T.C. + Hysteresis measurements on brand new LT1027DCLS8-5 voltage reference

[Andreas]

Hello,

What is the easiest way now to implement a home grown 7 or 8 digit converter these days?

I can only tell how I do it:
Select the best of a batch of 5V references. (T.C. < 1ppm/K and hysteresis < 1 ppm in near room temperature).
Combine with a LTC2400 and a LTC1043 divider.
Put a uC to the whole for transfer data to a PC and hold calibration constants and measure temperature of the reference.

Adjust NTC for temperature measurement (is optional since the true temperature is not essential).
Adjust T.C. of VREF with 3rd order calibration to better 1 ppm over a 10-40 deg C temperature range.
Adjust linearity to better 1 ppm with simple square approximation of the error curve.
Optional adjust Offset over temperature.
Adjust full scale (nominal VRef value at 25 deg).
Age the whole for 5000-10000 hrs.
Eventually repeat the calibrations.

What do you get after 2-3 weekends adjustment and the run in phase:
A unipolar +0..10V input range (up to now not very high impedant).
around 1-2uV resolution with a integration time of 1 minute.
(so more a old style Solartron than a modern instrument)

Better stability than a typical 6.5 digit instrument. 
Allan deviation shows around a factor 5 worse than a 3458A.
(ok the 3458A is much faster with 100NPLC compared to 1 minute integration time).

Standard deviation against a LTZ1000A over 42 days is around 0.25ppm for my best devices.
Drift on ADC13 is nearly the same as on my LTZ#1.
(around 1 ppm/year + additional 1 ppm seasonal changes between ADC+LTZ1000A due to humidity).

With best regards

Andreas

[Andreas]

Hello,

its time to go on topic again and show some overview over the measurements:

From 10 references I found 3 (eventually 4) usable ones for my purposes. (marked green).
(T.C. < 1ppm/K at least in the 18-32 deg C range and hysteresis < 1 ppm).

The "hysteresis" on these devices is mostly a large initial drift of 2 ppm/day,
where I hope that it gets reduced after some more temperature cycles.
I tryed several preconditioning either with running some days at room temperature
or (with asterisk) with a 15 mA load which is switched on 1.5 hrs within a cycle of 2 hrs.

I will have to compare the results with my other references especially with the AD586LQ and the LT1027CCN8-5.
In every case the LT1027DCLS8-5 (average T.C. around 1.4 ppm/K and Hysteresis around/below 2ppm)
 is much better than the LT1236AILS8-5  with around 3 ppm/K and larger Hysteresis.

But I fear that the LS8 package is not so good compared to a metal can package.
Especially when mounted directly on a PCB.
So beware of a LM399 or LTZ1000 in a LS8 package.
Unfortunately I have no LT1027 in metal can package to compare with.

With best regards

Andreas

[zlymex]

Speaking of preconditioning, there was one article by V. S. Orlov(from Datron? Attached) describing a method called 'soft' thermal shock.

Some one tried this with good result(http://bbs.38hot.net/forum.php?mod=viewthread&tid=25193)

[Andreas]

Hello Andreas,

Very nice setup and measurement. It seems to me that temperature curves of LT1027 have quadratic shape.
Have you measured the noise of that 3.6V(50% of LTZ1000 by LTC1043)?

Hello,

now I did 

First results:
Setup: LTZ1047B#4 with LTC1043 divider followed by a LTC2057 buffer amplifier.
The LTC2057 buffer has a 825R series resistor + 820pF at the output. (Improves noise for the LTC2400)
When used with the 0.1 .. 10 Hz filter amplifier with a input impedance of 1000R this gives a voltage divider with factor 1.825 which has to be considered for the readings.

First picture 1/f (0.1 .. 10 Hz) noise of LTZ#4 (around 1200nVpp) with 2:1 divider
reading 545 nVpp * 1.825 = 994 nVpp at output of unloaded divider.
So around factor 1.6 against a noiseless divider. (would be 600nVpp)


2nd picture: 1/f noise of input from LTC1043 divider shorted to GND.
reading 213 nVpp * 1.825 = 389 nVpp  at output of unloaded divider.
(noise floor of the filter amplifier alone is < 200nVpp).

