MAX6226 voltage reference

[hgl]

This is a short paper from Maxim about the long-term stability of the MAX6126AASA25

[dietert1]

The Maxim paper is interesting in that they confirm aging may take up to 10 000 h. Maybe this is the proper time scale for references running at room temperature, while for hot references like the LM399 and LTZ1000 the time scale seems much shorter, more like 2000 h.

Regards, Dieter

[Andreas]

Hello,

I (dont) like these "artificial" tests "with sockets (non soldered)" at constant temperature and humidity keeping all stress from the references away.

What has this to do with a "real life" application in industrial environment?
Even a short period (perhaps 2 days at the week end) with different humidity which has a time constant of several days (more like 5-7 days) has more influence than all other effects during the whole 3kHrs.

with best regards

Andreas

[2N3055]

The Maxim paper is interesting in that they confirm aging may take up to 10 000 h. Maybe this is the proper time scale for references running at room temperature, while for hot references like the LM399 and LTZ1000 the time scale seems much shorter, more like 2000 h.

Regards, Dieter

AD588/688 also settled mostly after 2 months (cca 2kh) and was fully settled after 3kh(cca 3 months).. It is a CERDIP case.
It has quite high internal dissipation, so you could consider it heated reference from a standpoint of thermal stress though...
I believe it is mostly mounting stress of the crystal, i.e. interaction between crystal and mounting material relaxation..

[kleiner Rainer]

Hello,

I (dont) like these "artificial" tests "with sockets (non soldered)" at constant temperature and humidity keeping all stress from the references away.

What has this to do with a "real life" application in industrial environment?
Even a short period (perhaps 2 days at the week end) with different humidity which has a time constant of several days (more like 5-7 days) has more influence than all other effects during the whole 3kHrs.

with best regards

Andreas

Hobbyists are not the market they aim for. In industrial applications, you want stability. Removing a sensor unit from a production line for calibration is expensive in downtime and handling costs. The aim is to calibrate it once (after burn-in), install it and forget it. The sensor is sealed, and if necessary fitted with special membranes to allow for pressure relief and to keep humidity out, combined with humidity absorbing material. Anything that needs extra precision lives in a controlled environment (temperature, humidity, EMC). My employer has such an environment for the precision measurements toolmaking (diecast and plastic molds) needs.
I suspect the measurement regime Maxim chose is based on some MIL Std. and as such it is easily reproducible by anyone with similar equipment. Changing temperature and humidity (simultaneously? one after the other?) requires more detailed specifications: rate of change for example. How comparable are the measurements if your climate chamber is slower in changing temperature and/or humidity?
Standards often also specify the number of specimens you must test to get statistically relevant results - this gets expensive for hobbyists if precision references are involved 

BTW "artificial" tests are the norm in industry standards work: take the LISN used in EMC work. The typical line impedance specified in the sixties is as artificial as it gets, but it is useful because everyone measures to the same standard, so results are comparable.

Sockets are not a problem, if you use the special type that is designed for burn-in: Teflon body with precision gold plated contacts - Fischer Elektronik still manufactures them:
https://www.fischerelektronik.de/web_fischer/de_DE/Steckverbinder/F03/Fassungen%20f%C3%BCr%20TO%20...-Geh%C3%A4use/search.xhtml

To sum it up: this reference is NOT designed for metrology grade applications. It is nice if it works for our voltnut aims, but there is no guarantee. It works for applications that need a "set and forget" approach as in industrial applications, where calibration after burn-in is done once. We are not the market. Period.

Greetings,

Rainer

[exe]

I ordered ten max6226 . I plan to make a pcb with eight ICs: four mounted normally (well, kinda, I don't plan to solder pins that are not internally connected or not supposed to be connected), and four to be mounted as a "dead bug".

Also ordered several humidity and temperature sensors from aliexpress.

[Andreas]

Hello,

I ususally bend the PCB to see if the reference is influenced by PCB stress.
If the bending ends with <5 uV influence on the 5V reference (1 ppm) I usually stop.
But this ends usually that I solder only the GND pin of the reference to the PCB (near the NTC for the temperature sensing) and the rest of the pins with thin VERO-wire.

with best regards

Andreas

[exe]

I'm back . I got the ICs, now designing a pcb. The board will have two 74hc4051 to switch between eight references.

[Andreas]

The board will have two 74hc4051 to switch between eight references.

Why not MAX4051A?
And where are the bypass capacitors from Pin 3 of the Muxes to Gnd?

With best regards

Andreas

[exe]

Why not MAX4051A?
And where are the bypass capacitors from Pin 3 of the Muxes to Gnd?

I didn't use MAX4051A because I have doubts it will make a difference. The real leakage is probably below 1nA. I'll check on breadboard if I don't forget.

As of bypass caps, it's not on any schematic in datasheets, I checked several datasheets. I concluded the mux is too simple to need a bypass cap. I also used 4052 switch on breadboard without bypassing, it worked just fine. If it had an internal clock, or brown-out detector or something, I'd include include it.

[Kleinstein]

Bypassing is a good idea. In the moment the CMOS muxes are switching they take quite some supply current. For the PCB it is easy to include.

I would consider to also have optional resistors to create an average over groups of some 2 or 4 reference for later use.
It makes not much sense to first burn in the reference and than later unsolder. A more realistic case would be cutting the PCB in parts and use smaller parts, like 2 references in parallel.

Usually the leakage currents should be OK, but one never knows for sure. The chips are not tested for low leakage and 74HC... is more like made in a low cost process / fab.

A place for a temperature sensor would be also nice for the PCB.

