether it could cause permanent shifts in the breakdown voltage. Probably of academic relevance only -  I just wondered if operating a Zener in the lowest possible capacitance environment could subtly improve its long term drift.

[David Hess]

I am quite sure MisterDiodes was referring to the LTZ compared to the LM399, the LM399 being somewhat more noisier and yes, while you can accomplish some filtering of a reference's output, it must be done very carefully or it can end up degrading the output's performance.

That must be it; I was thinking of in comparison to bandgap references.  The LTZ1000 has an advantage in that the zener operating current can be controlled while other zener references like the LM399 are fixed.  It is easy enough to distinguish bandgap and zener references by their output noise which is handy since manufacturers do not always distinguish them by type in their selection guides; for the later, noise is at least 5 to 10 times higher.

Other than the LTZ1000, do any integrated zener references allow adjustment of the zener current?  Even the 723 regulator fixes its zener current internally.

The question I was asking really relates to the LTZ or discrete Zeners (and therefore getting seriously OT). I was just curious about the effect that low impedance parallel capacitance might have on the physics of Zener / avalanche breakdown (at the microstructure level) and whether it could cause permanent shifts in the breakdown voltage. Probably of academic relevance only -  I just wondered if operating a Zener in the lowest possible capacitance environment could subtly improve its long term drift.

I have never heard of any such effect even though shunt capacitance is sometimes added to reduce susceptibility to EMI.  The junction is irrevocably altered or damaged depending on perspective when initially operated in breakdown so I doubt the lower AC impedance from a shunt capacitor will have any additional effect and the junction's series resistance places a lower limit on the impedance seen by the junction anyway.

[Edwin G. Pettis]

The easiest and most effective way of reducing LM399 noise is to parallel them, up to about five, but then you're getting close to the cost of an LTZ which will have a lot lower noise than even five paralleled LM399s.  So if you really need that low of noise, just get an LTZ, it's easier.  A LM399 is good for about 6-7 digits, with a little extra higher frequency filtering to help keep that last digit from bobbling around too much.  The high frequency Gaussian noise can be filtered some and averaging the output of the Gaussian noise portion will be good enough to clean up the LSD but it will have no effect on the 1/f noise, that cannot be averaged out unless you employ a very long averaging period (no 100 NLPC is no where near long enough for 486 seconds).

The LM399 will likely have achieved its very low drift term after several years of operation, that isn't the problem, it is measuring the DC value hidden underneath all of that noise.  I would have confidence in Mickle T's readings to probably a few PPM; 10.001735 ±.00005 at most, everything after that is essentially junk buried in noise and no amount of math is going to improve that, the actual reading accuracy is the limit and that cannot be changed by any manipulation of numbers.  The uncertainty of the measurement also puts concrete limits on any measurement or calculation, the ±4 PPM accuracy of the 3458A is not the only limit at work here.  There certainly would be less question of the measurement if it was an LTZ1000/A because of the much lower noise present.

[MisterDiodes]

I am quite sure MisterDiodes was referring to the LTZ compared to the LM399, the LM399 being somewhat more noisier and yes, while you can accomplish some filtering of a reference's output, it must be done very carefully or it can end up degrading the output's performance.

Sorry - didn't make that clear.  LM399 Noise spec is 20uV, 0.5ppm TC, drift is typically lower to mid ppm first year, although can improve with age - they are also tolerant of frequent power cycles.  LTZ1000 noise spec is 1.2uV, 0.05ppm TC, drift is typically low ppm per year; usually gets better with time.

You get what you pay for.

If you're building a Vref for ADC or DAC in the 20 bit +  range,usually LTZ is better choice.  '399's can be good performers but noise is an issue - you -could run 5 or more '399's in parallel to reduce noise (more than that is diminishing returns), but at that point might as well run an LTZ to lower parts count & cost.  Trying lowpass filter a precision Vref can be a fool's errand - The best way to combat noise is do everything possible to NOT generate it in the first place.

[Andreas]

Sorry - didn't make that clear.  LM399 Noise spec is 20uV, 0.5ppm TC, drift is typically lower to mid ppm first year, although can improve with age - they are also tolerant of frequent power cycles.  LTZ1000 noise spec is 1.2uV, 0.05ppm TC, drift is typically low ppm per year; usually gets better with time.

You get what you pay for.

Sorry,

but you are comparing apples with pies.

the 20uV(eff) is broadband noise of the LM399. (10Hz - 10kHz).
Whereas the 1.2uV(pp) is 1/f noise (0.1-10Hz).

My LM399 typically are around 4 uVpp (0.1-10Hz).
(some better and some also worse).

with best regards

Andreas

[MisterDiodes]

Hmm,

here in the volt nut section we don´t throw away any gear nor any digits 

with best regards

Andreas

Hmmm indeed - There would be a lot of Professional Engineers / Metrologists / VoltNUts that would debate the issue of fantasy digits with you - but for for good engineering practice it is best to always keep track of your measurement uncertainty, measurement confidence, and keep your results reasonable to your test setup without adding "phantom accuracy" along the way. 

Which by the way is something not up for debate - as a professional engineer, if I were to submit measurement results to an ISO 900x quality control audit team, the very first thing they check is the equipment used to make a measure and basic measurement accuracy, the last calibration date, the next calibration due date, the calibration history, any extrapolated calibration estimates based on past calibration cycles, and so on. 

