Parallel voltage references?

[Marvin]

Inspired by Daves calibration video I searched thru my chips and found 2 of each MAX6325/MAX6341/MAX6350 chips I sampled from Maxim back in 1999 MAX6341 have 1998 datecodes and other are 1999. I chucked them on my breadboard and behold - those 2.500 and 5.000 are SPOT on my Fluke 87 V that I just bought but was manufactured in 2010 (I had a friend visiting USA, the price of Flukes is CRAZY in the EU).

Now I am thinking about making myself a voltage reference board with 2.500, 4.096 and 5.000 voltages, but instead of using one chip is it possible to parallel them up? With current output DAC's they have been using 2 and more in parallel to gain lower noise floor and chipamps use parallel LM3886 to get more output by placing a 0.1 or some other low value resistor in series with each chip output. But coming from IT and not electronics I am not completely sure how to think about this with voltage references.

Any thoughts?

Datasheet @ http://datasheets.maximintegrated.com/en/ds/MAX6325-MAX6350.pdf

[SeanB]

Why, you put them in parallel across the power supply and have the outputs separate to test points.

[Marvin]

I have them all 6 in separate now (I'll do a plethora of measurements in the evening 2.500 vs 4.096 and 4.096 vs 5.000 difference measuring should theoretically give me the best ones out of the six and my Fluke). But I was thinking like in DACs, when you use two (or more) DACs the signal to noise ratio gets 3dB better and that's what I am wondering now. If I already have them if I could use them in parallel then the "noise" or signal error should be technically better? If one drifts and the other doesn't then the other should make the drift smaller or not?

[saturation]

Its not likely to succeed, because the drift of references of a model type tends to be positive or negative, not both, so they'd not likely cancel each other out.

Further testing to 1mV precision, you won't detect the drift, its typically in the uV range at the least.   If a reference had an annual 30ppm + or - drift,  [ the good news is internal drift eventually plateaus with age, so the rate of change will eventually stabilize and be less than the quoted spec],  say on your 5V reference as example, that would be 150uV per year; it can take 10 years of continuous drift in the same amplitude and direction before it would register 1mV change.  In that same 10 years, the rate of change will also drop, so the drift of 30ppm will also decrease, and this offsets the likelihood that the actual value of the reference is not 5.000 000V to begin with, so that it would really take 10 years to register a 1mV change.

[alm]

At the very least it should cut down on noise. I'm not very familiar with these particular references, but I wouldn't expect the drift to be the same across multiple devices, since that would be trivial to compensate for by the designer. Whether the drift is truly independent and will be reduced by a factor sqrt(n) for n parallel devices remains to be seen. The late Bob Pease, who knew a thing or two about voltage references, designed a circuit with a bunch of parallel LM399s for improved noise and stability.

[BravoV]

Though not in parallel, maxim has an app note on stacking them to lower the noise -> How to Reduce Reference Noise by Half

[saturation]

https://www.eevblog.com/forum/general-chat/interested-in-hobbyist-voltage-reference-standards/msg13911/#msg13911

Here are the archived discussions and links for more details.