Lowest drift, lowest noise voltage reference (ADR1000AHZ)

[branadic]

Hi folks,

after I played with the simplified version of W/F7000 schematic with LTZ and achived -0.0106 ppm/K after t.c. trimming for the bare ovenized zener, I now played with a very new, lowest drift, lowest noise voltage reference I had on the desk. Actually I already had it for some months already, but it always takes some time to get all the components and board design ready.

I treated it the same way as mentioned in the linked thread before, thus first setting the z.t.c. temperature to a specific temperature with the oven turned off. Next step was to turn on the oven and adjusting the oven to that specific z.t.c. temperature.

I then captured some low frequency noise data with a 0.1 ... 10 Hz LNA (80 dB, 100 nVpp or 16 nVrms) and an Advantest R9211E used as a DSO.

Afterwards I've measured the remaining t.c. of the board and trimmed the t.c. in multiple steps. The last trimming was slightly overcompensating things., turning the t.c. into opposite direction. The measurement is already limited by the lab temperature stability and the t.c. of the meter. Thus, it needs another run of adjustment and another run in the thermal chamber plus comparison to F7000-2 in the next run, but results are looking good so far.
Compensating the ambient temperature influence, I can find the t.c. already being at -0.03713 ppm/K, but I'm sure we can do better than that.

The board is currently missing the gain in the boost stage, but uses the amplifier as a buffer only. This is part of the next steps that will be performed soon.

-branadic-

[branadic]

After some tweaking I'm now down to about -10 ppb/K. With some more iterations I'm sure I could get it even smaller, but hey the result is not too bad, or is it?

-branadic-

[Kleinstein]

The temperature effect looks good. However this is the easy part - expecially for a rather limited tempearture range.

The noise performance also looks really good, though the time window is relatively short to really tell. Sometimes popcorn noise happens on a long time scale and some 100 seconds without a jump are well possible even for a LM399. The standard to LF noise charactrization is 0.1 to 10 Hz, because this is a window that can still be measured with AC coupling. However the LF noise that really matters is often the even lower frequency part, like 1-100 mHz. This is especially the case for a precision DC voltage reference. This is really hard to measure - so it is very understandable to measure the sharter time scale. It could still help to have a longer window.

[branadic]

At least in the observed time frame, a t.c. measurement takes about 10h 40min, no popcorn noise was present.
The only real measurement setup for very low frequency noise I know of is propably a JVS and a nV-meter. As the LNA already limits the bandwidth to 0.1 ... 10 Hz, longer time frames with this setup don't make any sense at all.

-branadic-

[Kleinstein]

I know the problem with the very low frequency noise. One could do a measurement with 2 identical reference, but it is quite some effort.

The TC measurements take quite some time, but the vertical resolution is not as good and the reference in the meter is involved too. So hard to tell which ref. is causing the change.

[ramon]

I usually liked to make measurements for 1 month at ambient temperature.

[branadic]

I repeated the measurement with my F7000-2 sitting at room temperature in parallel. I can state with some confidence, that the reference shows no obvious t.c. in the range 15 ... 31 °C. Both, the LDLN reference in the thermal chamber and the F7000-2 at room temperature, track each other.

-branadic-

[branadic]

I recognized that I've never published results of the final reference. After my last post I've trimmed the output voltage to 10.000 0x V and finally trimmed t.c.
The 10 V boost stage added some small higher order portion to the t.c. as already seen on my similar LTZ reference.

-branadic-

[Castorp]

I have taken noise measurements down to lower frequencies, fully in the 1/f region. ADR1000 seems to be at least 2x quieter than good samples of LTZ1000. There's also very little spread among different units. But this is all too preliminary and there are still a number of things that prevent me from publishing the results.