3rd picture: wideband noise 10Hz ... 100kHz as FFT.
Around 350 Hz the switching frequency of the LTC1043 with -117 dbV (1.4 uV)
I long searched for the "switchmode power supply" at 80 kHz until I recognized that this is in reality the (100kHz typ.) chopper frequency of the LTC2057 buffer amplifier. Amplitude -67dBV (450uV).

with best regards

Andreas

[zlymex]

Thanks very much Andreas for the test.
It seems to me that LTC1043 added not much noise when divided by 2, a good sign.

The question is, how to calculate the noise added.
1. by multiplication factor? 994nV/600nV=1.66. However, it cannot be apply for shorted input where 389nVpp/0nVpp has no meaning. Or it should be 389/200=1.945?
2. by Root-Sum-Square? sqrt(994^2-600^2)=792nVpp, much higher than sqrt(389^2-0^2)=389nVpp
3. by simple addition/substraction? 994-600=394nVpp, 389-0=389nVpp, although agreed, but I suspect that this is the way to calculate.

[Andreas]

Hello,

if the noise sources are independent they should add as root sum squares.
But I think there is some charge injection from the LTC1043 switching capacitor
towards the buffer op amp of the LTZ1000 reference. So maybe this is non linear.

By the way: I am matching the capacitors of the LTC1043 to be within 1% (typical 0.1-0.2%) to minimize settling time.

With best regards

Andreas

[zlymex]

Hello Andreas,
Were you suggesting that the noise generated by LTC1043 depends on the input voltage?
That is to say, the noise is smaller when input shorted?

[Andreas]

Hello,

no I think more that the noise of the switches of the LTC1043 input
influences the buffer amplifier o(LTC2057) f the LTZ1000 output.
(will do further measurements on week end).

With best regards

Andreas

[Andreas]

no I think more that the noise of the switches of the LTC1043 input
influences the buffer amplifier o(LTC2057) f the LTZ1000 output.

Hello,

obviously that is not true.
If I measure the noise at the output of the LTZ there is no significant difference
 between idle LTZ and LTC1043 connected to the buffered output. (first row of measurement table).

I also removed the 825R series resistor on the LTC1043 buffer to see wether the voltage divider against my 1K input impedance
has a influence. (column 1 vs column 2 of measurement table).
Also no significant change. (Except the fact that ~230nVpp are very close to the noise floor of ~120nVpp of my filter amplifier)

So all in all the buffered LTC1043 adds around 370-400nVpp on the output signal.

With best regards

Andreas

[Kleinstein]

For comparing noise measurements, it's often better to use the RMS values instead of the peak to peak values. They are scattering less. For peak to peak measurement more data would be needed to reduce scattering. Also some kind of median / robust mean (drop the larges and smallest quarter and average the rest) would be better than the plain average.

For the shown data, there seem to be quite some interaction of the LT1043 stage and the LTZ reference:

With a shortet input, the LT1043 stage gives a noise of about 50 nV(RMS), including the output buffer amplifier. With the LTZ as source, the noise increases from an expected 76 nV to about 131 nV. Thus the additional noise would be at about 110 nV (due to adding as squares !). So most of the noise would be from the interaction.

 One reason would be that the noise of the LT1043 stage depends on the input voltage. It might be interesting to use a different input source to the LT1043 stage instead of the LTZ. For the noise measurement some batteries / NiCd accumulators might be worth a measurement.

[Andreas]

Hello,

For comparing noise measurements, it's often better to use the RMS values instead of the peak to peak values.

thats why I give both values. The Vpp for comparison with the data sheet. The RMS value for easier comparison.

One reason would be that the noise of the LT1043 stage depends on the input voltage.

Good Idea.
I just checked this with 2, 4 and 6 NiMH cells (2.7V 5.3V and 8.0V).
The result shows no significant dependancy from the input voltage.
I needed a while to get reproducable results with this measurement.

The reason are the different battery holders with springs (most probably steel)
and thus thermo voltages up to 3uVpp measured at the output of the voltage divider.
And this even within my metal cookies box!

Surrounding the battery holder with a microfibre cloth and wiggling the contacts
before measurement gave reproducable results in the 400nVpp range.
So noise measurements also depend much on the mechanics.

With best regards

Andreas

[Kleinstein]

So the LT1043 by itself seems to be not adding extra noise. But there still is the higher noise, when dividing the LTZ signal.

