[FT] xDevs.com KX LTZ1000A fairy-tale, or the story of little jumper.

[TiN]

Let's start the new segment, fairy-tales of metrology 

Intro

One of engineers doing precision measurement projects at home for own enjoyment contacted me while ago, asking some questions related to nanovolt-meter amplifier front-end and voltage references. His country of residence is Belarus, and I mention this because their customs are really strict about international shipments. Basically everything over 20 $USD mark is troublesome to import and due to high fees, often much more than package value itself. So his instrumentation desk was mostly using some old instruments, made in Soviet Union. We had nice talk about electrometer of his, which is an interesting bit of gear on its own.

On DMM front he has Datron Model 1271, 7½-digit DMM, smaller brother of Model 1281 which our other contributor, Todd, had covered well before. Quickly talk was biased towards voltage references. Buying famous Linear LTZ1000 was not an easy option due to shipping/customs troubles, even worse considering existing lack of access to calibrators or metrology labs nearby. Even though this 1271 is fine instrument, it was not recently calibrated and absolute accuracy was unknown. Since this situation was very familiar for me few years ago, I decided to give a hand and help this engineer to “import” verified DCV standard into his home hobby lab. Agreement to build one KX LTZ1000 voltage reference was made.

Buy some parts, put on PCB with less than 30 components, measure the output and send the board thru mail. Sounds easy job, no rocket science, right?

Parts source and assembly

To make this whole little project more interesting for me, I wanted to try two new ideas, which were not attempted before in previous builds.

*A.* Try 1K?/13K? hermetic resistor network (custom Vishay VHD200, oil-filled +/-3ppm/K two-resistor network) with 0.05% matching. This might give some interesting data, if tightly coupled oven setting thermal point resistors help to improve tempco of the reference.

*B.* Try alternative chopper amplifier, Analog Devices ADA4522-1. Usually I use LTC2057, but both op-amps are rather overkill for this particular circuit, as noise of LTZ1000 zeners itself is magnitudes higher than opamp contribution. But maybe other issues pop up with AD chip. Let's see.

As a base my latest Rev.B02 version of KX LTZ1000 PCB is used. This module 4-layer FR4 ENIG-plated PCB use two single op-amps in SO-8 package, accepts +10-20VDC input

Before soldering resistors down into circuit, I did a quick measurement to check that they don't give some wild deviations.

Custom VPG 120 ?, R3 , -0.8 ppm - +26.4 C, 0.36 ppm/K
VPG VHD200 1 K?, R4 , -75 ppm - +24.0 C   2.5 ppm/K
VPG VHD200 13 K?, R5 , -5 ppm - +24.3 C   1.4ppm/K

First small SMD parts were soldered on, and checked. Used LTZ1000ACH chip, manufactured 36 week of 1993 was soldered on, with maximum soldering time around 0.5 second per pin, to reduce excessive heating. Foil precision resistors were populated last. Assembled module ready for the initial test.

Before we jump in to actual test data, it's important to have good confidence in measurement equipment and lab tools, to make sure that any unexpected phenomenon comes from device under test, not the equipment itself. This also means that even brand new, freshly calibrated equipment must be verified and tested before any solid conclusions to be made.

To satisfy this requirement, I had chance to calibrate my older LTZ1000ACH/LTZ1000CH based references against 15-day calibrated Fluke 732B DC Voltage standard in mid-August 2016. Two Keithley 2002 8.5-digit multimeters we calibrated to same Fluke reference same time, to act as transfer standards and guard-banding tools. I used these DIY standards and DMMs to transfer absolute DCV voltage into my verified and improved HP 3458A, which confirmed to be stable by daily 24/7 operation and monthly checks to both hot and cold external LTZ1000A modules. As a result, I have high confidence that absolute DCV accuracy in my home lab is within +/-2 ppm.

Now with verified gear, initial test conditions are fairly simple. Purpose of first test is to make sure of reference normal operation, and measure initial voltage, drift if any and stability. Output should not deviate more than +/-1 ppm over few days period, and temperature coefficient (change of output voltage from change of ambient temperature) should not exceed +/- 0.1 ppm/K.

Power supply : 12VDC 2A cheap mains power brick.
Measurement  : HP 3458A, calibrated August 31 to +/-2ppm DCV.
Ambient condition : Ambient room temp with aircon, +/-3 C. Monitor by BME280 sensor
Datalogging : Raspberry Pi with linux-gpib + python + NI GPIB-USB-HS
Analysis : D3.js plotting library on xDevs.com site

Test 1

Yes, it's alive, measuring output voltage +7.067329 VDC +/-2 ppm.