[exe]

As of temperature monitoring, I bought two temperature and humidity sensors, Si7021 and SHT31. I quickly tested the first one, humidity seems to be working, but temperature resolution is too low (like ~1C or so). Perhaps, need to fix initialization code.

About averaging. I think I can just do it in software. This way I can mix arbitrary references on the board and see their combined performance (but that's assuming my K2000 is stable).

Making a snap-off board sounds like an interesting idea, because that allows selecting references with smallest drift. Currently I don't have much plans to use the references other than just discover their properties. I'm also a bit skeptical about performance of these references.  May be in the future, if max6226 proves it's worth it.

[exe]

Pre-final version of the board. Changes:

1. Added a small decoupling cap shared by both switches just in case. Routing is getting a mess...
2. Finished routing, reduced board size, etc.
3. Added reverse polarity protection (diode)
4. Added an output capacitor after switch for filtering output.
5. Top copper pour connected to +3.3V.

I also include schematic of the reference. The header allows kelvin sensing, and also allows activating (with a jumper) a loading resistor. I plan measure the voltage changed with 1mA sinking current. The loading resistor is quite close to the reference, hope its heating won't affect the reference much.

What I want to try now is to make a dead-bug mount footprint for four references out of eight to see if this makes any difference.

PS Can 74HC4051 switch signals slightly below Vee? Like, a millivolt or two? Datasheets say no, but I'd expect it to handle some 0.2-0.3V beyond rails.

[Kleinstein]

I would also expect the 4051 to also handle up to some 200 mV beyound the supplies. There would be some some additional leakage though from "diodes" slightly forward biased.  At least the HC4053 / LV4053 / CD4053  do work that way.

The voltage regulator is quite close to the 4051 - this may cause some gradient there.

[Andreas]

PS Can 74HC4051 switch signals slightly below Vee? Like, a millivolt or two? Datasheets say no, but I'd expect it to handle some 0.2-0.3V beyond rails.

Hello,

Yes some millivolts should be no problem.
As long as the input protection diodes do not get conductive.
Above 100mV the leakage will increase.

with best regards

Andreas

[exe]

The voltage regulator is quite close to the 4051 - this may cause some gradient there.

You guys have sharp eyes, and I didn't even post the schematic . Moved the reg a little bit.

This is the board I'm going to order. There is a room for improvement, but as long as there are no critical bugs it's fine.

PS board thickness will be 1.0 mm so it's easier to bend in bending test.

PPS I think text needs to be on silkscreen layer, trap for young players

[exe]

The board is fully assembled and all channels seem to be working. Supply voltage after lm317 is 3.32. Now I'm writing a script to do automatic measurements.

[exe]

Very first plot of all references. Vertical scale is 50uV. I guess we all want to see deviation in ppms from a mean value, right? .

[Kleinstein]

There is quite some correlation between the curves, especially around 60% of the x axis. There seem to be a common effect of some effect of the measurement system.

How are the voltages measured ?

Except for the yellow and brown curve the drift so far looks acceptable, but hard to really tell.

[Andreas]

Very first plot of all references. Vertical scale is 50uV.

And the time scale? I interpret 2 seconds per axis-tick (but this would mean many hundreds of measurements per second?)

with best regards

Andreas

[iMo]

There is quite some correlation between the curves, especially around 60% of the x axis. There seem to be a common effect of some effect of the measurement system..

@exe: You have to plot the temperature of the pcb as well. The correlation there comes from a pcb/env temperature change, I guess..

[exe]

Last couple of days I was busy with trying to make meaningful plots out of the data I've collected. I'm attaching the test output, which is smoothed by moving average of 50 points. Sample rate is 6 samples/min (i.e., every 10s all references are scanned and measured). So far I'm very happy with results. I wanted to have some +-5ppm stability and it seems I got it. The setup is quite poor, I'll describe it below. So, probably, real performance is even better than on the plots.

As of temperature and humidity collection: working on it. I have issues with it, it gives very limited resolution. Figuring out why.

For debugging reasons the board is just laying on my table, and cats walk on it from time to time, causing disturbance in measurements. I plan to make an enclosure for it. As of connection to DMM, I use very cheap leads for now. I plan to replace that once I make enclosure and setup temperature/humidity sensor properly.

[exe]

I added temperature and humidity logging yesterday. Now making an enclosure. What surprises me is how stable voltage was. It seems tempco is about +1ppm/C. The disturbance after 7am is when I switched on fan, it was creating air disturbance. Temperature variation over night is much smaller than I expected, even with windows open.

Current connection is quite crappy (those red and black crocodile on the photo). Unfortunately, I cannot find banana plugs in my lab to make a good connection. I wonder if I can take some better DMM leads, cut off tips, and add a female 2.54mm connector so I can plug it in. Or those 2.54mm headers (commonly used for breadboarding) are not suitable for measurements? I can make them tight fit...

[exe]

I've made an enclosure, and also test leads (on photo).

There is +-0.5ppm or so variation from measurement to measurement. To remove that, I apply heavy averaging (moving average of 100 measurements), otherwise it's hard to see eight plots simultaneously. I attach what a plot looks like without averaging.

Some facts about max6226:

*. Price raised from 7.83 to 8.32 euro (before VAT) on Mouser.
*. Average voltage today across all measurements is 2.500032252142955
*. It looks like ic3 has some popcorn noise

[dietert1]

Precision work at 2.5 V! Point to point RMS noise should be a fraction of a ppm, lets say 1 uV. I think that is a very good result when using banana plugs. I am also surprised you get that with those standard CMOS MUXes. Good news! Unit 1 seems to drift in relation to the others, so it may be the second one to reject, but that's a bit early to tell.

Regards, Dieter