Ultra Precision Reference LTZ1000

[e61_phil]

What happens to the stdev/pp ratio for real (shorter) measurement series?

For example in my current observation of two LTFLU 10 V references i see a stdev = 0,029 ppm of the daily averages, so times 6 gives an expected pp of 0,174 ppm. After 38 days i have an observed pp of 0,11 ppm (minimum -0,065 ppm, maximum 0,046 ppm). So that gives a ratio of about 1.6 (overestimation of pp by 6x formula). In this case the factor seems to be more like 4 instead of 6.

Maybe this case is not a random walk, but something else. Maybe air pressure changes.

Regards, Dieter

Maybe it was a bit misleading to say "Seems to me, that there is no need to take values lower than 6 to estimate the pp noise" (Therefore, I deleted that sentence in the post above). What I wanted to say was: 6 is already conservative. It is most likely less than 6x std. dev. (6 / 1.38 = 4.34)

[mrflibble]

I played with some simulated random walks. To evaluate "how wrong" 6x the standard deviation compared to the "real" peak to peak value is, I created 1000 random walks. Every walk was a million steps long. I calculated the standard deviation and the peak to peak value for every walk and compared them.

Quick note regarding peak-to-peak value of a simulated random walk.
You may also want to try:
- do random walk of N steps, storing each location x[n] at step number n.
- sort all x[n]
- for maximum, take sorted list entry at index floor(0.95*N)
- for minimum, take sorted list entry at index floor(0.05*N)

Notation: x[n]is location at step number n, assuming start position x[0]=0.
Adjust the (0.05, 0.95) constants to taste.

Similar story for the mean value. Since you sorted that list already, you might as well take the median, so x[ floor(N/2) ].

PS: request from those of us who do find this sort of thing interesting, but are a but strapped for time... I did not read the entire thread today, so I'm sorry if I missed the context that would have made it obvious... But: for walk_example_1 and walk_example_2, I couldn't get the information of what it was. I.e, is this a single trial of N=1E6 steps. (yes). Or is this the sum of all M=1E3 trial runs. (No). To be sure it was a single trial the image filenames actually were more informative. That  and eyeballing the number of bins ~ 50. And eyeballing the area of the histograms .... "mmmh, if I put this blob here and that blob there I get an about level line around the count of ... 20000. Okay. About 50 * about 20000 equals about 1E6. Check, it is a single trial. Or you could just put that information in the plot title. Just a suggestion. I am as guilty of forgetting to do this as the next person. Reason I started doing proper annotation, titles and whatnot even in my own projects that never see the light of day: yeeeaaaars from now my future self will thank my current self for making it easy to follow what the hell I was doing at the time.

Oh yeah, random other note: (PPS? ) If you want to solve for the expectation value of the (min,max) values of a random walk you can use the infinity-norm of the x-vector, where again x[0] is start, and X[N] is end position. So in this case x is an (N+1) dimensional vector. max(x) == inf-norm(x). Using that you can solve for min,max analytically. Or do it the proper aka lazy way and just solve for max, then do some handwaving and claim symmetry relation between min & max.

Something similar can be done for the (0.05*N, 0.95*N) values, but that might get too involved for a quick experiment. If interested, lookup "order statistics".

Related linkies:
https://en.wikipedia.org/wiki/Norm_(mathematics)#Maximum_norm_(special_case_of:_infinity_norm,_uniform_norm,_or_supremum_norm)
https://en.wikipedia.org/wiki/Order_statistic
 

[e61_phil]

@mrflibble: Thanks for the hints. I already played with the sorted "path" to estimate only 99.73% (6 sigma) for peak to peak. But the difference is so small that I didn't cover it in the first post. 95% would only be equal to 2 sigma -> factor 4 instead of 6.

You're absolutely right with the shown examples. I promise to make better titles in the future

[dietert1]

When i was thinking about my data and about this difference between peak-to-peak expected from single point standard deviation and observed peak-to-peak values, maybe this is caused by looking at the difference of two reference voltages.

I guess the rare peak values of one reference will "collide" with the rare peak values of opposite polarity of the other reference even more rarely. So one would see 1.4 * the RMS of one reference but only 1 * the peak-to-peak values of one reference. Such a model could explain my observation until now. Maybe the expected higher peaks will appear after taking a very long data series with hundreds of points.
In other low noise tests where one device (meter) is much better than the other (DUT), the proposed factor 6 should hold better.

Regards, Dieter

[e61_phil]

Hi Dieter,

I'm not sure if you had that in mind, but I think your example is something like two normal distributions (one for every reference) which have similiar properties. Therefore, I made two distributions with the same µ and sigma (0 and 1). It seems to me, that the ratio between peak to peak and the standard deviation hasn't changed. (And I think that was expected because of the central limit theorem)

Best regards
Philipp

[dietert1]

Yes, that's what is expected, except i was wondering how long it will take to observe those peaks. And somehow i think it should depend on bandwidth. I could try what i get when looking at 24 * 38 hourly averages instead of 38 daily averages.
Maybe i should just be patient and wait for the first 2000 hours as everybody else.