And how we looking here? If this would be integrated reference, such as band-gap reference on internal DAC/ADC references, result noise 1.1 ppm could be simply miraculous. However for top-end ovenized buried zeners, which for Linear LTZ1000 is extremely bad. To remind you specifications, this circuit should be able to provide 0.05 ppm/K drift rate and approximate noise levels of 1.2 uVpk-pk. We are getting 7.4 uV instead for noise. Tempco is also horrible, about 1 ppm/K, 20 times worse than expected!

Test 2: TC correction resistor add

For second test we work on mistake and try to address issues on data we saw earlier.

* Improve stability of ambient temperature, so we can isolate TC issue for now.
* Test with 394K R2 populated, even though we use LTZ1000ACH and this compensation not needed.

Some of our readers not familiar with LTZ1000 performance may scream at me right now, how you know that this is bad result. Vertical ppm scale (left green scale is +/-1ppm, which is 0.0001% accuracy!) is blown out to make this amazing reference look bad, you say?

Alright, I hear you. And here's the answer, same condition measurements with good and checked LTZ1000CH voltage reference module, using same schematic and same PCB, just different parts. Horizontal time span of graph below is 11 hours.

Reference test result on another LTZ1000CH KX module

See, noise of this good module is mere +/-0.1ppm, no random jumps or other weird crap, everything nice and stable. That's how you want it, not jumpy-jumpy . Of course if you reduce your accuracy demands to 0.002%, everything looks nice even with jumps.

But it does not make sense to pay 250$ USD for such accuracy, so let's continue our journey.

Test 3: Reduction of oven temperature test

Alright, how about reduce LTZ1000ACH oven temperature inside chip, by changing 13K? resistance on divider to lower value? Let's see

Fluke 300K? -1 ppm/K wirewound resistor was soldered in parallel with 13K? on board, and test was repeated.

Test 4: Op-amp kelvin connection test

Here I had cut separate trace between Pin 2 U5 opamp and R6 resistor, and jump-wired connected guard ring point directly at pin 4 of LTZ1000ACH chip. This makes almost Kelvin-like connection for all inverting node connections. Why I did so? Hm, intuition?

Jumpie-jumpie is still there. Noise of LTZ chip noise itself is riding on top of those jumps. Overall it looks nice and stable, just the pesky jumps, which is pink-noise of zener junction. Output voltage changed from +7.064808 to +7.066329 VDC, which is +215 ppm.

Test 5: Replacement of LTZ1000 chip

Now, if we dig a bit into physics, this jumps are nothing but pink-noise, likely caused by zener effect breakdown is due to quick bursts of electrons going thru thin junction on high voltage. This breakdown is just about 7 volts, and that's close to out near 7.06V output of zener inside this LTZ1000ACH. So it does look that no external component tweaks could fix this... Let's see on test result of TEST RUN 5:

t's good now! What is the magic? There is none, actually, just engineering. Now we getting some solid data. Noise over multiple hours is just around 0.2ppm, which translates into around 1.2 uV[~pk-pk~], initial tempco evaluation from +26.6  C to +30.5  C unable to reveal any tempco clearly, no large jumps either.

Bonus

I also built thermal box to test tempco, so jump into full article link down below

Total amount of time for this project ~25 hours, including ~50 hours of datalog time by HP 3458A.

Hope this data helps to better understand hidden icebergs and dangers big enough to sink a project in high-precision field, even for something as simple as direct voltage reference module. It's not enough to buy fancy parts and plonk them on the board to get 1 ppm accuracy.

I have it running tempco test now, from 20C to 55C in 0.1C per hour step.

Full article

If you have own unexpected story, do tell us. 

Important update, calibration offer by community

One of our members, mimmus78 ordered KX LTZ1000 PCBs and sent some to other interested tinkerers at cost, so to support this great initiative, additional "calibration service" is provided by me, plesa and Alex Nikitin to support this.

Rules to participate are simple:

STEP 1. Build your KX-module, publish short worklog in this thread (few photos of assembled PCBA, photo of it running with DMM hooked up showing 7V)
STEP 2. Leave it running for 200 hours, so initial drifts get stabilized. Module should be enclosed in some box without vents.
STEP 3. Measure it again, record the temperature and voltage and put a label with values on the box
STEP 4. Ship the box to me (Taiwan) or other member (Europe) after agreement via EEVBlog PM. You pay shipping.
STEP 5. I'll power your box with +15V for 48 hours to have everything stabilize
STEP 6. Test 7V output voltage in temperature span +20....+30C, record the graph
STEP 7. Unit will be shipped back (from me via EMS Express (~30-40$USD, 4-7 days to USA, bit more to EU)). You pay shipping.

