Lowest drift, lowest noise voltage reference (ADR1000AHZ)

[Kleinstein]

For the 7 to 10 V step only about 1/3 depends on the resitors. So a 1 ppm drift in the 10 V translates to 3 ppm change for the resistor ratio. The averaging over 2 / 4 resistors each also helps. Leaving out 2 resistors out of the 8 would allow to skip on the outer ones that seem to be ones that can be a bit differente at least for the NOMC.

[Andreas]

Also interesting to see the difference between buffered and scaled output.

Hello,

I made a additional chart with the difference of ppm drift between the 6V6 and 10V output of ADR1000A#1
(so effectively the ratio drift of the statistical divider)
Up to now it is a rather linear drift over time.
But if I compare it with PT1 filtered rH% values of my room data logger it could also correlate to rH%.

From my LT1027 references with DIP8 package I know that the influence from rH% needs several days (more than 5 days) to correlate with voltage output drift. So I used a PT1 filter with 1 week time constant. The values are logged every 15 minutes.

with best regards

Andreas

[Andreas]

Hello,

update on ageing drift of my ADR1000A#01 + #02 now at 7 kHrs
see also:
https://www.eevblog.com/forum/metrology/lowest-drift-lowest-noise-voltage-reference/msg4547027/#msg4547027

still a relative large drift against a typical LTZ1000.

Humidity is now decreasing (a little) and the 10V output divider drift is slowing down somewhat.
But I fear at the moment that the humidity is not the main factor for the output divider drift.

with best regards

Andreas

[openloop]

Decided to check the "low noise" aspect of my ADR1000 references.
See how they stack-up against my LTZ1000s.
Conditions: Two independently battery powered reference boxes connected to measure voltage difference, HP34420A, 10mV range, 100 NPLC (@60HZ), 4 hours of readings, no filters.

Notice the change in vertical scale.

[branadic]

I passed 13000 hours, attached an update.

-branadic-

[Dr. Frank]

Here's an update of the drift measurement of my ADR1000 #3.
After 1 1/2 years of operation, the drift is slowing down, but still at a rate of about -2ppm/year.
I observed temporary hysteresis of about +2ppm when I shut off the references for some time, in around June 2022.
I shut it off for 8h and then measured the recovery over 17h.
Its voltage returned to its former value within less than 0.2ppm, so I assume that this hysteretic effect might be not so pronounced anymore.

Frank

[miro123]

Marco reps posed youtube video

[iMo]

Btw, - as seen in the above video - burn-in with 140C at 6 o'clock position and then using it at 12 or 3 o'clock position - is that a good idea?
PS: a pity ADI is so slow with the release of the ADR1001, as it would be much simpler and cheaper exercise, imho..

[Andreas]

hmm,

I fear I will not have the time to view the looong video. (i have to watch my references for ageing).
Can someone summarize the quint essence?

BTW I have found some github data:
https://github.com/marcoreps/ADRmu

with best regards

Andreas

[Kleinstein]

I fear I will not have the time to view the looong video. (i have to watch my references for ageing).
Can someone summarize the quint essence?

For a lower TC without the heater active it seems to help to have smaller collector resistors (e.g. 30-50 K range instead of 61 K from the datasheet example circuit) than in the data sheet.
This does no come to a surprise - it was known that the zero TC point is at higher current (for the zener or transistors).
The main point was that the reference boards mostly seem to work well (noise in specs and not much visibe drift, one sample with popcorn noise).
Preaging at 140 seems to cause at least initially more drift - so likely not a good idea.  I think others made a similare experience with aging at a rather high temperature.
The DCDC converter as some positive aspects:  isolation via a simple though relatively large toroidial transformer with relatively large spacing between the windings and the core. No details on the performance shown and I think there may be room for improvements ( at least a partial electrostatic shield, reduction in stray field).

[miro123]

The DCDC converter as some positive aspects:  isolation via a simple though relatively large toroidial transformer with relatively large spacing between the windings and the core. No details on the performance shown and I think there may be room for improvements ( at least a partial electrostatic shield, reduction in stray field).

I question this approach nowadays.
1. The battery technology improved last 50 years.
  - capacity/volume went up
  - reliability went up too - long term stability. number of cycles no memory effect just to name few.
  - BMS made last 50 years huge leap too
  - price went down - e.g. I've bought 100Ah LiFePO4 cells - each has the cost of single LTZ1000
2. Typical use case of Vref - 99,99% Idle/aging  + 0.01% measurements.
3. Cost - BOM and effort to build own dc-dc vs COTS battery backup

[Kleinstein]

Li cells are well performing, but there are regulations on transport, that can make them a hassel. NiMH cells are not that great an alternative and a bit tricky when it comes to charging.