Regards, Dieter

[Grandchuck]

There was a very interesting post by TiN on April 20 regarding his design and development of a reference board for one of his 3458As:

"I've swapped reference in experimental 3458A to explore the low-noise theories. It went with few trap pitfalls, but meter is running now and logging first data."

How is that experiment proceeding?  Any results that can be shared yet?

[notfaded1]

There was a very interesting post by TiN on April 20 regarding his design and development of a reference board for one of his 3458As:

"I've swapped reference in experimental 3458A to explore the low-noise theories. It went with few trap pitfalls, but meter is running now and logging first data."

How is that experiment proceeding?  Any results that can be shared yet?

Is this the board with 4 LTZ1000's on it?  If so that did look interesting to see what kind of data it produced.

Bill

[TiN]

So far noise data aligns with expectations on QVR ref.

Battery power:



SMU power:



But once installed in 3458A it brings havoc, most likely due to huge power consumption (190mA +15V, 20mA -15V) and heat dumped next to resistance current source circuits and DC preamp. Half of the ranges of DCV and resistance are ruined after transplantation. 
I did not expect it to be easy plug and play project anyway, more problem to see, better learning to get ..

Calibration report.

I want to try with ref sitting externally to meter and with different power/compensation schemes, but had no time to do that yet. 

To remind - single goal of that design is to provide lowest noise 7V reference possible, NOT the ultra-stable or battery powered DC standard. In fact I'd expect annual stability be much worse than usual 2ppm/year that even standard 03458-66509 can do after oven mod. 

[Kleinstein]

Why does it take so much power ? I would expect way less than 50 mA for a LTZ1000 reference circuit.

Not directly related to high precision, but what would be the minimum current needed ? I would expect some 3 mA for the zener (a little less than normal) some 0.5 mA for the OPs and transistor currents and than the heater, depending on the environmental temperature. For power conserving one could have the heater control transistor(s) not at the far edge of the board, but close to the LTZ to also help heating. At low current the transistor has more power than the resistive heater.

[notfaded1]

Despite the power and heat it's just such a neat idea... perhaps you could extend the connections for the PCB outside the case and build a little case for it so it can stabilize in there.  It seems you really don't have to fool the 3458A at all since it boots with it inside.     I'd be curious to know if not having the regular reference inside the case at all would have any effect the opposite way ie. not hot enough???  One things for sure... it was one of the coolest looking reference pcb's I've seen in while.

[Grandchuck]

Why does it take so much power ? I would expect way less than 50 mA for a LTZ1000 reference circuit.

Four LTZ1000s.

[niner_007]

Why does it take so much power ? I would expect way less than 50 mA for a LTZ1000 reference circuit.

Not directly related to high precision, but what would be the minimum current needed ? I would expect some 3 mA for the zener (a little less than normal) some 0.5 mA for the OPs and transistor currents and than the heater, depending on the environmental temperature. For power conserving one could have the heater control transistor(s) not at the far edge of the board, but close to the LTZ to also help heating. At low current the transistor has more power than the resistive heater.

A lot more than that, I measured an LTZ1000 board, taken directly from a 3458A, using 30mA at 11V, maybe the heater temperature being higher could explained it, but it seems the heater circuit uses a lot of power

[Kleinstein]

The A9 board from the 3458 is known to run at a rather high temperature. Still 30 mA is quite a bit, as the heater power goes up with the square of the current. TiNs is at nearly 50 mA per LTZ1000 so higher than that. Is that because of more heat loss through the board ? With a highly conductive board I would definitely prefer the A version.

Another part could be the heater resistance. There is quite some tolerance range in the data-sheet. A lower resistance heater would naturally need more current. So not all chips would be the same. The variations in the heater resistance should also effect the value for the "400K " TC compensation resistor. So not such a surprise that the datasheet example 400 K does not fit all.

[Dr. Frank]

TiNs 4 boards use about 170mA in total for heating, i.e. 42mA each LTZ.
I don't remember, which type he used (A, non A), and which oven temperature.

I recently assembled my first two LTZ1000A references at about 52°C (12k/1k) inside a thermally isolated tuner box, LTZs again not individually isolated, and each consumes 18mA @ 12V in total only. All my non A references at the same oven temperature consume about 23mA in total.

So I really wonder, what the hybrid of TiN is doing ...
Probably a completely different type of reference, off from the LTZ1000, is needed to really achieve lower noise in the 3458A.

Frank

PS: At 36:22min in his video you can see, that he uses the A version, and 13k/1K oven setting, i.e. about 70°C.
42mA is far too high, but maybe the stabilizer and averaging circuits consume a noteworthy part of that.

[dietert1]

Does the reference board have a temperature sensor? In order to reduce power consumption one could wrap the board to protect it from the fan air flow. An easy test, and then you know whether it's the heaters. Later one can think about whether it's a problem for low thermal EMF to leave the whole board running hot.

Regards, Dieter

[Pipelie]

@TiN
time to wake up and reveal the answer!

As far as I know, to achieve ultra low noise, TiN isn't just put 4 LTZ in parallel he also increase the Iz of LTZ significantly.

