How does the HP 3458A mitigate the LTZ1000 noise?

[benj38]

My apologies in advance for a possibly too simplistic question.

Looking at the LTZ1000 datasheet, I see a noise specification of 1.2 microvolt p-p.
https://www.analog.com/media/en/technical-documentation/data-sheets/LTZ1000.pdf
I know that there are many aspects to noise, but let's just focus on this one for a moment. I guess this number correlates with the graph titled "Zener Noise" with an x-axis running from zero to sixty seconds.

Now, for a typical zener voltage of 7.2V, 1.2 microvolt is about 0.17ppm. 8.5 digits means a resolution of 0.01ppm. Hence, it looks like the least significant digit should be all over the place, and basically meaningless. What am I missing here?

Does HP select individual units with at least an order of magnitude less noise? Alternatively, do they average readings to get a stable 8th digit (this would mean that one can assume the noise has a mean value of zero)? Something else?

I would love to be enlightened 

[Andreas]

What am I missing here?

a) you have a integrating ADC which filters out most of the noise.
b) look at the 3458A noise specification (over NPLC setting)

[Kleinstein]

The 1.2 µV LTZ1000 noise is for 0.1 to 10 Hz peak to peak noise. To get 0.01 ppm noise from the 3458 is at zero input and 100 PLC integration which is some 0.3 Hz of bandwith. In addition this is RMS noise - so some 6.3 times higher. When actually measuring 10 V the 3458 noise is higher as more of the reference noise adds to the result. It usually won't be the 0.1 - 10 Hz band but more like 1 - 30 mHz. Overall one would not really get to the 0.01 ppm noise level.
There is also the DUT noise - in many cases the 10 V reference would have a comparable or higher noise than the meter. So there is limited use for a meter internal reference that is much better than the signal sources one can expect.

In the 3458 they actually missed out a bit on removing some of the LTZ1000 higher frequency noise that they could have easy filter out. Not a big deal (maybe 10%), but an easy part they missed.  A second part (aliasing around 3 Hz) is already tricky to handle and may not be worth it.

[EC8010]

It's not a simplistic question at all. It's more informative to look at the graph to the left of the 60s graph; the noise spectrum. That shows that the noise of the LTZ1000 reference is near-enough white down to 3Hz. If you use an aperture of 10PLC, you are averaging over 200ms (50Hz mains), which is the period of 5Hz. Thus, at 10PLC, you are averaging very nearly white noise, which averages to zero over an infinite time. But even that white noise has a low frequency component that won't average out in 10PLC, so you're right; a problem remains.

Ultimately, all voltage references are noisy and whilst you can filter out high frequency noise, you must always use a low-pass filter (to allow the DC through) so very low frequency noise always comes through. That means you have to attend to the noise of the voltage reference. Read a few data sheets and you'll quickly find that they say to shield voltage references from air currents that can change temperature. Another issue is mechanical stress, which is why PCBs for references have cut-outs around the reference. You can prove both of these effects to yourself by lagging a voltage reference with foam and noting the improvement in noise and by pressing a scriber onto the top of a (plastic-encased) reference and seeing how you can change the voltage. There's been lots of discussion about the control circuitry around LTZ1000 to maintain constant temperature. After that, you can select references to find the most stable one. Finally, you use your meter in a stable temperature-controlled environment.

[TERRA Operative]

In the 3458 they actually missed out a bit on removing some of the LTZ1000 higher frequency noise that they could have easy filter out. Not a big deal (maybe 10%), but an easy part they missed.  A second part (aliasing around 3 Hz) is already tricky to handle and may not be worth it.

I wonder if it is worth building a replacement vref module for the 3458A with these improvements?
I have idly wondered about the benefit of a multiple LTZ1000 vref board too?

[MiDi]

It is not only the voltage reference that contributes to the overall noise, the reference gain stages, the ADC and the AFE contribute additional noise.
Lowering the reference noise in 3458A gives some improvement, especially near FS and LF: A9 ADR1000 (330nVpp 0.1-10Hz) vs LTZ1000 (factory) comparison

[benj38]

Just for me to understand -- since I never had the privilege of using a 3458A -- when one uses it to measure 0V (or as close to it as one can due to thermocouple effects etc.) say at 10PLC, is the last digit stable between readings, or does it change at every reading basically spanning all possible values from 0 to 9? What about at 100PLC?