There is one possible alternative: use 2 sets of batteries and relays to switch over. One set is used to power the device and the 2nd set can be charged with an automatic (low glitch) switch over if needed.  So no more transformer in a DC/DC, but 4 relay contracts to set the isolation. The battery to case part can still add quite some capacity - so look out for a low capacity battery  .
Another point could be the variable heat production during a charge cycle. So it is not that ideal too.

Those references are usually made to be run 24/7. So a battery only solution is not really an option. It is nice to have the battery option though. So use mains power most of the time and battery operation for the really critical measurments and on transit if one wants to keep it hot all time.
A good DCDC would allow to keep it mains powered for more of the less critical tests (e.g. 24/7 difference data to a 2nd ref).

[miro123]

Humidity is now decreasing (a little) and the 10V output divider drift is slowing down somewhat.

you can apply correlation function to fanout any dependency, delay times and trends. That said does statistical methods works better with higher amount of sample datapoints.

[Dr. Frank]

Anyhow, it's a pity that Marco does not join this community to share his evaluations, although he evidently has used knowledge especially from this thread. 

[DavidKo]

Andreas: I have watched it in the train to work (2 trips required) 

The PCB should be universal for LTZ and ADR (ready for critical divider from normal resistors, resistor array or Vishay), it should have output current protection and lost power indicator. Both old and new refences are evaluated. What I have found interesting is T.C. measurement and tuning of resistors. During the resistor tuning the whole setup is in external oven to include the T.C. of the whole setup, but at the end he uses the internal heater. In video is said, that the github data should be updated to see further development of values.

Design is on 6 layer PCB. The cap for reference is glued - I expect that screwing will be better due to the different temperature expansion of FR4 and plastic.

[Kleinstein]

I don't know the plastics used for the cap, but most types are relatively compliant compared to the FR4 board. Ideally the cap should not be massive plasitcs, more with small voids, as sometimes used with 3D printed parts to save material. Not sure how it is really made.  The force from the cap should be relative moderate. In addition the temperature is expected to be relatively constant, partially regulated by the reference.  I would have preferred to glue the cap only at 3 local spots to allow even more compliance. Not sure that srews would be much better.

Tunig the ref. TC for the whole circuit in one step is a bit odd:  much of the reference TC will in the final configuration be suppressed by the oven part. The final gain part would still be 100 % effective.
So it does not help to do the test for the full circuit.

It should still be OK to do the trim this way: the final gain stage needs a very good TC (e.g.  < 2 ppm/K preferrably better). This way it would not really interfere with the trim of the unheated reference that does not need to be that good. Even the initial about 20 ppm/K range is OK and 5 ppm/K well good enough.  One could use an alternative way by keeping the environment temperature constant and just add a disturbance to the set point divider. So change the oven temperature and choose R2 (e.g. 30 K to 100 K range) for low sensitivty to the set point divider.  I would consider a jumper from the set point divider to a resistor (e.g. 100 K range) and trimmer (or external control voltage) for that test to shift the set point. This way no extra temperature controled box is needed.
The test in the variable temperature box with the heater enabled would mainly test the final gain stage.

[branadic]

First a quick comment about the history of the ADR1000. What we've learned from Eric Modica at MM2021 was:

"... I had just taken over a project to second source the LTZ1000, which you all know. And this was obviously before the merger with Linear Technology. The primary reason ADI started to look at second sourcing the LTZ1000 was because, one of our VP saw an ADI AppNote that used the LTZ1000 as a reference for a 20 bit DAC and that was the AD5791. So by the time I took over the project, there had been a couple of test chips to look for various candidates for internal processes to use at ADI to match such a part..."

So in my understanding the goal was not to create an update or improved successor, but something similar to the LTZ1000 and by "...We got lucky?..." the final reference showed lower noise.

Given the charts from this thread in the video showing measurements after 8000 hours it must have taken Marco quite some time to create his video, almost a year.

A few technical comments. What I could extract from the video is that Marcos references show quite large noise compare to other measurements done before, which makes me wonder why. A possible source could be the DC/DC converter and to verify that powering the reference circuit from batteries could help. Also the measurement seems to be performed without protection from ambient noise, here is where your lovely cookie box comes in handy.

Furthermore, he build a reference with an early engineering sample with datecode 1727. As far as I understand this is a sample without the later introduced PMO which is a 150 °C, 168h both bake and burn, which Eric explained to be the cause for jumps that appear like popcorn noise.

