LM399 based 10 V reference

[Kleinstein]

In this circuit PWM is used to set an amplifier gain. The very extreme PWM values can be problematic. So starting slightly below 7 V the lower limit would be a little over 7 V (maybe 7.1 V). The upper limit is likely set by the supply, so maybe 11.5 V.
The other limitation of the circuit is a relatively slow settling - not a problem with a fixed reference though.
The setting is also not necessarily very linear.

The circuit looks a little odd at two small details: one is the choice of OPs. My guess is it would work better with the OPs swapped. The AD8675 is a really good OP, but still not that suitable for a 300 K source. The AD8675 might than need a little more than 12 V to work at 10 V.

The filter at the output amplifier is only rather high frequency. I would expect at least a little more than 10 nF there.

[branadic]

It's just a board for experiementing and for getting a gut feeling how to improve things, not a final schematic with all components defined forever. The board also includes some options to play with.
After waiting forever the boards arrived today.

-branadic-

[try]

Hi enut11,

Expanded graph sections from the results above. All measurements are spot checks.
The blue line is Vz. Graph shows variation from test-to-test.
The red line is 10v output. Graph shows variation from 10v. I occasionally tweaked the 10v output to bring it back to 10.00000v.
Some of the graphs show off-chart excursions. These are simply missing data from when the Refs were traveling around Australia for the Aussie Cal Club.

I am zeroing in on Ref #5 as the best of this lot.
enut11

the variations of the 10V-output of your reference look unnecessary high. I have just one LM399-based 10V-reference which is permanently running and most of the time it is connected to a 34401A. I can't provide you with a fine documentation such as you prepared but I can tell you that my 10V output drifted 2ppm in the first year and 4,5ppm in the second year in absolute* terms.
The output drifts really slowly over time. During summertime with wild temperature swings in my appartment from 15 degrees Celsius to 30 degrees the 34401A shows variations of maybe 1,5 to 2ppm per day which provides a kind of overlay to the long-term trend.

*Absolute terms means, that the reference was 10V-2ppm at t=0 measured with a (calibrated!) 8508A from PTB's DC cal lab at the maker fair.
At t=1 year it was measured against a calibrated 3458A from PTB at the maker fair (10V flat)
At t=2 years it was measured against a calibrated 3458A from PTB at the PTB and at the maker fair (10V +4,5ppm).

Swings like 6ppm in one month only seem odd to me. As you wrote that you reset your reference after some time I assume your references contain a trimpot. Depending on the adjustement range and maker a trimpot could be a source of drift evil.
I would replace the trimpots by fixed resistors and would not modify the references anymore.

When you track your references and simply take notice of the adjusted value of 10,000 00 V at reset time you loose the drift information over the last period und you will be unable to compile a continuous drift chart.
You might consider recording the voltage BEFORE adjustment as well. 

There is nothing special in my reference. The resistors used are just Yageo MF in TK25.

Best regards
try

[branadic]

Most of the parts were in the junk box, so I could start to assemble the basic circuit. Still there are a few components missing, that I need to order.

-branadic-

[WillTurner]

JimmyJo built a reference with a 7V to 10V gain stage using a Vishay NOMCA resistor array
https://www.eevblog.com/forum/metrology/lm399-based-10-v-reference/msg1604887/#msg1604887.

These packages contain eight 25ppm/C resistors, tracking to 5ppm/C. Note that the datasheet specifications are given as "typical".
Individual resistor values of interest include 1k 2k, 5k, and 10k.

Then z01z suggested
Quote from: z01z on June 14, 2018, 05:38:26 pm

It's possible to do the ratio with one package, as 1.5k/3.5k.

Yes indeed, and surprisingly (to me at least), the 10/7 ratio is independent of the individual resistor values.



Has anyone calculated an overall temperature coefficient for this circuit?

[z01z]

These packages contain eight 25ppm/C resistors, tracking to 5ppm/C. Note that the datasheet specifications are given as "typical".
Individual resistor values of interest include 1k 2k, 5k, and 10k.

Diligentminds had an interesting post about these, saying that after initial burn-in, they are much more stable.