[Kleinstein]

There are plenty of reports of the 3458 noise and the noise is often not that much better than the specs.
In the 10 V range one can expect some 100 nV RMS for 100 PLC and some 300 nV for 10 PLC. So the last digit will still move a bit around, though not all over the place in the 100 PLC case, but nearly in the 10 PLC case.

[MiDi]

The Allan Deviation for input short on 10V range shows minimum noise at around 1h of averaging for 1NPLC readings (~24h data for each line).
From that plot you can estimate peak to peak values by multiplying with 6 for each averaging time.
For 1h averaging it is around 20nVpp, a change in 9th digit by 2-3.
Keep in mind that at those long averaging times other effects come into play, mainly TC.

[CalibrationGuy]

In the 3458 they actually missed out a bit on removing some of the LTZ1000 higher frequency noise that they could have easy filter out. Not a big deal (maybe 10%), but an easy part they missed.  A second part (aliasing around 3 Hz) is already tricky to handle and may not be worth it.

I wonder if it is worth building a replacement vref module for the 3458A with these improvements?
I have idly wondered about the benefit of a multiple LTZ1000 vref board too?

No need to wonder. XDevs already did that upgrade with 4 references and recorded the upgrade live.

TomG.

[TERRA Operative]

Unfortunately the xdevs website has gone offline with a political message shown instead.

I'll take a look on youtube.

[Kean]

Unfortunately the xdevs website has gone offline with a political message shown instead.

I'll take a look on youtube.

If you have a login from Illya for the xDevs site, you can still access the docs repository.
I see a few xDevs projects named QVR* which look like different attempts at a quad voltage reference for 3458A, but I think you cannot access the text of articles.
The main article would be https://xdevs.com/article/qvref/
Some comments about QVR are at https://github.com/tin-/adr1000
An older video about it at

[benj38]

Unfortunately the xdevs website has gone offline with a political message shown instead.

I'll take a look on youtube.

WayBackMachine to the rescue: https://webcf.waybackmachine.org/web/20241208162536/https://xdevs.com/article/qvref/

[benj38]

So if I understand correctly:

• The 3458A does average (i.e., integrate) to get rid of some of the noise. This is inherent in the way the ADC works.
• This averaging works because above around 3Hz it is basically white noise with a mean of zero.
• The more PLCs are used by the ADC, the more averaging one gets, and more of the (relatively) higher frequency noise is removed.
• Nonetheless, the noise spectra at lower frequencies cannot be gotten rid of, even at the maximum 100PLC setting.
• In the end,  as I conjectured in the initial message in this thread, the remaining noise overwhelms the last digit which is fluctuating all over the place at 10 PLC, and even at 100 PLC according to the graphs by @Midi (I am a bit confused still, since @Kleinstein says it does not fluctuate much at 100PLC).
• To further remove noise one can use, e.g., the math function to average readings over much longer time periods, but this is limited by other effects like thermal drift and the non-necessarily-zero-mean characteristics of the noise at these very low frequencies.
• Obviously, the reference is not the only source of noise, but as I wrote at the beginning, I was just focusing on that aspect in my question.

[Kleinstein]

The averaging works as a low pass filter and does not need the noise to be white. It is just easier to calculate with white noise.
With 1/f noise one has however the problem that more averaging can also require a longer time window to look at and this way not gain much.

With a shorted input one has amplifier and ADC noise. A small part is also from high frequency reference noise that can alias down to near DC. This small part (some 10% at 1 PLC or 5% at 10 PLC as an estimate) of the reference noise could be filtered out. I doubt it would be worth modifying the ref. in the 3458 for this relatively small improvement. If one would build a new ref. module anyway it would however be good to add a little fitlering. E.g. TIN in his 4 x ref. module could add a single capacitor at the averaging point.

The effect of normal LF reference noise is proportional to the voltage measured and can dominate at full scale and relatively long time scale, as it contains quite some 1/f noise. There is not that much that can be done about this LF reference noise, except a lower noise ref. and maybe more in parallel.
With 100 PLC and relatively large voltage the 3458 can have less ADC noise than reference noise.

[benj38]

The averaging works as a low pass filter and does not need the noise to be white. It is just easier to calculate with white noise.

But the noise must have a zero mean. Correct?

[Kleinstein]

A low pass filter removes the higher frequency part. This is by definition does not contrain a DC part.