Ultra Precision Reference LTZ1000

[Kleinstein]

Doubling the drift rate every 10 K of higher temperature is a rough rule of thumb for many thermally activated processes (also works for biological processes like milk getting old). However I am not to sure if it actually 10 K for the LTZ1000, it could be 7 K or 12 K for doubling. In addition the initial phase can be a superposition of effects going up and down and partially compensating. So it might be a good idea to have some burn in, so that only the slowest process is dominating. After that there should be drift in one direction only if there are not other processes coming in (like radiation damage, hysteresis form temperature cycle).

A low temperature also slows down this initial burn in - so it is a 2 sided thing to go for only 45 C. So it might even be better to do a 2 (or more) step burn in: first at a higher temperature and only after that at the final temperature.

As there is a certain chance to get a bad reference, it is difficult to give an upper limit on drift / stability without an individual test / burn in. The chances are low, but there are a few bad ones, that are not detected at the manufacturer. So without an individual test, there will be a few outliers. Not sure about the fractions accepted by DMM manufacturers, but I would expect tighter specs there. The lower temperature will only improve long time drift, but not the chance for rare outliers. It might even make it slower to find a bad one and it slows down the initial burn in.

Even with tested good samples, it really helps to wait for after the initial burn in. Especially the very first part (e.g. first 100 hours)  is not that well defined. Often testing would include that initial burn in phase anyway.

Chances get much better to guarantee a certain limit if 2 refs are used and compared. With 2 refs one should be able to detect any sudden jumps (that will not happen in both) and it is also very unlikely that 2 units will have the same unusually high drift rate. So those rare bad units could be detected with a good certainty at least (one still does not know which unit is bad), but at least is warned to need recalibration.

Comparing 3 units over a reasonable time should give an indication if there a bad ref and which one i would likely be. However there is little to tell about the final long term drift (e.g. years 2-..), as chances are that the units will show a similar long term drift. Only with different temperatures there might be something, but likely not enough statistics.

[ap]

The ball park 10K for an increase of 2 in aging is driven essentially by the Arrhenius law. Details on the web. However, re. the Fluke 7000, the drift is not only given by the LTZ1000, but also by the gain-defining amplifiers in the amp stage. Vishay non-hermetic resistors have been used here, and these seem to be higher drift than the hermetic foil or WW resistors one could use (and Fluke used elsewhere on their non-Wavetek references) , even though due to statistical (distribution of drift rates over several resistors) the demonsstrated drift is less than one would expect from a single resistor of the type used (the spec value is actually not that good; there are better non-hermetic resistors available than the one used; on paper at least).

[rigrunner]

My LTZ has been powered on for > 350 hours now so time to log some data and see how it is doing.
Logging is with Datron 1271 on 10V range set to 8.5 digits resolution, one measurement taken every 30 seconds.
LTZ is powered at 15V from a basic 7815 supply.

[chuckb]

LTZ1000 non A chip microscope images

I had an old LTZ1000 chip that I knew I would not use so I sacrificed if for the advancement of our knowledge.

It uses a similar air entrained epoxy as the LTZ1000A. This must be a low stress attachment. I believe DiligentMinds noted that the LTZ1000A has an extra thermal interface material under the chip.

In the images you can see that the LTZ1000A sits higher above the TO-5 header than the LTZ1000 chip.
The chip was very clean and we had a good microscope so we took a few images.
There is a little detail inside the Super Zener symbol I had not noticed before.

[chuckb]

More fun with a microscope and a digital camera.
LT1088
The LT1088 True RMS sensor uses a similar chip to the LTZ1000. The hermetic package will open cleanly with a 200 watt soldering iron applied to the lid.
The LT1088 operation is detailed in Linear Technology Application note 22.

With the lid off you can clearly see the two chips with the OTT die attachment process.
Extra credit for the person who identifies the very special lens used for the “LT1088 hermetic package.jpg” image.
The die has heater rings that are similar to the LTZ1000. It uses 4 of the 8 transistors (diodes) in the middle of the die to sense die temperature. It does not look like there is a Zener in the middle of the die.

[Muxr]

Great photos. I want that machine they use for attaching pins to the die. Imagine the dead bug possibilities!

[chuckb]

If you used a 16 pin ceramic package similar to this you would have space for an LTZ1000 die (if you could get one), a LT1013 die and a Vishay MTR hybrid resistor network. Now your entire LTZ1000 system is in a hermetically sealed package. It would only cost you a few thousand per package! And you could regulate the package temperature with a TEC for minimum power.

But then there would be nothing left to tweak. What's the fun in that?