All steps are mandatory, to make sure that reference is good and stable, otherwise it would be no much point in ppm-level calibration, if initial aging drifts are not removed first.
This is standing offer, without schedule or specific date deadlines, unless my homelab changes or other force majeure events occur.

Even though I don't have official calibration performed on my equipment, thanks to volt-nut community and multiple cross-checks I'm confident that my DCV accuracy is around ±2 ppm.
And to make this even better, few other EEVBlog members also backed up this initiative to provide same calibration offer in their regions.

Calibration by	Accuracy / Standard	Region/Country	Contact for participants
xDevs.com	<2ppm / 3458A-mod + 2*K2002	Asia/Taiwan	EEVBlog forum PM
plesa	<4ppm / 3458A-002	Europe/Sweden	EEVBlog forum PM
Alex Nikitin	<4ppm / 3458A-002	Europe/UK	EEVBlog forum PM

Article was also updated for same information

[lukier]

Interesting story.

I would thought that such expensive parts go through rather stricter testing (than TL431 for example). I remember in your 3458A article the original reference board from eBay also had jumpy LTZ1000.

Is this jumpy LTZ1000 brand new or second hand from eBay?

Slightly worrying, as I would like to build a LTZ1000 reference at some point, but no way I'm buying a dozen chips or so to find one working to the specs 

[Nuno_pt]

This jumpy LTZ was the one from the A9 board that TiN bought on flebay for is 3458.

The resolution is in the full article in TiN's website, he then replace the LTZ with another one and no more jumps.

[TiN]

lukier

I think you get answers from article . If it's more comforting, all 6 LTZ chips I bought directly from Linear site are not jumpy. 

[Nuno_pt]

To make this whole little project more interesting for me, I wanted to try two new ideas, which were not attempted before in previous builds.

*A.* Try 1K?/13K? hermetic resistor network (custom Vishay VHD200, oil-filled +/-3ppm/K two-resistor network) with 0.05% matching. This might give some interesting data, if tightly coupled oven setting thermal point resistors help to improve tempco of the reference.

TiN do you notice any diference yet, with the exchange of the voltage divider in relation to previous method, of two independent resistors?
This way the tracking of Tempco should be better since they are inside the same encapsulation.

BTW, in a conversation with Edwin Pettis he told me, that the Engineers from LT told him that for the last few years they recommend the all LTZ's to be drive by 15K/1K, but this was already cover in the LTZ thread .

[lukier]

I think you get answers from article . If it's more comforting, all 6 LTZ chips I bought directly from Linear site are not jumpy. 

Oh got it I thought that's yet another unit, so I worried about LTZ1000 reliability.

[zlymex]

Good story and nice photos/charts. Slightly worrying for me too because I did not test all of my 2nd hand LTZ1ks fully.

TiN is the thermistor hermetic? It seems to be very stable over time.

[TiN]

zlymex
It's teflon coated. You can see photo of it in the article. Overall it does look more stable than my previously used Honeywell HEL-705 Pt RTDs (on same K2510 setup). I had to increase current for RTD at 833uA compared to just 100uA on this YSI 44031.

[doktor pyta]

One time I was repairing Data Precision 8200 DC calibrator.
I did many measurements to prevent good old precision components from heating during soldering.
Unfortunately the problem was LM399 (see plot).
Maybe it is a problem with contact resistance inside IC.
I assume that Data Precision would not release a product with such issue, so it is very likely that this LM399 broke down during normal use.

[VintageNut]

I have been monitoring one of my Fluke 731B for a few days continuously with a DMM7510 on 15 NPLC with averaging of 10. In the morning the measured voltage is 9.999961 and late evening the voltage is 9.999978. This cycle is consistent from day to day. At first I thought it was temperature but after some colder weather followed by warmer weather, I think the cycle is humidity.

I have no humidity measurement equipment now so that has to be on the list to obtain sooner than later.

[Nuno_pt]

For humidity, pressure and temperature I think that the BME280 is hard to beat. And you can do like TiN, sample it together with your measurements via GPIB. 

[zlymex]

I have been monitoring one of my Fluke 731B for a few days continuously with a DMM7510 on 15 NPLC with averaging of 10. In the morning the measured voltage is 9.999961 and late evening the voltage is 9.999978. This cycle is consistent from day to day. At first I thought it was temperature but after some colder weather followed by warmer weather, I think the cycle is humidity.

Then you need a third equipment to judge, either a DMM or a voltage reference,

[TiN]

Or send that 7510 to us for check  3458 vs 7510... I think that's not equal fight   

[Kleinstein]

The Fluke731 and DMM7510 should have a similar class reference inside. So it could be either one that is doing the drift.
The reference in the 3458 is not better than that, and near full scale is could be very well the reference that sets the limitation. So 3458 vs DMM7510 would be LTZ1000 vs LTFLU so this may come down to binning or individual quality.