The way the Lt1043 divider circuit is build, it is sampling the input only for sort intervals, just before switching. This will make it sensitive to noise at higher frequencies. So noise in the higher frequency range (e.g. around 1 kHz) might get mixed down to the LF band and thus this way increase the noise. So it might be a good idea to have an low noise filter at the input of the LT1043 divider. Also a series resistor to limit the speed how fast the capacitor is recharged could be an important point to limit the effective bandwidth.

[Andreas]

Also a series resistor to limit the speed how fast the capacitor is recharged could be an important point to limit the effective bandwidth.

The LTC1043 itself has a large series resistor within the switches (300-700 Ohms).
So with the 1uF capacitors there is already a low pass in the range of 300 Hz.
Of course there is some charge injection outside the series resistors.

With best regards

Andreas

[Kleinstein]

The on resistance and the 1 µC capacitance helps to a certain degree, but there is still some higher frequency noise from the LTZ to be mixed down to the LF range. With something like a 300 Hz band limit my estimate is not to expect that much noise, but still a small contribution. It might be still worth to add some filtering (e.g. 1 K and 10 µF) between the LTZ and divider, as the low noise NIMH cells did show the much lower noise level. Good filtering without adding resistor noise or drift / offsets due to capacitor leakage might be tricky and it can also interact with the buffer amplifier.

I don't think charge injection should be the problem - this is more like a source of offset and possibly drift.
The charge injection current might limit the amount of RC filtering that is useful.

The other possible source could be RF noise going back from the LT1043 to the LTZ reference / buffer amplifier and causing trouble there.

[Andreas]

Hello,

did a simple test to check the charge injection: I shorted the 10nF capacitor of the R/C-Oscillator of the LTC1043 to GND.
(Stopping the oscillator and thus all switching of the LTC1043).
Result: dramatically increasing (factor 2) of the noise.
So the only explanation is the current noise of the LTC2057 buffer amplifier.
(The input node of the OP gets high ohmic by stopping the oscillator).

After removing all shorts I got again the normal values.
(ROut = 825 R so correction factor 1.825 is again needed).
I also checked a 2nd sample with LTC2057 buffer amplifier  (V10_1501#2).
This sample had even with working oscillator nearly double the noise of sample #1 (V10_1501#1).
The standard cirquit with LTC1050 (sample V10_1102#2) shows 1.75uVpp together with the LTC1043.
(Datasheet value with low impedance input is 1.6uVpp for the LTC1050 alone).

With best regards

Andreas

[Andreas]

As in all choppers, if you add a capacitor from the (-) input to the output, the chopper clock "feed through" will be dramatically reduced. 

Hello Ken,

I have a unitiy gain cirquit. So the capacitor in my case is a zero ohms resistor.
I do not think that I can improve this.
Do you have other experiences?

With best regards

Andreas

[Kleinstein]

I won't expect current noise to be a major issue, as the switched capacitor stage is still reasonably low impedance (e.g. < 1 K range). This gets different with the clock stopped or a much lower clock.

The difference between the two circuits could be due to different clock speed or on resistance of the switches.


To reduce the effect of mixing down noise from the input to the LF range one might consider using both half's of the LTC1043 chip, used with opposite phase. This would sample the input nearly all the time and thus might reduce feed through of noise from the input.

[Andreas]

Hello,

I have to correct my measurement from 02.08.2016:
(The difference between LTC2057#1 and #2 sample did not let me sleep).

I recognized that on sample #2 of V10_1501 the oscillator did not work.
Reason was a defective 10nF capacitor for the oscillator of the LTC1043.
So the noise corresponds to the measurement of #1 with shorted 10nF capacitor.

After repair the noise is around 430nVpp so very similar to sample #1.

You are probably buffering a voltage that is across a capacitor-- and this capacitor will attenuate the [~100Khz] clock noise quite a bit anyway...

Attached the cirquit of the buffer stage.
the +Input is from a 1uF foil capacitor.
When the oscillator of the LTC1043 is stopped by shorting the 10nF capacitor (not shown)
then the flying capacitor is in parallel with the 1uF output capacitor.
(The switch is shown in the "high" state of the oscillator pin.

Edit: in the actual build with LTC2057 C28 is populated as 0 Ohms Resistor and R26 is not populated.

To reduce the effect of mixing down noise from the input to the LF range one might consider using both half's of the LTC1043 chip, used with opposite phase.