[branadic]

This ceramic packages are available of the shelf, no problem. LT1013-Die is available at least at Texas Instruments... LTZ1000 is somewhat more difficult to get as bare die at least in small quantities. So maybe you have to buy a complete wafer.
However, this is only half of the truth. The more expensive part is the assembly, the die attachement and the developement of the processes and their installations. Nothing you would do with manuel pick and place apparatus or bonding machines.
At the end of the process you need to solder the metal cap to the ceramic package using vacuum and/or inert gas and AuSn solder preforms. Additional costs are for the tools.

Conclusion: Interesting way to go, but not worth the money you have to invest unless you have a big budget and no other idea how to spend it.

[TiN]

Thank you, chuckb, for great LT1088 teardown. Probably they used same or similar process when making LTZs, LT1088s. I wanted to play with that chip for a while already, but not managed to find resources yet.

Hermetic packages for LTZ ref, even while it make no financial sense (probably running cost of such project is more than getting LHe PVJS), at least we can have great fun speculating about it.

On more practical note, I travelled with one of my LTZ1000CH based modules to a friend, and reference output of the module was within -0.8ppm of my measurement average after 1 month of 24/7 operation. Measurement was taken by friend's 3458A, which I got repaired and calibrated to my references last year November. So even with quick setup it was reassuring to see <2ppm data from different references on term of 8 months. But the interesting thing was once I connected Fluke 8846A parallel to 3458A and LTZ output, readings jumped to down -16ppm! All usual settings were done on meter, manual 10V range, NPLC100, High-Z input impedance. Even switching 8846A input terminals to rear (essentially disconnecting LTZ from input) still corrupted output at -11ppm. Huh? Only hard power off with rear mechanical switch "fixed" the output back to expected level.

Connecting 34401A in parallel with exact same cables powered from same mains breaker did not change LTZ output even a ppm. Same with my K2002, no effect. Only 8846A caused this. LTZ module was powered by battery at all times. So another thing to watch out hooking multiple DMMs to direct zener output without buffering.

[TiN]

Volt-nut candies has arrived.

In case somebody forgot how genuine LTZ reference looks like, here the close-ups.



Bonus shot, compared to LTZ1000CH from 1990:



P.S. No problems ordering just 2pcs qty each type directly from LTC website

[chuckb]

Extra credit for the person who identifies the very special lens used for the “LT1088 hermetic package.jpg” image.

The 1X in the lower right corner was supplied by the 200mm f/5.6 Medical Nikkor lens. It's a slick lens with a built in ring flash.

[TiN]

Extra credit for the person who identifies the very special lens used for the “LT1088 hermetic package.jpg” image.

The 1X in the lower right corner was supplied by the 200mm f/5.6 Medical Nikkor lens. It's a slick lens with a built in ring flash.

That's a nice lens.  I was looking to buy one few years ago, but they still fetch pretty penny, so I stay happy with my Sigma 150mm f/2.8 macro.

[borghese]

I'm going to make my first LTZ based board. Has anyone
tried the thermal performance with board thickness 32mil and copper weight of 2 oz? What is your opinion?
Best regards

[Kleinstein]

There is a limited advantage in having a thicker Cu in getting the ring more effective. However it would need even thinner lines to bring the signals out.  Usually the layout should be in a way that wire / board trace resistance should not be significant.

[d-smes]

Does anyone ground the metal can of the LTZ? 

[babysitter]

I have guard pour on my board, however, its not connected to the can.

[schmitt trigger]

Chuck;
thanks for the naked-IC-porn photos! 

It definitively shows that you have substantial experience in the field.

To me these photos epitomize the Linear Tech spirit: an extreme attention to the most minute details, to obtain the very best analog performance.

[d-smes]

I had stability problems with my LTZ1000 build so I thought I'd do some long-term data logging and see if it settled down.  I initially decided to monitor it with both meter and data logger.  Here are the results.  Both set for 100 PLC.  Temperature is from a LM35 sensor that sits 2cm above the aluminum case that holds the LTZ1000 (and provides ~1.5K additional heating above true ambient).   All I can say is Keysight sure doesn't make meters like HP did!

[julian1]

Good information. I love my 34970a.

[d-smes]

I guess this is a good example of what people have said in this thread:  You can't use a meter with a '399 voltage reference to measure the performance of a LTZ1000.   I'm just surprised the 34465a made such a good thermometer!  Even when temperature is stable between the 16 - 21 hour mark, there's a great deal of noise that made me wonder if my LTZ1000 was operating correctly.  But in view of the 34970a measurements, it seems fine.

Do others with 6.5 digit meters experience similar issues?