Then the very next thing that would be examined is any subsequent calculations based on those measurements, and at any point was -any- apparent accuracy increased by the improper use of extra digits - for instance using 13 digits of precision to calculate a measurement made with a 6ppm accurate DMM.  That would be cause for immediate loss of certification right there.

Everything gets boiled down to double-check that a certain testing procedure, if we are delivering a device that operates to some PPM accuracy - really does operate to that level, and is traceable back to NIST at every step.  The whole idea is to cross-check at every step that accuracy is not lost or gained.  High Resolution is not data if accuracy isn't there to match.

That is the only Head's Up there, and absolutely no offense intended to you or Branandic, and I apologize if it was taken that way.  It wasn't meant to.  Keep Building and Have Volt Nut Fun!  But along the way it is helpful if you can be a student of good metrology lab practice, and learn what passes for data - and not - in the real world.

The biggest thing to keep in mind is your initial measurement uncertainty, as that limits the rest of your calculations.  As I wrote in my Price of Chasing PPM thread - the best way to measure a precision Vref is via 732a/b, mechanical null meter and Kelvin Varley divider.  Not only can you do that measure under battery power only, you get your 10Hz bandwidth limiting done for you.  It's much harder to get a good .1 to 10Hz bandwidth measure from a 3458a.  It can be done, but the noise the DMM adds is an issue - besides the fact a 3458a isn't really considered the best voltage transfer device for absolute measures.

And OF COURSE the best way to INCREASE measurement accuracy, DECREASE uncertainty and INCREASE measurement confidence is to acquire more equipment    So I will 100% agree with you there.  More accurate equipment always helps!

Now the mark of a REALLY good Volt Nut is if their life savings is depleted buying lots of 732a/b AND they get buddies to do the same and store everything at VoltNuts house...

[MisterDiodes]

Sorry - didn't make that clear.  LM399 Noise spec is 20uV, 0.5ppm TC, drift is typically lower to mid ppm first year, although can improve with age - they are also tolerant of frequent power cycles.  LTZ1000 noise spec is 1.2uV, 0.05ppm TC, drift is typically low ppm per year; usually gets better with time.

You get what you pay for.

Sorry,

but you are comparing apples with pies.

the 20uV(eff) is broadband noise of the LM399. (10Hz - 10kHz).
Whereas the 1.2uV(pp) is 1/f noise (0.1-10Hz).

My LM399 typically are around 4 uVpp (0.1-10Hz).
(some better and some also worse).

with best regards

Andreas

The comparison still applies:  at 10Hz and below (what you normally need), LTZ is still much lower noise, and at 4uV '399 noise you'd still need 4~ 5ea '399s in parallel to get the noise performance of an LTZ (TC is still 10x worse on '399) - and at that point it's lower BOM count and cost to just build an LTZ in the first place IF THAT is appropriate to the application.  AND you get better TC also.

There are of course times when a '399 will work - and of course use it! But for lowest noise - which is what we're after most of the time - the LTZ's will get you there faster and cheaper a lot of the time, as a practical matter.

No matter what, you still get what you pay for...

[Kleinstein]

To get the noise down from 4 µV to 1.2 µV , that is a little more than a factor of 3, one needs something like 10 or 12 of the LM399. Reduction of noise goes with the square root.

Things can get even more in favor for the LTZ1000, as the LTZ1000 at higher current has a low noise corner and thus the 0.1 -10 Hz value is not yet fully 1/f noise, but still has quite some white noise contribution. So for the often more relevant (but not so easy to measure) 0.01-1 Hz range the LM399 should be at 4 µV again while the LTZ1000 should be below 1 µV, more like 0.5 µV.

If can still find a few (like 4) of the LM129 (essentially LM399 without the heater part), there would be the option to use a few of them together (series or parallel) and add temperature stabilization (to a lower temperature than the LM399) by hand - this way temperature stabilization could also apply to the scaling resistors. Still this needs a lot of testing, compared to the more or less ready made LM399, that needs only little support circuit.

[MisterDiodes]

To get the noise down from 4 µV to 1.2 µV , that is a little more than a factor of 3, one needs something like 10 or 12 of the LM399. Reduction of noise goes with the square root.

Correct.  I was going with the benefit of the doubt, comparing a group of lower noise '399's - which can be as low as a few uV on the 10Hz spectrum if you're lucky.  Really though it is diminishing returns after you get about 4 or 5 '399's in parallel, and averager amp issues, etc.  Plus even with parallel '399's the TC isn't anywhere near as good as LTZ.

[Andreas]

Hello,

Ok now we are near truth for the noise.
LM399 is typically around factor 2-3 more noisy than a LTZ1000.
1.2uVpp typ against 4uVpp typ 1/f noise
and
4uV eff typ against 7uV eff typ for the broadband noise
(4uV taken from the 40nV/sqrt(Hz) from the diagram with bandwidth of 10kHz).

For the T.C. we have still to work on. (have you really measured your T.C.s?)
From my measurements at least near the typical "lab temperature range" my LM399
have around 7-10uV change over a > 30 deg C range. (0.05 ppm/K)

A undadjusted LTZ1000A (without R9) showed 35uV change over a 22 deg C range (0.2 ppm/K)
so not all LTZ1000 have the typical 0.05ppm/K value.

So the price/performance ratio of the LM399 is not too bad.

With best regards

Andreas