Then z01z suggested
Quote from: z01z on June 14, 2018, 05:38:26 pm

It's possible to do the ratio with one package, as 1.5k/3.5k.

Yes, then I realized that for having a statistical divider, it is better to have more parts. For example, the single R in the upper resistor is the most sensitive to changes.

Yes indeed, and surprisingly (to me at least), the 10/7 ratio is independent of the individual resistor values.

This depends on how ideal your opamp is, you wouldn't get away with say, using a 1Ohm or 1GOhm resistor.

[Vtile]

Edit. Silly me LM399 is drawn partially in the shcematic, not ref. zener type Y.

Not exactly this circuit, but I think what is better variations of that as a servo controller, if I have understood correctly. There the zener leg is connected to 10V output so it is self stabilizing (or oscillating if something goes wrong). Old basic design found at old datasheets of precision OPs etc.

No temp. co. calculations were done, but it shouldn't be too hard to put the values with error coefficients to some math environment.
Post #8,#19 shows picture of the results.
https://www.eevblog.com/forum/metrology/oh-no!-i-got-infection/msg1248065/#msg1248065

Conrad Hofmans original post, where I did get inspiration.
https://www.eevblog.com/forum/metrology/old-fashioned-zener-10v-reference/

JimmyJo built a reference with a 7V to 10V gain stage using a Vishay NOMCA resistor array
https://www.eevblog.com/forum/metrology/lm399-based-10-v-reference/msg1604887/#msg1604887.

These packages contain eight 25ppm/C resistors, tracking to 5ppm/C. Note that the datasheet specifications are given as "typical".
Individual resistor values of interest include 1k 2k, 5k, and 10k.

Then z01z suggested
Quote from: z01z on June 14, 2018, 05:38:26 pm

It's possible to do the ratio with one package, as 1.5k/3.5k.

Yes indeed, and surprisingly (to me at least), the 10/7 ratio is independent of the individual resistor values.



Has anyone calculated an overall temperature coefficient for this circuit?

[branadic]

After some progress the "ION" is up and running!
The missing parts arrived and were soldered into place. Silly me, I routed the power supply of the output buffer (LTC2057) to the wrong pins, but was able to botch some wires to correct for this. (corrected this in rev1 of the board files)
I then had some trouble with my AVR Dragon. Since much time has passed without using it, I didn't remember that the VTG pin is not able to supply the micro with the necessary power and as a result I wasn't able to program it. It took many hours and a lot of search on the web to get me to this point. And it took me another amount of time to get the latest version of AVR Studio (~25k/s download in 21th century is pretty slow, seems there was a bottleneck somewhere in between).
However, now that I solved all this issues I was able to program the Tiny this morning, to set the fuses and to power up the circuit. Hurray "Number 5 is alive" and pressing the buttons changes the ouput voltage. I was able to set it to 10,00000V on my meter.
So this is a good point to start further investigation and improvement on that reference, such as implementing a temperature compensation scheme in software (if necessary) and to find out if something similar is worth for a LTZ reference.

-branadic-

[Andreas]

After some progress the "ION" is up and running!

2nd on this interesting theme within this forum 

Since I use a external (well aged) LM399 for this tests and thus have larger wiring I had also to populate a additional 100nF output capacitor. (on the bottom side).

As population I use 14V for the VCC regulator (for my NiMH battery packs) and 2*LTC2057 for the OP-Amps.
Software is the 22 bit version for the PWM. (8 bit in Hardware + 14 bit dithering).
I have deactivated the key inputs since the NTC (at the moment without function) is using one of the inputs.

The EEPROM constant was adapted to my LM399#3 and so I should have exactly 10 V at the output.
But I got around 420-430 uV more than that.

As reason for the difference I found out the non-symmetrical delay times of the ADG419.
The rising 10V signal has 97.4 ns delay and the falling 10V edge 93.1 ns delay against the 5V PWM signal.
This 4.3 ns difference together with the nearly 10 kHz PWM signal (20MHz / 8 / 256) generates calculatory 42 ppm difference.
Bingo: 42 ppm are 420uV at 10V.