[Dr. Frank]

I guess this is a good example of what people have said in this thread:  You can't use a meter with a '399 voltage reference to measure the performance of a LTZ1000.   I'm just surprised the 34465a made such a good thermometer!  Even when temperature is stable between the 16 - 21 hour mark, there's a great deal of noise that made me wonder if my LTZ1000 was operating correctly.  But in view of the 34970a measurements, it seems fine.

Do others with 6.5 digit meters experience similar issues?

What issue do you mean?
I think you let yourself fool by the different visual appearances of both graphs, which are at first differently formatted, e.g. the thicker line width for the 34970A.
Then, both instruments feature different resolutions over GPIB, i.e. 100nV for the 34465A, and 1uV only for the 34970A.
If you look closer at the short-term noise, the 34465A varies about 3uVpp, and so does the 34390A. The finer resolution of the 465A gives a more fluttery appearance.

Also, the time resolution of the 34490A measurement seems to be much lower, than the 34465As, maybe there's an additional running-average filter, which additionally reduces the apparent noise.

I made a comparison between the 34465A, 34470A and the 3458A, using a stable 10V reference (5442A), equivalent to the LTZ1000:
https://www.eevblog.com/forum/testgear/keysight's-new-34465a-(6-5-digit)-and-34470a-(7-5-digit)-bench-multimeters/msg889217/#msg889217

which indicates, that concerning noise, the 34470A performs similar to the 3458A, only due to the LTZ1000A based references, whereas the 34465A (which has identical hardware as the '470A, otherwise), shows more noise due to its LM399 reference.

Latter is valid for all LM399 based DMMs, as the 34401A and your 34390, which share the identical DMM circuit.
The 34465A/470A are very similar to, or based on the 34410/411 design. (Same Multislope IV architecture, many identical crucial components)

Therefore, the Keysight instrument are very probably made as good as under the former brands.


In the end, you made a trivial conclusion, that you can't really evaluate an LTZ1000 by an LM399 based instrument, that applies to to your 34970A as well.

You did not describe, how you have set up your instruments. The temperature sensor measures the LTZ1000s temperature, but obviously neither the real room temperature, nor the instruments individual temperatures.
Probably, you have stacked them, 34465A on top of the 34970A, so that the 465A is heated additionally by the 970A.
Both instruments are specified to have a T.C. of about 5ppm/°C. So the change of 10uV or 1.4 ppm for the 34465A, in face of the 2°C change RT (?) change, or much more, if you really stacked the instruments, is very well inside specification.

That might explain the seemingly better temperature stability of the 34970A.
Therefore you should have a better / more temperature stable experimental setup, before you draw the wrong conclusions.

Btw.: I also measured the 16h stability and noise of my 34401A, using the same 10V reference. That looks quite similar to your measurements on the 34970A:
https://www.eevblog.com/forum/testgear/hp34401-measurement-of-linearity/msg358701/#msg358701

Frank

[HighVoltage]

I had stability problems with my LTZ1000 build so I thought I'd do some long-term data logging and see if it settled down.  I initially decided to monitor it with both meter and data logger.  Here are the results.  Both set for 100 PLC.  Temperature is from a LM35 sensor that sits 2cm above the aluminum case that holds the LTZ1000 (and provides ~1.5K additional heating above true ambient).   All I can say is Keysight sure doesn't make meters like HP did!

What cables did you use to connect the 34465A and the 34970A
Picture please of your setup?

[d-smes]

Thank you Dr. Frank for your comments.
There are indeed two data sets acquired simultaneously and later imported and manipulated within Excel.  34465A data was free-running (auto-zero ON, NPLC = 100) while the 34970A Data Acquisition unit was scanning ~30 channels every two minutes with a two second delay after each relay switch event.  34970A also set to 100 NPLC.  Data records may also have ~60 seconds of time skew as the meters were started and stopped manually and normalized to the same time scale within Excel.  And yes, I did thin the 34465A lines and add a data point so I could could try and understand if 34465 apparent finer resolution was an artifact of internal meter filtering or true ADC step size.

I agree I was somewhat fooled by the higher sample rate and resolution of the 34465A.  During the period of steady temperature, I took this to be purely noise whereas the 34970A data looked to be the expected bit toggling when the measured value is near the threshold of an ADC step; superimposed noise being on order of 1 uV.

By way of background, my initial LTZ1000 build was with LTC2057 choppers and I had noise, drift, and jumps in output voltage.  Adding capacitors all over the place seemed to help, but when I'd run simulations with the added capacitors, loop stability was compromised to the point where I couldn't understand why physical hardware didn't break into oscillation.  I then replaced LTC2057's with LT1677's (I didn't know about the LT1006 at the time).  That made a huge difference in noise but I still had the drift and jumps in output voltage like the 34465A plot I originally posted.

