3458A autozeroing

[splin]

The HP journal article has this to say about the ADC:

At lower resolutions the ADC's rundown time becomes a more significant portion of the overall measurement time and therefore pulls the ADC away from the theoretical limit. At higher resolutions the 1/f noise of the ADC forces a measurement of zero several times within the measurement cycle to reduce the noise to the 8Vz-digit level. This also reduces the measurement speed.

How are we meant to interpret this? Are they referring to the fact that if you specify NPLC greater than 10 it actually averages n x 10 NPLC measurements and that *if* autozero is enabled it will subtract n x 10 NPLC zero measurements? The word 'forces' is confusing me given that autozero is optional.

Also I see that using the APERTURE command, an integration period of 1s can be used. Presumably it then alternates each measurment with a 1s integration time zero measurement if AZ is enabled. But how does this impact 1/f noise performance - is it much worse than  5 x 10NPLC AZ measurements (50Hz line frequency)?

[Kleinstein]

The 1/f noise of the integrator and input amplifier limits the usefulness of very long integration times. For this reason they choose to not provide real 100 PLC integration, but use 10 PLC signal and 10 PLC zero followed by averaging instead. This gives auto-zero function and suppresses 1/f noise below about 2.5 Hz. 

Using 10 times a 10 PLC integration without a zero measurement in between would not be effective in suppressing 1/f noise, as there is no analog  zero function (like in the 3457) . So from a certain integration time one expects lower noise if AZ mode is used, even though half the time is lost to the zero phase. The exact turn over would depend on signal noise.

The downside of  averaging short integration times is that there is more time lost to the rundown and there is more noise from the rundown part.  For the HP input design, faster AZ cycles could also increase input bias. So there is a kind of optimum - that seems to be around 10 PLC for the 3458. Modern DMMs like the Keithley DMM7510 seem to use even faster (e.g. 3 PLC) switching between signal and zero.

I am not sure how the aperture mode is realized, but a would assume it should be just a different way of setting PLC setting.

[splin]

Thanks Kleinstein. I understand the principles but my question mainly related to the HP Journal statement that '1/f noise forces...', given that AZ is optional - I wasn't sure if they meant that some zero measurements are forced regardless of the AZ option. Given the specified reading rates v NPLC v AZ, I see that can't be true so I guess the HP Journal description is a simplified explanation and not meant to cover all situations.

I also see that the APER 1s v NPLC50 noise issue has come up before: https://www.eevblog.com/forum/metrology/hp3458a-ocomp-delay-related-problems/msg1464930/#msg1464930

In that thread TIN did some resistance measurments with APER 1s and NPLC 100, but it isn't stated if AZ was enabled or not but I assume it would have been the same for both cases. I would have expected the std dev NPLC100 measurements, with twice the integration time of APER 1s, to have been SQRT(2) better than shown if AZ is off and better still if AZ was on (due to 1/f suppression), but most results aren't that much better than the APER 1s tests so I'm not sure what to conclude.

[TiN]

AZ was enabled on all my measurements. I got significantly more noise when use APER 1 (when ADC integrates for 1 sec, instead of NPLC10x10 chunks) in resistance mode, than NPLC100.

[chuckb]

I had just finished compiling some HP3458A Allan Variance data this afternoon. The Auto Zero function makes a large difference is the voltage reading stability. For my equipment, the stability improvement with AZ on was nearly a factor of 100 better at 10 seconds.

This was a stable temperature environment. I had the 3458A running at 10 NPLC with a copper wire shorting the input terminals. I had AZ on while recording. I tested the 0.1V, 1V and 10V ranges. The noise floor was basically exactly the same.

Then I shut Auto Zero off on the 10V scale and the zero reading was just all over the place.

One of the attached files shows the results from the HP34420A and EM Electronics A10 preamp testing for comparison. The two plots for the A10 preamp are different because they had different temperature stability.

New attachments are under construction.

More details of nanovoltmeter and preamp testing in this discussion -
https://www.eevblog.com/forum/testgear/nanovoltmeters-performance/msg1625434/#msg1625434

[splin]

I had just finished compiling some HP3458A Allan Variance data this afternoon. The Auto Zero function makes a large difference is the voltage reading stability. For my equipment, the stability improvement with AZ on was nearly a factor of 100 better at 10 seconds.
...
Then I shut Auto Zero off on the 10V scale and the zero reading was just all over the place.