Hoping that the 4.3 ns remain constant over power supply variations and temperature ...

with best regards

Andreas

[Andreas]

Got a unexpected result on PSRR:

Supplied by a programmable linear power supply at the VCC node
(yes: the LT1763 can be back suplied/paralleled without damage).

from 12V to 18V  1.054 mV difference on 10V output.
This gives 176 uV/V or 17.6 ppm/V in average.
from 12 to 14 V the value is even 21.5 ppm/V,

compared to less than 1 ppm/V for the LM399 above 10V (mainly for the heater)
we have to take much more care on the supply voltage stability.

with best regards

Andreas

[Kleinstein]

The supply to the CMOS switch can have quite some influence on the switch. This can be in two ways: one is extra charge injection and the other is a change in the delay for turn on and turn off.  So a really stable supply to the DG419 would be important.

A 200 µV/V PSRR is not that bad.  It needs some care but should not be so hard.

[Andreas]

It needs some care but should not be so hard.

Lets do the math for the current cirquit.
LT1763 has up to 60mV change in bandgap voltage (1220 mV) over a 165 K temperature range.
so worst case up to 300 ppm/K.

Cheap 0805 1% resistors have 100 ppm/K
worst combindation will be 500 ppm/K for the 14 or 15V supply.
Over a temperature range of 30 deg C (which I want to have) this gives 1.5% possible supply voltage variation.
14V * 1.5% = 0.28V max change over temperature.
with 21.5 ppm/V ("typical", since only 1 sample measured) we can get 6 ppm change on the 10V output over the 30 deg C range.

For me that means that we have to check every single power supply cirquit for T.C.

with best regards

Andreas

[splin]

It needs some care but should not be so hard.

Lets do the math for the current cirquit.
LT1763 has up to 60mV change in bandgap voltage (1220 mV) over a 165 K temperature range.
so worst case up to 300 ppm/K.

Ironic considering it's used to power a < 2ppm reference! Those (two) LT1763s are very expensive - over $4.2 each from most distributers (but only $2.3 from Arrow which is great if you are in the States).

Personally I'd have used jellybean opamps and pnp transistors to achieve similar noise (why don't they specify <10 Hz noise for many/most voltage regulators?). TC and time drift would be way better and saving $9 in the process which would have been better spent on a second LM399. It would need a bit of care to ensure it starts up properly though.

[EDIT] Just to be clear, I'm sure the circuit met the designer's own requirements/trade-offs who might, for example, happened to have surplus LT1763s.

[Andreas]

Hello,

by the way: a jelly bean LM317 has specced 100 mV worst case change over temperature range
of the reference voltage (1250 mV) resulting in nearly 500 ppm/K.

So the LT1763 is not too bad.
Besides other advantages (low power consumption,  low noise, low voltage drop).
So ideal for battery powered equipment like voltage references.

with best regards

Andreas

[Kleinstein]

If the supply would be a real problem, there would be still the option to use the existing LM399 reference and use an OP as an amplifier from there.

The regulator may not be that bad: at least in the typical curve there is quite some curvature / slope at the upper end of the temperature range. In addition there are resistors that are better than 100 ppm/K and the calculation was worst case for the resistors.

I don't think the supply should be so critical that one would need to have a super low noise regulator. Keeping the emissions from the DG419 switch away from the reference and OP could be the more important point.

A totally agree that it would be nice to have data in the DS on the low frequency noise (e.g. 0.1 Hz to 10 Hz or even lower). However testing would take quite some time by definition. A band-gap reference typically used has quite some noise, but at least it should be reasonable predictable, so that typical data would already help.

[Andreas]

Hello,

another notice to the measurement of trise and tfall:

the loading by the 10:1 scope probe (10Meg / 15pF)
of the PWM signal at the processor changes the output voltage by -9 ppm.
The loaded PWM signal at the analog switch output is changed by +18 ppm.

So the results above for trise, tfall are already changed significantly by scope probe loading.

with best regards

Andreas

[MiDi]

Personally I'd have used jellybean opamps and pnp transistors to achieve similar noise (why don't they specify <10 Hz noise for many/most voltage regulators?).