Unfortunately there is only one oscillator pin for both switches.
And I do not really know what happens if you exchange input with output
because of the charge injection compensation cirquitry.
At least the "gain" of the cirquit changes if you do that.

with best regards

Andreas

[Kleinstein]

The charge injection circuit could add a little offset / gain difference. Combining the two parts would than result in a small modulation at the output. But this should not be so much with the relatively large capacitors.
The advantage would be a much lower (e.g. a factor of 10-100) sensitivity to noise from the band around the clock frequency.

With the rather low clock frequency (10 nF should result in about 300-400 Hz). The clock frequency is about in the cross over range of the filtering through on-resistance and the caps. So it might be worth testing the noise with a faster clock (e.g. 2 nF cap). A higher frequency would also make extra external filtering more practical, without adding to much resistance.
What is the reason to choose this low frequency ?

[Andreas]

The low frequency is
a) recommendation (10nF) in the data sheet. (typical 400-500 Hz)
b) after my experiences the precision parameters (linearity) get worse if you change any of the capacitor values.

With best regards

Andreas

[Andreas]

Hello,

to come back onto topic: the rest of my LT1027DCLS8 measurements.

From 25 references there are 4 good and 3 "usable" references regarding T.C.
Interestingly: the best 4 all show lower readings of LTZ#4 voltage. (so have VREF at the upper end of the distribution).
The other way round is not true so a higher output voltage has not necessarily a lower T.C.

In average the T.C. is around 2 ppm/K (against 3 ppm/K for the LT1236 in LS8-package).
The measured "hysteresis" is mostly a ageing drift which takes place at higher temperatures.
I hope that this drift will settle down after a few cycles.
I will build 4 ADCs (2 with PCBs from branadic and 2 with my stressless mounting method) to check ageing drift.

All in all I am not so lucky with the LS8-package. There seems to be a problem with the die attach which creates those relative large ageing drifts/hysteresis.
I will have to compare with my AD586LQ and LT1027CCN8-5 measurements. But from remembering the LT1027 in plastic package had far low "hysteresis".
And the AD586LQ had lower T.C. in average.

with best regards

Andreas

[Kleinstein]

Die attachment with some kind of epoxy glue could cause some drift / aging. Typical epoxy has is amorphous with a glass transition temperature somewhere around 100-150 C. So heating above that temperature (e.g during soldering) can cause a new aging cycle and the speed of cooling can have an influence. Ideally one would expect a more stable behavior when cooling down to about 70 C is slow. After fast cooling there could be quite some structural relaxation in the glue.

For the LTZ1043 it could be just charge injection that causes more nonlinearity / errors, when the frequency is higher. Adding resistance to get more filtering would also increase the errors due to change injection. So the way around would be more with larger caps than the 1 µF - but this makes things bigger. Still 2.2 µF or 4.7 µF may be still acceptable as film caps.

[Andreas]

Hello,

just built the first 2 ADCs with the new references.
2 different designs.
One older from me which I used already with LT1236 (the upper on the overview picture).
And one PCB with design from Branadic.

The PCBs with differences in detail:
Branadics PCB contains already a pre-regulator 10.5V and the USB-Interface on board.
The mounting of the LT1027DCLS8-5 (#7 of the measured ones) is done with
the recommended cut outs according to data sheet / application note.

My design uses external pre-regulator (9.31V on the picture but will replace with 10.5V).
USB-Interface is externally done.
Mounting of the LT1027DCLS8-5 (#2 of the measured ones) is done with my
stressless in hole mounting method.

Software is flashed on both devices.
I will have to calibrate the NTC temperature measurement and then the T.C.
of the reference with final thermal shielding.
And finally the linearity of the ADC needs to be compensated.

With best regards

Andreas

[David Hess]

That galvanically isolated USB interface is a work of art.  I just about jumped up to cheer when I saw it. \o/

[Andreas]

Hello,

after having made one stability measurement with the first device (ADC21) without T.C. calibration

https://www.eevblog.com/forum/metrology/ultra-precision-reference-ltz1000/msg1004639/#msg1004639

I also made a measurement after T.C. calibration of my own design.
I hoped that the stability would be below 0.25uV (after 2:1 divider).
But obviously there is still a large ageing drift (independant from temperature).
From the Allan deviation the standard deviation goes up to 0.5uV (2:1) or 1.0uV
(before divider) for measurement times of 100 minutes.

I hope that the drift of the LT1027DCLS8-5 will settle down
 to the values that I am used from my AD586 devices.

With best regards

Andreas