That doesn't look right - a factor of over 30x worse without AZ at 1s is much higher than I would expect. This paper investigating the 3458A noise performance:

https://www.researchgate.net/publication/306114861_Keysight_3458A_noise_performance

found AZ increased noise by around sqrt(2), commensurate with the noise added from the zero measurement.

I'm at a loss to explain what might be going on, but I'm sure there are some here who might have some plausible ideas...

[chuckb]

I agree that it does not look correct. I would say one of the plots is off by a factory of 10.
The data collection program used GPIB to put the data directly into a spreadsheet. Then I transferred that data to a text file to feed into Stable32. It's all cut and paste.

I must have switched the DVM to the 100V scale when I turned off Auto Zero. I'll get the plot corrected.
Thanks for pointing the problem out to me.

[Kleinstein]

There is a trade of in using the AZ mode: it increases the higher frequency noise by a factor of sqrt(2)  because use of calculating the difference  (though with some extra averaging on the zero the factor could be slightly less).  However averaging on the zero measurements can interfere with 1/f noise suppression. There was quite some discussion of this for the Keithley meters (e.g. DMM7510 etc.).

On the other side the AZ function also suppresses 1/f noise - so the lower frequency noise is reduced. Depending on the time scale used (time difference in Allan deviation) the noise with AZ function can be lower or higher. For short times the AZ mode could be higher by up to the factor of 1.4, but while the non AZ mode should have the Allan deviation to go up from some 0.1 seconds or so, in the AZ mode the Allan deviation can go down for quite a while, up to 10s or even 100s of seconds. At the long time scale a factor of 30 and more is well possible.

The old data looked odd - the noise (in absolute numbers) should be lower in the 1 V and 0.1 V ranges. So maybe autoscaling was still active. The numbers also would make more sense for the 100 mV range, as they where way too good for the 10 V range.

[chuckb]

The corrected DVM AV stability plots are attached. I checked two HP3458As with similar results. Compared to the original HP3458A, the second meter seems to have half the noise on the lowest two ranges.

The original problem with my analysis was the result of my cutting and pasting segments of noise data out of order. I have automated that in EXCEL now.

Based on all this, I'm leaving the meters in Auto Zero all the time now.

[Kleinstein]

The new data look a lot more plausible.  I am a bit surprised to see so much difference between the two meters. In the 10 V range there is only a small difference in noise performance. However in the 100 mV range meter #2 is considerably better. So it looks like the input amplifier of #1 seems to be considerably more noisy.

[chuckb]

The quiet DVM is the newer meter. It's about 10 years old. The other unit is over 25 years old. The next time I have the covers off I will check the details of who made the parts in the front end.

[David Hess]

I went over the datasheet specifications and the text in the journal and I think what they mean is that the normally discrete automatic zero integration period at the beginning or end of the signal integration period is (forced to be) distributed within the long total aperture time which makes perfect sense for suppressing 1/f noise.  So if you disable automatic zero, you also disable suppression of 1/f noise.

The confusing part of the text is "This also reduces the measurement speed." but I think they mean that it reduces measurement speed just like automatic zero reduces measurement speed.  The specifications say 10 NPLC takes 166ms to produce 6 readings/second without automatic zero and 3 readings/second with automatic zero.  But 100 NPC makes 0.6 readings/second (36/minute) without automatic zero and 0.3 readings/second (18/minute) with automatic zero just as expected so there is no "extra" time involved.  Instead the automatic zero time is distributed through the total measurement time instead of being done in one step which I assume that just means that ten 10 NPLC measurements with or without automatic zero were made and averaged.

There is a subtle difference in the two modes of operation.  At 10 NPLC and faster, the aperture time is continuous.  At 100 NPLC and slower, the aperture time is chopped up taking twice as long to complete which is why the specifications do not even list an aperture time for 100 NPLC and slower.  If it was not for 1/f suppression, this would result in even more 1/f noise than the total signal integration time would otherwise include because the lower frequency cutoff is halved.  Measurements of operational amplifier 0.1 to 10 Hz spot noise face the same problem and it is common to limit the measurement time to 10 seconds maximum adding a second filtering pole at 0.1 Hz; a longer measurement picks up more noise below 0.1 Hz.

This is where the whole 100 NPLC with automatic zero suppresses 1/f noise but also picks up more low frequency noise problem comes from.  It is only spending half as much time integrating the input signal for a given aperture time as 10 NPLC does but it is the aperture time and not the signal integration time which determines the low frequency cutoff.