Think the noise <10Hz from vregs does not really matter for circuits incorporating e.g. opamps with their high PSSR at low frequencies.
Issues arise for higher frequencies where most circuits suffer from decreasing PSRR.

[David Hess]

At low frequencies, power supply rejection is high so low frequency noise is not a problem.  At high frequencies, passive filtering is effective so high frequency noise is also not a problem.  The exceptions are circuits which have no power supply rejection like single ended logic and low noise oscillators where the noise in the power supply limits the ultimate performance.

[Andreas]

Hello,

first T.C. measurement on branadics PCB.
This time the VCC is externally supplied (14V voltage regulator at room temperature).
Temperature sensor is at the ADG419 switch.

Box TC:
0.0366 mV max change over 22.43  deg C  = 0.1632 ppm/K

Linear regression curve:
-0.144 ppm/K

3rd order regression curve @25 deg C:
-0.124 ppm/K

the resulting error to the linear and 3rd order regression curves is not significant different.
so for a temperature compensation of the switch alone a linear correction would be sufficient.

with best regards

Andreas

[Andreas]

Hello,

2nd measurement
this time with 14V generated from LT1763 on PCB.
Temperature sensor is again at the ADG419 switch.

Box TC:
0.030 mV max change over 23.2 deg C = 0.131 ppm/K

linear regression curve:
-0.117 ppm/K

3rd order regression curve @25 deg C:
-0.117

so the result is slightly below that with external stabilized VCC.

when looking at the stabilized voltage over temperature
it gets clear that the negative TC of the voltage regulator (-65 ppm/K) together with the negative PSRR value of -22 ppm/V gives 0.02 ppm/K reduction of the negative T.C. of the switch.

Box T.C. of LT1763
0.021 V
23.2 deg C
-64.8 ppm/K

with best regards

Andreas

[branadic]

Thanks Andreas, that's not to bad. So as far as I understand it is possible to improve the output voltage by T.C. compensation in firmware. The question left is towards linearity of the PWM.

-branadic-

[Andreas]

Hello,

I do not think that linearity will be a large issue.
The different switching times may give a offset.

Another question for me will be noise.
Although with my 6.5 digit instruments I did not see anything unusual up to now.

with best regards

Andreas

[Kleinstein]

The -60 ppm/K for the regulator is a relative good case - it can be considerably higher, though normally not as bad the worst case.
Even with the worst case TC of the supply, the overall TC would not be much worse than the TC of the LM399.
Besides the effect via the supply, the DG419 could also have an TC from a change in the T_on and T_off.

Another point worth looking with this circuit is the relative high resistance of the filter. So bias current from the OP and maybe leakage could be a problem. Just a constant bias would cause an extra offset (e.g. 100 pA * 300 K = 30 µV).

For a reference circuit linearity of the PWM is not that important. It is more about long term stability and to a smaller extend the TC.
With the rather large resistance nonlinearity should not be that bad. As an crude estimate something like the difference in high / low resistance of the DG419 divided by 100 K could be a crude upper limit. So maybe 1 Ohms / 100 K or 10 ppm.

[Andreas]

the overall TC would not be much worse than the TC of the LM399.

Near room temperature the typical T.C. of a LM399 is not worse than that of a LTZ1000.
So a additional T.C. of >100 ppb/K matters.

with best regards

Andreas

[Andreas]

Wideband noise measurement of branadics PCB.

51uV RMS over 10 Hz - 100kHz
with only 210uV peak/peak. (so shurely no normal distribution).

in zoom we see that the 10kHz clock from the switch is dominant
-> 40uV RMS or -88dBV for the fundamental.
Frequencies above 10kHz are filtered by the switch/filter.

As comparison the input noise of the used LM399#3
59uV RMS with much higher peak to peak noise and significant parts above 10 kHz.

and finally the noise floor of the used low noise amplifier according to AN83 together with a 10.8V NiMH battery.
0.43uV RMS so enough distance to the measured values.

with best regards

Andreas