Ultra Precision Reference LTZ1000

[Marco]

You can buffer the reference with a separate opamp and switch the output into the capacitor with a small resistance ... once it's charged, flip the switch again.

[Mickle T.]

I have a two of noise measurement setups now. The first device have a cheap russian 800 uF 50V wet slug tantalum capacitors battery (10$), two PWW resistors and EM Nanovoltmeter. The second scheme is an adaptation of AN124. Both devices gives a similar results on the all of my voltage references, but #1 have a <10 minutes settling time and 20 nV p-p input short noise, #2 only 20-30 minutes and 160 nV respectively.

[janaf]

Mickle T 

The $1.100,00 cap, OK, I'm interested to map the noise but not THAT interested. But reading the app note makes me feel a bit better. It's not supposed to be simple to measure noise at this level.

Once beyond the initial amp stage, a gain of 10 or 100, the rest should be relatively simple. As I wrote before, a 8-bit ADC should really be enough. I've mostly been measuring rms, estimating peak to peak. To me the rms value is more relevant anyway. Eventually, I want to see what tinkering with various options (like the huge R2) will do to noise. But even with an analog or digital scope, once the signal is low-pass filtered to below 10Hz, the difference should IMO be academically small.

I was also thinking of another solution; set up a second LTZ1000, trim it to have output level the same (within millivolts) as the DUT and measure the direct difference. The difference signal can then be amplified a lot, like 60dB, without A/C coupling involved. But frankly it sucks to have to build two setups......

I have now got a 18.000uF 80V E-cap charged to 80V. The first minute, the leakage was 1.5mA. After half an hour down to 0.1mA. I'll leave it for a day or two and see. When lowering to 7V it may be usable :-)

A related side track. TI have a new 32-bit D/S ADC as preview since February: ADS1262 / ADS1263. But very little data is available so far. One nice feature is that the sample rate is specified from 2.5Hz up to 38kHz. The accuracy will of course not be anywhere near 32 bit but resolution / noise may be useful....

[janaf]

Another thing you might do is take a look at the HP/Agilent/Keysight 34420A's input low-noise amp.

Will definitely take a look. I think Keithley had a really low noise amp too.

[janaf]

The *reason* that Jim used the low-noise dual-JFETs in his design is that it is near impossible to get both low-noise and high input impedance at the same time in a tiny op-amp die.  This is a matter of physics-- for a JFET, you need a large die area to get the noise down, and this is not going to happen in an op-amp IC.

I'm beginning to see that. As you say, you can get the low noise or the impedance but not both in an op-amp.
It seems you can get pretty far though, maybe "good enough" to measure the 200nVrms 0.1-10Hz but margins are very small. As mentioned before, I might also be happy measuring something like 1-10Hz, as long as the measurements are comparable & stable.

I downloaded the 34420A manual from Keysight but the scanned schematics where of practically unreadable quality. Any other source?

[Kleinstein]

For the problem of building an amplifier to measure LF noise of the ref. there is a solution in a German language forum:
http://www.mikrocontroller.net/topic/207061#2060389

The Input stage is just a LT1037 OP-Amp with a selected low leakage 3000 µF Elko. Not even a special tantalum type, but likely good AL types. selected for low leakage. The drawback is that it takes something like overnight (possibly a few day for the caps - so keep them charged all the time) settling.

Preparing and selecting a few caps may be worth a test - no special gear needed, just a battery (e.g. 9V), a meter with nA resolution and a few days of time. Something like 2200 µF 63 V cap's are not that expensive and still useful if they don't turn out to be good enough.

One conclusion was that the rather complicated circuit from AN124, combining an OP amp with JFETs was not such a good idea. So there is no need to think so complicated.

[janaf]

Thanks, i have some few e-caps on full voltage now, I will put them on 7V (or 9V) later.

[fmaimon]

Preparing and selecting a few caps may be worth a test - no special gear needed, just a battery (e.g. 9V), a meter with nA resolution and a few days of time. Something like 2200 µF 63 V cap's are not that expensive and still useful if they don't turn out to be good enough.

No need for meters with nA resolution. Just do as Jim Williams did and put a 1k resistor in series and read it's voltage.

[paulie]

Thanks, i have some few e-caps on full voltage now, I will put them on 7V (or 9V) later.

The two chinese 1000uf caps seen in my photo used for the filter required 2 days to form and then about an hour to stabilize each time after that. If they go a month unused then must start over. Of course as mentioned previously my requirements were an order of magnitude less stringent.

[Edwin G. Pettis]

@Klienstein,

Oh really, you think you know more about the subject than Jim Williams, perhaps you should get in touch with Linear Tech and let them know there is a replacement available for Jim.  I do not intend any insult here but it is obvious that some here do not grasp the underlying design principals involved  in trying to measure very low noise levels accurately.....note that word - accurately, not in the ball park measurements.