"... what I’ve observed is that bake and burn is helpful primarily for low frequency noise reduction in this context. We have had tier 1 manufacturers tell us the same thing when we send them ADR1000s. Prior to instituting a bake and burn most of the parts look okay from a low frequency noise perspective, but if you look at it enough you find one that is extremly noisy and actually remedies with the bake and burn. So only being a process hobbiest myself, the only explaination I can give there is that you have these dangling hydrongene bonds at the surface of the SI and sometimes that amounts to surface charge and if that ends up in the vacinity of the zener, if you are unlucky enough, then this is the kind of thing that you get. And it’s really the only thing that would move at such low temperatures where you are running a bake and burn, say like a 150 °C..."

Nice to see Marco adopted some of the concepts from W/F7000, that were proven to work with results shown here on eevblog too, such as zero t.c. adjustment. However, some of the t.c. curves make me wonder as I have never observed any kink in my curves when sweeping through the temperatures.
It would be nice if he would present his results not in some special Python or Grafana format, but as charts with proper image file extension. Finally, it remains interesting to see more longterm results from more specimen in general, so I'm curious to see how his samples will do in the future.

-branadic-

[miro123]

A few technical comments. What I could extract from the video is that Marcos references show quite large noise compare to other measurements done before, which makes me wonder why. A possible source could be the DC/DC converter and to verify that powering the reference circuit from batteries could help. Also the measurement seems to be performed without protection from ambient noise, here is where your lovely cookie box comes in handy.

I'm going to experiment with ZVS/ ZCS half resonant converters.  I'll post the result in separate tread. they will be together with transformer bode100 plots - just to avoid the lucky shots in inductors winidings

It would be nice if he would present his results not in some special Python or Grafana format, but as charts with proper image file extension. Finally, it remains interesting to see more longterm results from more specimen in general, so I'm curious to see how his samples will do in the future.

I tend to disagree here. My reasons are listed below
1. Giving access to raw data allows other engineers to analyzed it in make own conslusions.
2. Files are in jupyther notebook format. De facto standard in EE world. I think this is essential  EE skill nowadays - like handling soldering iron, Ltspice matlab and CAD product.
e.g. jpyther screenshots shows few wrong assumptions made by Marco. Source https://github.com/marcoreps/ADRmu/blob/main/results/ADRmu3_TC.ipynb

[branadic]

I agree that access to the raw data is nice for sanity check by other, I disagree though giving the raw data only in a very speciallized format. It is unscientific to provide data and just say: "analyze them on your own". It would be way better to provide them in a way so that everyone understands what you want to present and that is a plot/chart/image. You can feel free to provide the raw data additionally.
I don't care whether or not jupyther notebook format is a de facto standard, you simply don't put such files into a review paper and tell the reviewers to plot whatever they like to see themself, do you?
And what if at some point the de facto standard changes? Hence I prefer raw data in plain text files, csv format or similar, which is platform independent and readable "For All Times, For All Peoples".

-branadic-

[Andreas]

Hello,

from the diagram you can clearly see that 16 deg/C per hour is much too fast.
The delay between reference output and Temperature is around half an hour.

with best regards

Andreas

[iMo]

Hello,
from the diagram you can clearly see that 16 deg/C per hour is much too fast.
The delay between reference output and Temperature is around half an hour..

That would require 10x longer sweep time, imho. Also 0.024ppm/C with the components used seem to be too good to be true..

[Andreas]

Hello,

update on ageing drift of my ADR1000A#01 + #02 now at 8 kHrs
see also:
https://www.eevblog.com/forum/metrology/lowest-drift-lowest-noise-voltage-reference/msg4623766/#msg4623766

Now at -6 ppm reference drift over 8 kHrs or -0.75 ppm/kHr in average.

Humidity is now decreasing but the 10V output divider drift is still rising.
Obviously the drift over time (ageing) is the main factor.

with best regards

Andreas

[branadic]

I passed 14000 hours.

-branadic-

[iMo]

While looking at your above pictures - that reminds me on the first ever measurement I did with myADR1001#1 - see below (from the thread on ADR1001). I ran the chip with temperature set to something like 115-125C (it was not my intention, btw). The measurement started perhaps 1 hour after first ever power up and took aprox 13 hours.
Almost the same data and shape as your long term measurement above. Interesting..
Could it be you may speed up the burn-in with higher set temperatures?

[branadic]

That is unlikely. Correlation doesn't necessarily mean causation. You can't speed up multiple thousands of hours to just a few tens of hours by higher temperature.

-branadic-