LT1021CMH based mini-reference

[branadic]

Having only 2 references for regarding differences over time is not enough.
You never know if there is a slow drift in between how the absolute drift over time is.
So you need at least 3 references to sort out the bad one.

Let me add here, that these three references should in theory be completely uncorrelated. Having three references from the very same batch adds some potential correlation to them. You wouldn't notice if all three references simultaneously jump by the same amount into the same direction, when only observing their differences.

-branadic-

[ejd.pol]

@Andreas:

Thank you for the explanation of the trimming process. Indeed, if trimming is done only for three fixed temperatures,
the TC for other temperature points may have a systematic deviation from zero.
Would you agree that, in theory, and for a large enough population of devices, the TC at these three trimming temperatures should have zero average?
(The average is never exactly zero of course, but presumably it should be significantly smaller than the datasheet value valid across the entire temperature range.)

Using a DAC to generate the trim voltage sounds like a neat idea! A processor can easily interpolate (polynomially to any order) between a number of predetermined table entries.
If the trim value is determined at, say, 5 C intervals from 0 to, say, 80 C, the table contains only 17 values. Assuming that the TC is a relatively smooth function of temperature,
third order interpolation of four table entries should suffice (hopefully!). That is a lot less work than determining the optimal 3rd order polynomial per interval.

But I guess one runs into the question of how much the DAC output drifts? I guess the DAC resistor chain be set by the reference to be trimmed?? (Similar to bootstrapping a zener...)

[iMo]

A "disciplining" of the LT1021 has a sense only when the DAC plus its own Vref_DAC reference is "much more" long term stable than the LT1021 itself.. It works nice with OCXO discipling where the GPS is indeed "much more" long term stable.. In order to discipline the LT1021 you would need at least a 399/1000 based Vref_DAC, imho..

[alm]

Thank you for the explanation of the trimming process. Indeed, if trimming is done only for three fixed temperatures,
the TC for other temperature points may have a systematic deviation from zero.
Would you agree that, in theory, and for a large enough population of devices, the TC at these three trimming temperatures should have zero average?
(The average is never exactly zero of course, but presumably it should be significantly smaller than the datasheet value valid across the entire temperature range.)

I don't see why. Why would the population mean be expected to be close to zero? Why not +10 ppm/K? Or -5 ppm/K?

[Andreas]

Would you agree that, in theory, and for a large enough population of devices, the TC at these three trimming temperatures should have zero average?
(The average is never exactly zero of course, but presumably it should be significantly smaller than the datasheet value valid across the entire temperature range.)

I would not make any assumptions about the curvature (local temperature gradient) of a trimmed device.
You can only expect that by trimming you get more higher order curvatures than untrimmed devices.
(compare a simple TL431 datasheet  with a LT1019 or LTC6655 datasheet for "typical" response over temperature).

with best regards

Andreas

[iMo]

Out of curiosity I put into my sim above the TC coefficients from Andreas' post https://www.eevblog.com/forum/metrology/ad587lw-10v-precision-travel-standard/msg1547132/#msg1547132
The no-comp shape identical with that from Andreas, and the compensation with diodes (an example only), see below.