So far I only have two units of HPM7177 with ADR1000, but there will be more soon. What I can tell for sure is that in those units the LF noise at full scale is no longer dominated by the Vref. With LTZ1000 it was, with ADR1000 it's some extra noise coming either from the ADC chip or the AZ op amps (but definitely NOT from the resistor networks). I also have a hint for long-term drift - both units have drifted less than 0.5 ppm in 9 months. They have been powered on and off many times, and one of them was severely abused with EMC tests (conducted RF, EFT, LF magnetic field). The 0.5 ppm/9months was taken with a rock-solid 732B that we calibrate at METAS every year.

I'll mention some of these findings in the coming MM talk.

[Castorp]

Snapshot of an ongoing test. Once again - that's with a 732B and only a small fraction of the LF noise comes from the ADR.

[Dr. Frank]

Hello,
Thanks to ScoobyDoo and branadic, I got 5 samples and some PCBs.
Here's the mockup board, where I made the first tests.

The lower zener voltage compared to the LTZ1000 now allows this Reference Amplifier to be trimmed to zero T.C. for a given oven temperature, by individual selection of the collector resistor R2. Compare this to the +50ppm/K for the unregulated LTZ1000.

That trimming process is the same procedure as for the SZA263 and LTFLU in the different FLUKE instruments.

I measured U_ref vs. temperature, which always gives a negative, flat parable. The zero T.C. point varies from 45 to 60°C for a fixed value of R2, and the T.C. varies from about +5 .. -5ppm/K over 24 ... 80°C only.

At a fixed divider ratio R4:R5 for around 52°C oven temperature for all five samples, the individual T.C.s are well below +/- 2ppm/K.
It follows that the attenuation factors for R3 and the most critical R4, R5 are at least 10 times greater than in the datasheet. These resistors now can have much relaxed stability specifications, only R1 and R2 should be PWW or BMF types.

I also measured the low frequency noise 0.1 .. 10Hz, like branadic and Castorp.
I use an LNA from Andreas, and always made "typical" measurements of 10sec long samples, so to have a direct comparison to my LTZ1000s, but also to the measurements done by Andreas.
100s measurement time or averaging give no essentially different results.
Other LNAs will give different absolute noise figures, so these results are not directly comparable .
 
The noise of the ADRs RefAmp evidently is about 50% of the LTZ1000, which several others also have confirmed.

ADR #1 would be the most quiet sample, but it suffers from  noise, or frequent dips, but with a very moderate amplitude of maybe 400nV_pp.
The ADA4522 buffer contributes a lot to the overall noise, so I have to investigate if it's only this specific unit, or what can be done to improve this.

I will add the noise figures of these samples, when measured with my 3458A, maybe now I can see a pronounced difference.

As my 34465A showed much higher stability / lower noise when I replaced its LM399 reference by a LTZ1000 clone PCB, my idea is to replace the LTZ1000 in my 3458A with an ADR1000 (*).
Maybe its noise figures will also go down, best by a factor of 2.
Castorp found that in his DVM other components overwhelmed the noise budget, so let's see what happens.

I would like to remind that TiN some time ago made a similar experiment on one of his 3458A, by implementing 4 parallel LTZ references, which would also yield 50% noise of a single LTZ.
I'm not aware what his final result was.

Anyhow, I think this new device will require quite a big time slot during the Metrology Meeting 2021, and I'm excited to meet the volt-nuts community personally and we'll have some fruitful discussions and comparisons   

Frank

(*) my 3458A already has a modified FW, with CAL? 2,5 = 6.50000V (7V reference lower limit).
Thanks to Poul-Henning Kamp for your advice how to correct the checksum.   

[Kleinstein]

Directly buffering the reference with an AZ OP can cause problems from input current spike of the OP effecting the reference circuit that uses a relatively slow OP. The usual solution is to have some filtering of the reference voltage. The OP is high input impedance and one can thus use a simple RC low pass, at least to filter out the spikes (MHz range) but also a little more (e.g. 5 K and 1 µF).

I don't think it would be really needed to fine tune the unheated TC. It is enough to start with a typical much smaller value like 5 ppm/K instead of 50 ppm/K. A factor 10 less sensitive resistors is already a big step forward and should be good enough. For the tests it is of cause good to have the unheated TC measured.