Seriously, Jim did it that way because that was the correct way to do it at the time, moving on a few years, yes, you might be able to sub a few of the components, such as a polypropylene capacitor for that tantalum and maybe some other minor changes but the circuit as done cannot be significantly improved on nor can you duplicate it with cheaper, lower quality parts.  This is not the jelly bean project some seem to think it is.

Again, you cannot measure a Gaussian noise source with an integrating or sampling ADC (that includes DVMs), it is not accurate, it does not repeat, it does not measure peak-to-peak nor RMS but will give you a sort of kind of average variable of the noise.  Unless you are using a very long period of measurement, tens of seconds at least, the average of a DC-10Hz noise band will not be accurate in any way, shape or form and in the case of a voltage reference, you need to know the noise characteristics accurately and that includes peak to peak as well as RMS, average is of little use.  Noise is the limiting factor in a reference, not just drift.

Pardon me if I sound a bit terse.

[Marco]

The long settling time is not so the capacitor can charge up to voltage [although that's part of it], but rather so that the dielectric absorption [DA] can dwindle down to zero.  DA, even on a huge high-voltage capacitor will look like a large leakage current until the process is complete.

That doesn't really matter if it has no significant component above 0.1 Hz ... a subsequent high pass filter gets rid of it.

Same as with the leakage current (as long as the output doesn't hit the rails).

PS. simply putting a zener across Rf would be the easiest way to quickly charge it.

PPS. flicker noise is surprisingly high for metal film resistors, but at ~0V across it it's unlikely to be an issue here.

[Andreas]

Hello,

rant: the tread is going more and more off topic.
The construction of a low noise amplifier would be worth a own thread.
(as can be seen on the length of the german thread).

Ontopic:
You never should draw a current more than about 1-2 mA from a unbuffered LTZ1000(A).
Otherwise the temperature setpoint will go to infinity introducing at least a hysteresis to the LTZ1000 output.
So your calibration will get lost and perhaps the LTZ is quickly aged.
Usually you can see the influence of such a "event" as drift up to a half year.
So charging a large capacitor from a unbuffered LTZ is a strong NO NO.

Even charging a 1uF capacitor from unbuffered LTZ will introduce a visible shift of the output voltage.
So I would not use a zener to clamp the 1k resistor further.
Instead I would introduce a further resistor during charging of the input capacitor to avoid the LTZ being toasted.
(And I do not complain about the additional time when measuring noise).

With best regards

Andreas

[Marco]

I said a zener across Rf , I was assuming the capacitor being in the feedback loop. When the capacitor isn't charged the opamp will go to the positive rail, with a (low leakage) zener diode across the feedback resistor it can charge the capacitor until it hits 7V.

[janaf]

rant: the tread is going more and more off topic.

Yes. New thread started: https://www.eevblog.com/forum/projects/low-frequency-very-low-level-dc-biased-noise-measurements/msg654777/#msg654777

Mod /Dave: imo, move basically all posts in this thread from #1174 and forward.

[lars]

For the first time I have built a reference with an LTZ1000. Got two unused LTZ1000ACH-PBF with date code 0733 from a friend. I followed the LT datasheet for the design except added a 22ohm cathode resistor ("R60) and R1 was 13kohm. I built the circuit on a prototype board (RE201EP from Reichelt). For R1-R4+R60 I had a parallell resistor with 100 times higher resistance that I could select with a jumper. For R60 I also could short the resistor. I even used a socket for the LTZ so I could switch between the two LTZ's. As I only have a 6-digit DMM (Keithley 2000) I got a resolution of about 1.5ppm. I was glad to see  that even if I switched jumpers, LTZ's or was blowing on the LTZ I could only see a shift of one digit (1.5ppm) going back to the same configuration.

I did the test with a switching lab supply (Korad KA3005D). As the circuit worked fine down to 9.2V I choose 12V. Current draw was about 20mA. From 10 to 15v I could see no change (less than 1.5ppm change).

This is my result of the sensitivity test (ppm voltage for 100ppm resistance change as in datasheet):

With R60 shorted (original circuit with R1=13k)
LTZ#1   
7.16822V
R1 -0.17ppm
R2 -0.42ppm
R3 -0.10ppm
R4  1.14ppm

LTZ#2
7.13327V
R1 -0.18ppm
R2 -0.42ppm
R3 -0.08ppm
R4  1.00ppm

With R60=22ohm
LTZ#1
7.24931V
R1 -1.08ppm
R2 -0.47ppm
R3 -0.01ppm
R4  0.00ppm
R60 1.12ppm

LTZ#2
7.21639V
R1 -1.15ppm
R2 -0.47ppm
R3 -0.00ppm
R4  -0.14ppm
R60 1.15

/Lars

[Andreas]

. As I only have a 6-digit DMM (Keithley 2000) I got a resolution of about 1.5ppm.

