9½ Digit Multimeter, feasible?

[dietert1]

Sometimes i think this forum doesn't support scientific discussion. Whenever i tried, my contributions were buried under a pile of .. We can hardly mention the 3458A except for praising it and a thread like this one is doomed.
Anyway let me mention that i got a nice result today regarding the making of a constant temperature, constant humidity and constant pressure oven.

Regards, Dieter

Edit: A 2018 NIST report of a 0.1 nV comparison between two programmable Josephson voltage standards (11 digits).
https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=925369#:~:text=This%20paper%20presents%20results%20of%20the%20first%20fully%20automated%20direct

[Echo88]

You beat me to it Dieter.
Im not that quite there yet with my constant T/P/H chamber which i intent to use for the ultra low noise Vref project with supercaps and foil caps, hope to post some stability results in a few days now that i got those BME280 working
Temp regulation right now assures an attenuation factor compared to ambient temp variation of about 250 using an analog PI-regulator.

Any opinion if a bipolar current source for resistor measurement would be beneficial in an imroved DMM compared to the regular current active/inactive for TEMF compensation like Fluke does with their "Tru Ohm" claim? I dont see the benefit.

[Kleinstein]

There is a small advantage in switching between a postive and negative current instead of current on / current off. It gives more signal with the same self heating or less heating with the same signal. It depends on the resistor range if this is relevant and if switching the current is used at all.
Allowing postive and negative currents also gives some consistancy check as 2 separate values.

[O.B.Don]

Sometimes i think this forum doesn't support scientific discussion. Whenever i tried, my contributions were buried under a pile of .. We can hardly mention the 3458A except for praising it and a thread like this one is doomed.
Anyway let me mention that i got a nice result today regarding the making of a constant temperature, constant humidity and constant pressure oven.

Regards, Dieter     

Dieter, please do not despair!  The contributions that you and the majority of the other people posting to this thread are useful and significant.  I am glad for the original post by laichh on "How far away are we from the 9½ digit multimeter?" as this has stimulated a lot of useful information and discussion on what it would take to actually implement a 9½ digit multimeter. 

I started a personal project slightly over 12 years ago to investigate the issues needed to implement such a meter.  I only became aware of the EEVblog about 3 years ago and this thread much more recently, and believe I have found a "home" and "like minded people" to discuss aspects of such an implementation. I hope that I will be able to contribute to this discussion in the near future.

Cheers!  Don

(Please excuse my post format, still trying to figure out how do do a "proper" quote in this system - - -)

[laichh]

Hi O.B.Don,

I am glad that my original question triggers such a chain effect.

Take a look at the datasheet of LTC2500-32, on page 8, the chart for DC histogram with different number of averaging from DF = 4 to DF = 16384, it shows that the standard deviation of the histogram shrink with respect to sqrt(DF). That's the power of math & statistic, just Google "Central Limit Theorem" to learn more.

[O.B.Don]

The challenge with the LTC2500-32 ADC is that digital filtering will reduce the spread of the data at a single value input from slightly over ±7.5 ppm with no filtering to about ±0.1 ppm with DF=1684 for a SSINC filter. 

Averaging or digital filtering however may not necessarily reduce the overall INL of the device, so the device's INL spec may ultimately limit the the "accuracy" of the device. 

Improvement in the INL, by averaging, is dependent upon the type of ADC, (delta-sigma vs SAR for example) and how smooth and predictable the INL curve is or isn't.  If the INL curve is sufficiently smooth and predictable it allows one to calibrate the ADC's range at a smaller number of points and then develop a calibration compensation equation that with sufficient noise and averaging may allow an effective INL that is better than the published specifications.

If an ADC has a low INL by design, then averaging of the signal along with enough noise to span several LSB digits of the ADC can result in more effective calibrated bits.

[Kleinstein]

The averaging is only effective like the square root of time if there is not extra 1/f or similar low frequency noise. The auto zero switching / chopping can avoid this for a large part of the ciruit, but there is still a part before the switches that is sensitive to thermal variations. Also the reference side has 1/f noise.  So the idea of using more time is somewhat limited. For the really long times there are usually also effects that make things worse the longer it takes ( the Allan variation curve goes up again). So it really makes sense to start with low noise.
Some of the very slow meters (Solartron 7081 , Fluke 8508, Datron1281) take 50 or 25 seconds for a single 8 digit reading. Much slower gets really inconvenient for the user too.

Very long time averaging only works well if the signal source itself is modulated. I have used long time averaging with a lockin amplifier up to 1000 s averaging per point and it was still roghly on the 1/sqrt(time) slope. However it usually does not work that long for DC.

[iMo]

The challenge with the LTC2500-32 ADC is that digital filtering will reduce the spread of the data at a single value input from slightly over ±7.5 ppm with no filtering to about ±0.1 ppm with DF=1684 for a SSINC filter..

Did somebody try with the LTC2500-32 actually?? I'm still looking for an opensource ADC around this chip (having them for years in my junkbox among the others famous "32bitters")..