Using the ADR1000 in the 3458 may not be so easy, as the voltage is lower and the meter may very well have points to complain about too low a votlage and CAL constants out of range.

[branadic]

Directly buffering the reference with an AZ OP can cause problems from input current spike of the OP effecting the reference circuit that uses a relatively slow OP. The usual solution is to have some filtering of the reference voltage. The OP is high input impedance and one can thus use a simple RC low pass, at least to filter out the spikes (MHz range) but also a little more (e.g. 5 K and 1 µF).

There is some R-C-R filter in front of the AD4522-1, though Frank uses slightly different values than I do

I don't think it would be really needed to fine tune the unheated TC. It is enough to start with a typical much smaller value like 5 ppm/K instead of 50 ppm/K. A factor 10 less sensitive resistors is already a big step forward and should be good enough. For the tests it is of cause good to have the unheated TC measured.

Once the unheated reference is fine tuned the demand for the resistors with heated references decreases dramatically, same as for all t.c. compensated references (LTFLU, LTZ, 1N829, 2DW23x, ...) with only one exception, LMx99 which is more than just a zener. This is true for ADR either way.

-branadic-

[branadic]

I'm constantly monitoring the ADR for 26 days now. Every morning the skript first performs an ACAL, measures the amplified ADR reference with R6581D, the difference to F7000-1 with K2182A-1 and the difference between between F7000-1 and F70002 with K2182A for an hour.
The reference is currently drifting downwards. Allthough some people might think this is due to the resistor network in the boost stage, I tend to believe it's the reference itself. But I don't want to bother the setup and add additional equpiment to prove that.

-branadic-

[ScoobyDoo]

Hello BRANADIC - You have probably observed the LTD graph from its datasheet.

If you power the ADR1000 without an accelerated ageing process - it will take at least 5-6 months before it settles to an acceptable level. It is possible to age the ADR1000 (as done with LTZ) but ageing process needs modification to get it working ... - the rest is proprietary recipe from the Mastercook.

cheers
ScoobyDoo

[Kleinstein]

The curve for aging is with 75 C set point temperature. With a lower temperature the time scale is expected to be even longer.

[branadic]

You have probably observed the LTD graph from its datasheet.

I'm aware of that diagram, but since the recipe of accelerated aging is a secret, it doesn't help a lot to know that there is one. The only recipe available by now for LTZ was given by Cern.

https://xdevs.com/doc/CERN/1101699_V1_Burn_in_of_LTZ1000.doc
https://cal.equipment/doc/CERN/1101699_V1_Burn_in_of_LTZ1000.pdf

Modifying this receipt requires knowledge of the die attach being used and its glass transition temperature, an info that is not available at ADI website.
Aging means cycling around its glass transition temperature to reduce internal strain and allow further curing.

-branadic-

[ScoobyDoo]

Hello Folks - this is our ADR1000 EVB - up and running.

We did some basic STDEV tests using 3458A - but running against limits (of 3458A) - 120nV STDEV (K=2) with an output voltage of 6.659V typical for the ADR1000.

The 120nV STDEV (0.018ppm) is just an indication - we expect it to be better using a setup with a null meter - but this would require an averaged ADR1000 buried zener bank or JJA/JVS.

Best regards
ScoobyDoo

[Kleinstein]

A 2nd ADR1000 refrence and than a low noise meter would be enough. This would measure the sum of the noise of the 2 ADR references plus the noise the meter in a low range (e.g. 100 mV). In the lower range the 3458 could be good enough. It should be good at least for the lower frequency part, as the DMM amplifier has little 1/f noise, but the reference is expected to have 1/f noise in the very low frequency part. A low noise amplifier (no need to be very high impedance) would be an option too.

[dietert1]

Yes, what is the numerical resolution of the HP 3458A in that measurement? 100 nV? If you want to go beyond what you have as resolution or stability of your meter, you setup a difference/bridge type  measurement.