I did the test with a switching lab supply (Korad KA3005D).

Hello Lars,

Thanks for the results.

Have I understood it right: you are making a 1% resistor change
and calculate the result linearly down to 100ppm change?

If you had 2 sockets you could measure the difference of the 2 LTZs having a resolution of 0.015 ppm in the 200mV range.

Mhm: my Korad KA3005D has a linear transformer + power stage
(only a digital display and keys to set the voltage setpoint by a D/A converter).

With best regards

Andreas

[lars]

Have I understood it right: you are making a 1% resistor change
and calculate the result linearly down to 100ppm change?

Yes, I have used Excel with the resistor values I have used and scaled it to 100ppm resistor change.

If you had 2 sockets you could measure the difference of the 2 LTZs having a resolution of 0.015 ppm in the 200mV range.

I will make another board later but want better resistors before I do that.

Mhm: my Korad KA3005D has a linear transformer + power stage
(only a digital display and keys to set the voltage setpoint by a D/A converter).

So I guess my also is linear (got it from Reichelt last week).

/Lars

[Marco]

Question about all the talk about dual PWM DACs a few dozen pages back ... basically there is no known monotonic by design way to combine two PWMs for higher resolutions? (Datron/Fluke calibrators seem to rely on calibration to correct for gain/offset errors in the resistor combiners of the MSB/LSB PWMs.)

[Andreas]

Question about all the talk about dual PWM DACs a few dozen pages back ... basically there is no known monotonic by design way to combine two PWMs for higher resolutions? (Datron/Fluke calibrators seem to rely on calibration to correct for gain/offset errors in the resistor combiners of the MSB/LSB PWMs.)

Hello,

you will also have to calibrate out the INL of the MSB PWM (at least).

On PWM INL results from different resistances in HIGH/LOW state (temperature dependant).
Further you may have different (HIGH/LOW) switching times (also temperature dependant).
Further charge injection errors.

So the adding of the 2 resistors may be your least problem.

If you didn't already see this article, you can find some valuable informations in there :

http://www.edn.com/design/other/4326640/DC-accurate-32-bit-DAC-achieves-32-bit-resolution

But be carefully: the values given are very optimistic.

And: the formula for adding the 2 PWMs to the output signal is simply wrong.
You will never get a output voltage larger than VREF * 65535/65536 by
combining 2 PWMs with a resistor divider instead of a summing amplifier.
So the half reference voltage will be reached at 0x80008000 instead of 0x80000000

With best regards

Andreas

[Marco]

you will also have to calibrate out the INL of the MSB PWM (at least).

You could low pass and buffer the MSB/LSB parts separately, then the switch resistance at least becomes irrelevant (offset voltage differences become relevant instead).

Any way, so no monotonic by design implementations? (ie. not like plain PWM, thermometer DACs and buffered Kelvin Varley with "small enough" offset voltage errors.)

[Edwin G. Pettis]

Regarding the 32-bit DAC story: http://www.edn.com/design/other/4326640/DC-accurate-32-bit-DAC-achieves-32-bit-resolution

This was later discredited as not working and not providing anything better than the current 24-bit DACs on the market so don't waste your time with this, it does not work, not even theoretically.  The noise floor is much too high in the first place to get beyond maybe 24 bits at best just for resolution, forget accuracy, with good luck it might just do 1 PPM accuracy with a lot of tweaking.

[Kleinstein]

Combining 2 PWM Signals is difficult. With some modern µCs very high timing resolution (e.g less than ns) is possible, and thus very high resolution PWM, like 24 Bit. 

[Andreas]

With some modern µCs very high timing resolution (e.g less than ns) is possible, and thus very high resolution PWM, like 24 Bit.

And how do you switch your precision analog voltage reference (e.g. 7V) with this PWM.
Usual analog switches have about 100 ns switching time and a part of this as break before make time.

With best regards

Andreas

[Marco]

It would probably be better to have two SPST switches and do your own break before make timing.

[acbern]

A PWM DAC (e.g. two stage), for use as a precision reference divider, working from 2 LZTs in series to be able to generate a voltage close to 10V without an amplifier (which you want to avoid) does not have to have good DNL/INL. All you need to be able is to set the voltage close enough to the 10V and keep that stable, i.e., the switching period... must be stable as well as the analog summing amp (it can be calculated that the summing amp drift is not the problem in a normal environment). The Datron 4910 solved this long ago, the schematics are available on the net. It is not a simple design though, and switching noise, filtering are just a few of the challanges.