I was therefore astonished when I collected the 34970A data-  Voltage jumps are gone.  Voltage drift highly correlated with temperature gone.   Perhaps I drew the wrong conclusion?  Attached are pictures of meter setup and connections to LTZ.  34970A connections are soldered to shielded Cat-6 cables (1 nF capacitor feed-thru's into boxs - there are two LTZ's side-by-side).  Temperature sensor is black TO-92 over left box.  34465A are EX-hook to RG-58 coax in Test Setup photo.   I originally stacked the 34465A on top of the 34970A with a 5 cm space between, but every now and then I'd get voltage "spikes" in the 34465A data which I correlated to relay switching events within the 34970A, so I moved them apart.

[Dr. Frank]

Thank you Dr. Frank for your comments.
There are indeed two data sets acquired simultaneously and later imported and manipulated within Excel.  34465A data was free-running (auto-zero ON, NPLC = 100) while the 34970A Data Acquisition unit was scanning ~30 channels every two minutes with a two second delay after each relay switch event.  34970A also set to 100 NPLC.  Data records may also have ~60 seconds of time skew as the meters were started and stopped manually and normalized to the same time scale within Excel.  And yes, I did thin the 34465A lines and add a data point so I could could try and understand if 34465 apparent finer resolution was an artifact of internal meter filtering or true ADC step size.

I agree I was somewhat fooled by the higher sample rate and resolution of the 34465A.  During the period of steady temperature, I took this to be purely noise whereas the 34970A data looked to be the expected bit toggling when the measured value is near the threshold of an ADC step; superimposed noise being on order of 1 uV.

By way of background, my initial LTZ1000 build was with LTC2057 choppers and I had noise, drift, and jumps in output voltage.  Adding capacitors all over the place seemed to help, but when I'd run simulations with the added capacitors, loop stability was compromised to the point where I couldn't understand why physical hardware didn't break into oscillation.  I then replaced LTC2057's with LT1677's (I didn't know about the LT1006 at the time).  That made a huge difference in noise but I still had the drift and jumps in output voltage like the 34465A plot I originally posted.

I was therefore astonished when I collected the 34970A data-  Voltage jumps are gone.  Voltage drift highly correlated with temperature gone.   Perhaps I drew the wrong conclusion?  Attached are pictures of meter setup and connections to LTZ.  34970A connections are soldered to shielded Cat-6 cables (1 nF capacitor feed-thru's into boxs - there are two LTZ's side-by-side).  Temperature sensor is black TO-92 over left box.  34465A are EX-hook to RG-58 coax in Test Setup photo.   I originally stacked the 34465A on top of the 34970A with a 5 cm space between, but every now and then I'd get voltage "spikes" in the 34465A data which I correlated to relay switching events within the 34970A, so I moved them apart.

Sorry to say, but your setup is really a mess, especially in aspect of low thermal voltage connections.
No wonder you're measuring this temperature dependent voltage on the 34465A, that is probably not the instruments T.C., but the e.m.f. of these totally inappropriate clamps, in comparison to the twisted pair soldered Cu cables which are used for the 34970A. You may at first also use soldered, twisted pair Cu cable for connection the 34465A.
These feed-through capacitors used as the voltage output of the reference are also not the best pick for low e.m.f.

Please replace all of that stuff with good CuTe jacks, where you have the choice of connecting with good cables (Cu with Au plated connectors) towards the 34465A, or clamping blank Cu cables with the jacks' screws.
I even had big problems on my analogue / DC precision measurements with my (10 times more stable) 3458A, when room temperature changed a few °C, so I moved the whole lab down to the basement.
Only there, at about 0.1°C stable environmental temperature, I was able to make measurements in the sub-ppm region, as you tried to do here, also.

Check for switch mode PSUs in your room, these mostly create these spikes you're seeing.
Use the additional capacitors from Andreas design, which requires an additional bodge at one point for stability.
 Adding only two 100nF capacitors, in parallel to the 120 and 1k precision resistor already improved my old LTZ reference greatly.

The shielding by your boxes looks nicely, but is probably not effective, as the external disturbances enter either via the supply, or the output. Both sides are not appropriately connected to any guard, as your PSU, as well as your DMMs don't have a guard. Maybe connecting to ground may help a bit, but in the end, that shielding is mostly useless, compared to the effects of e.mf. generation.

So I recommend that you at first completely revise your assembly before attempting to make precision measurements.

Frank

[Andreas]

 Adding only two 100nF capacitors, in parallel to the 120 and 1k precision resistor already improved my old LTZ reference greatly.

Hello,

of course you get the best benefit of those 2 capacitors when soldering directly to the corresponding (base+emitter) pins of the LTZ1000.

with best regards

Andreas