For example with the two LTFLU references i made last year, i am observing a standard deviation of day-to-day averages of about 270 nV. This is the difference of the two references and comparing one day to the next day. So there are two factors sqrt(2) involved and i estimate the noise of each reference to be about 130 nV. The combination of both LTFLUs should be a reference with 90 nV standard deviation for day averages. I am getting similar numbers for hourly averages.

Of course i would also prefer a direct comparison to a much better reference, but that will be expensive.

Regards, Dieter

Edit: The LTFLU references i mentioned include 10 V gain stages, so it's a difference between two 10 V reference voltages.

[Dr. Frank]

Hello,
I've performed the first two stability measurements on the ADR1000, and updated the 1h noise figures as determined with my 3458A in the table above.
It's been placed inside an aluminum case with proper jacks.

The references show a strong negative drift in the first hours already, so confirming branadics findings.
Room temperature and internal temperature of the 3458A was stable to about +/- 0.1°C, so that's fully the ADRs drift what we see here.

I have then chosen 1h of mostly constant output voltage and determined the noise (SDEV) for 1h and for several 10 minute periods.
I have to emphasize again, that these figures are always a convolution (quadratic sum) of the different noise sources like ADR1000, plus 3458A amplifier, plus its LTZ reference and plus the A/D noise.

The 1h noise (185nV for ADR#3, 168nV for ADR#1) is remarkably lower than measurements on other LTZ based DUTs.
Of course it can't be as low as 50%, as the 3458A contributes a lot to the noise budget.

This also confirms the SDEV~ 120nV measurement by ScoobyDoo, which were probably made on a shorter sampling period.
I've even seen 80..90nV r.m.s. over 1min (16 samples).

In the 2nd graph you can recognize quiet and more noisy 10min periods, ranging from 133nV to 237nV, latter showing sort of small bumps.
This seem to indicate different instability mechanisms, the first being the usual zener noise.
So what might be the 2nd mechanism?

Qualitatively, you'll find that picture identically in differential mode measurements, of course on a bit lower level, as only the instabilities of both references come into play. (10V level, 146nV @10 min., 189nV @ 1h)

So it would be interesting to run two ADRs against each other over 24h, to confirm that the noise is cut by 50%, and to investigate, whether these small bumps vanish.
In this case, the ADR could really improve the short term stability performance of the 3458A, when replacing its LTZ1000A.

(I need to change about a dozen further ACAL limit parameters inside the FW, so that the lower reference voltage is accepted.)

I guess I will also bring this mockup board and Andreas LNA to the Metrology Meeting, on next weekend!

Frank

[TiN]

I got some 0.5 ppm jumps from first few chips runs, sadly. In these Log 1, Log 2 two socketed ICs tested starting from power on. Circuit and PCB is very same xDevs FX with 13k/1k Tset and 120R.

Logged by 3 x 3458A + 4th (red line) on 10V boosted output.
DSV-files : Log 1, Log 2
Downward drift also present in logs, as others show.

We had crazy heat here in August, so ambient temperature stability was also very poor.

[Kleinstein]

There seem to be some popcorn / RTN type noise ( sudden jumps by some 0.05 ppm).  With rare event one can get intervals without and intervals with such jumps. I consider these jumps problematic, especially if they are rare, as averaging ober some 10 minutes is not very effective against such noise.
For short time tests this can be OK, but for a long time experiment you never know whether one is more on the upper of lower values.
It is still the question, if this from the ADRs or maybe the LTZ1000 in the 3458. I don't think it is the amplifier or ADC of the 3458 - though at the 0.05 ppm level INL errors may reach this, much and still be in spec. Still the main suspects are the 2 references.

edit:
TiN's cuves show the jump, at least some comig from the DUT - don't expect 3 times 3458 to jump in sync, even with something like EMI.

[quarks]

thanks a lot for sharing information

[MegaVolt]

I seem to have missed the big ADR1000 sale )) Where do you get them?