[Dr. Frank]

@TiN
time to wake up and reveal the answer!

As far as I know, to achieve ultra low noise, TiN isn't just put 4 LTZ in parallel he also increase the Iz of LTZ significantly.

Probably not... if Illya has dimensioned the neg. currents correctly, to exactly compensate the reference currents,
then each I_z should be about 4mA.
PS: In the video, 120 Ohm are probably used for the zener current, that gives typically less than 4mA I_z

190mA minus 20mA then gives  about170mA for the 4 heater currents, plus the summing OpAmp. It's that easy.
The oven temperature can be measured by digitizing the startup voltage of the BE diode of the reference transistor.
This changes by -2mV/K, so its difference between cold and hot state gives quite precisely the final oven temperature.
A pity, that nobody else determines the real oven temperatures.
Frank

[Andreas]

So I really wonder, what the hybrid of TiN is doing ...

I fear the ceramics together with short legs is doing a good cooling job ...

with best regards

Andreas

[KT88]

I fear the ceramics together with short legs is doing a good cooling job ...

...but may cause some mechanical strain on the other hand side...

[niner_007]

So I really wonder, what the hybrid of TiN is doing ...

I fear the ceramics together with short legs is doing a good cooling job ...

with best regards

Andreas

I do wonder if a ceramic PCB is a sensible approach for an LTZ1000 reference, when the reference is doing the temperature stabilization, typically Fluke et all use a ceramic PCB, with a non heated reference, or with the heater heating the entire ceramic PCB, no? Maybe not much better can be done from what was already achieved here, prior, with standard PCB construction. Would be interesting to see thermal pictures of the ceramic PCB, I think someone has posted them before.

[branadic]

It's Rogers, not a ceramic PCB, so thermal conductivity is expected not to be that great compared to real ceramics.

-branadic-

[Dr. Frank]

I would like to summarize about the newly built two LTZ1000As, which are the first ones to me, after 7 non-A types. At first, these references will be powered from a common linear 12V supply, that are about 11.5V for the circuit.

To mitigate random 'jumpers' and for pre-ageing of the die cast, both LTZ chips got a high temperature Burn-In. I can't disclose any details.
I only assembled the reference and a buffer amplifier, w/o 10V amplification. The output got a  ferrite core to suppress external noise.
The mechanical assembly is similar to the previous 5 LTZ1000 CH. For thermal insulation, I salvaged 3mm thin Polystyrol from Doner Kebab boxes, which is readily available, and can be cut much more nicely with a scalpel, and also saves some space.
Its purpose is to mitigate fast thermal changes and create an isothermal chamber inside the tuner box to mitigate thermal e.m.f.  for the whole assembly, but not only for the LTZ chip, like the plastic caps do in the hp3458A reference. The legs of the A chips were not cut short at all.

Temperature difference between ambient and internal is about 2°C, due to the 400K/W thermal resistivity, whereas the non A chips create 7°C difference.
All my references use a 12k/1k divider. By choosing 100µs sampling interval on my 34465A, I precisely measured 53.0 and 51.6°C oven temperatures. The chip is heated up within 3 seconds, and the supply current is 18mA only.
Initially, I also experimented with 12.5k/1k to account for the 10°C self heating of the A type. This gives 10°C higher oven temperatures, i.e. between 60..65°C. A 13k/1k divider probably gives over 70°C, in contrast to the data sheet.
Therefore, the 12k/1k gives best drift performance of < 1ppm/yr., similar to any other standard references like 732X and Datron 7000.

I measured the 12h stability of both references, also to determine some noise parameters (1h and 10min.). The first measurement started at power on and indicates that the whole assembly needs about 15min for stabilization to < 0.1ppm of nominal.

The stability is mostly at +/- 0.1ppm_pp, at certain 1h time frames even +/-  0.05ppm_pp or 0.028 ppm_rms, which seems to be less than the other LTZs.
The references probably need some further settling time to deliver better short term stability.

As you can see, I used mixed types of PWW (econistor from G.R.) and BMF (FLCY from AE) resistors, but did not determine any T.C.s upfront.
The overall T.C.s were 0.02 and 0.03 ppm/K respectively, w/o any additional compensation resistor R9.
The measurement for LTZ#7 with 51,6°C oven temperature is quite interesting, as at about 40°C internal temperature the oven regulation fails, i.e. the low T.C. of the regulated oven develops abruptly into the +50ppm/K of the unregulated reference. This confirms that the self heating is about 11°C.

'A' types at nominal 50°C oven temperature can be used up to about 38°C room temperature 'only', but this is no disadvantage in a standard lab environment between 18..28°C.

Frank.

[babysitter]

Hm, do I glimpse a bit of inspiration from my box?
How did you wire the ferrite toroid - common mode choke? source material?

BR
Hendi

[Andreas]

All my references use a 12k/1k divider.

Hello Frank,

You simply had luck that the voltage of the LTZ (7.1V) is more at the lower end of the tolerance.
With 7.3V zener voltage the oven would fall out of regulation already at 30 deg C.

with best regards

Andreas