[VintageNut]

Good points Tin. I will connect my model 2000 in parallel with the 7510 and let it run for a few days to see if the drift is detectable on the 2000 as well.

I am reasonable certain that the drift is the 731B but only proper measurements and analysis of the data will make it certain.

I accept the challenge to find out for sure.

Do you collect your data for this type of analysis with Python?

[Andreas]

lukier

I think you get answers from article . If it's more comforting, all 6 LTZ chips I bought directly from Linear site are not jumpy. 

Hello Illya,

You have to look harder on the noise.
My LTZ#5 has relative large jumps of typical 1uV up to 1.8uV.

The others have also jumps (below the 1.2uVpp) which are only visible when zooming into the record.
E.g. LTZ3 with a 300nV jump.

All my LTZs are sourced via distributor or DigiKey.
So even if they are costly you have to monitor them for popcorn noise.
The noise screening tests at the manufacturer typical use a very short time window.
(<< 10 seconds)
So the detection of popcorn noise is usually not possible.

With best regards

Andreas

[Theboel]

I am sorry Ilya and others if this not relevant but anybody has experience about same jump/popcorn noise from LM399 ?

[d-smes]

BTW, in a conversation with Edwin Pettis he told me, that the Engineers from LT told him that for the last few years they recommend the all LTZ's to be drive by 15K/1K, but this was already cover in the LTZ thread .

Is this new info?  I thought you only wanted to run them that hot for the "A" version in high ambient environment.   Can you link to where this recommendation was discussed?  I skimmed through the LTZ1000 thread but couldn't find it...

[TiN]

The choice is up to application. Lower noise at cost of long term stability - run high temp. Better long term stability (thats what most people here after) at cost of noise - reduced temp.

[TiN]

Also I realized something this morning during my ride to work.
Since now I have bad, but functional LTZ1000A chip, which I'd not really use in voltage reference builds, it opens a road to second fairy-tale...

We don't really care if it dies, right? Now if I can find LTZ-set resistors which I can sacrifice without breaking a bank, we can make one special module to run cryotest.
That is, enclose module into metal box with weak thermal coupling and submerge everything into fresh 26L dewar with LN2, without heater circuit on.
It could answer question I was asking myself for years, if LTZ1000 at 77K can deliver better than 0.1ppm noise. If it's positive, then that would open new box of pandora making cooled LTZ-references  .
LN2 is also cheap, so that experiment could be repeated by other volt-nuts if proven worthwhile (for example to do nV LNA testing, when lowest noise possible is a help).

Why need sacrifice resistors? Because they are unlikely to be very good and stable after 230K thermal shock.

[VintageNut]

Have you placed one of these reference boards in an ordinary freezer at 0 deg F to test the idea?

[TiN]

I'd expect difference at 0 would be just down in measurement errors and uncertainty.

[VintageNut]

I'd expect difference at 0 would be just down in measurement errors and uncertainty.

Does the noise reduction only happen at some magic very low temperature or is the noise reduction a linear function of temperature?

[TiN]

Delta deviation in ppm per each °C is shown on graph.
Also for easier grasp, here's summary data on chamber temperature span from +16 °C to +31 °C (temperature stability +/-0.005 °C). Perhaps bit easier on table below with calculated temperature coefficient and output deviation delta in ppm units. Voltage reading at +16.000 °C was taken as a baseline.

As we can see, temperature coefficient stays very close to +/-0.05 ppm/K specification. Also one has to remember than this test is essentially a comparison of internal HP 3458A LTZ1000-reference versus external. So accuracy of such measurements is already +/-0.1 ppm at best. To perform better measurement, one would need measurement equipment more stable than old aged HP 3458A, which leaves very few options, like very expensive JJA systems or characterized bank of Fluke 732A/732B/734A DC standards and low-noise null-meter.



RAW samples data in DSV-format is available as well, if you reader desire to conduct own analysis.

DSV-data log with temperature chamber test

I still have it running to raise temperature up to 50 °C, so will update data after test samples arrive

[Cerebus]

Because the Zener in the LM399 is operated at such a low current [no matter how much current you pump into the device], you will see this "popcorn noise" more often with that part.

This is worth emphasising because a lot of people miss this. Internal to the LM399 is an, internally regulated, constant current supply that pushes 250 uA through the zener. Any excess current is shunted off by a pair of transistors that just turn the excess into heat. This doesn't mean that you can get away with feeding the LM399 with an unregulated current but you won't gain anything by going over the recommended 1 mA.