Is a Solartron 7081 worth the money?

[tggzzz]

AFIK there is a way to get the result in a floating point format with plenty of digits, however only a limited number of values that actually occur. This might be an issue in the 8.5 digit mode as the integration time is not that much longer there and the internal resolution of the ADC should only increase linear in time (if done the simple way).

I DUmp the values using the RS232 interface, and they have 9 digits in the mantissa.

I've repeated my earlier measurements with the 7081 on the 100V range rather than the 10V range. On the front panel display there is one digit less, as you would expect. Dumping the values gives, for example, 9.99972406E+00,  9.99972592E+00,  9.99972406E+00,  9.99972902E+00, 9.99972034E+00. On the graph below it can be seen that the resolution is effectively 1uV, but the datapoints are rarely exactly quantised; again that is what you would expect with "spurious resolution" in the floating point numbers.

The standard deviation is around 10.5uV whereas on the 10V range it was around 1.5uV. This tends to suggest that 1uV of the 1.5uV is due to the 7081, with the remainder down to the Zener source - but I haven't thought about the maths and physical processes in details.

[Kleinstein]

There are 3 main noise sources: the Input amplifier, the ADC itself and the zener Reference.
To check for the noise contribution of the reference, one compares measuring with a shorted input and with an external low noise reference. The difference is due to the references noise.

Changing to the 100 V range inserts an 1:100 divider and put the input amplifier to x 10 mode. The reference contribution stays the same (as it is proportional to the actual input signal), but the noise of the input amplifier is scaled with a factor of 100 and there is additional noise from the 100 K output impedance of the divider. Noise of the ADC itself is scaled with a factor of 10.
So there is not that much to learn from just comparing the 10 V and 100 V range. Usually the noise in the 100 V range has contributions from the ADC, the amplifier and the divider, that can all be important. So this is the most difficult range.
So all we can say from comparing the 100 V and 10 V readings is that the 1.5 V noise in the 10 V range is at most 1 µV from the amplifier and ADC. So the reference noise must be between about 1.1 µV and 1.5 µV (RMS).

To get the noise contributions, usually the 100 mV range with shored input will give mainly noise from the input amplifier. The 10 V range shorted gives mainly the ADC - a small amplifier contribution could be corrected. The 10 V Range with an external voltage gives additional noise from the two references.

[e61_phil]

A possible partial test of the INL is by measuring two voltages and the sum of the two voltages. In the simplest form this is reversing the voltage source, knowing that the sum in 0, but it can be done with other test points too. In principle this should work down to the limits set by thermal EMF (and similar) when switching between the voltage. Noise may require quite a few repeats. Usually this needs a well isolated (e.g. battery powered) low noise source for two test voltages in series. A single source and a resistive divider is the most convenient setup.

If I understand it correct one could build a chain of similiar resistors. 10 times 1k for example. I would use resistors with small tolerances and (more important) low TC (some S102 perhaps). In the next step I would measure the voltage on every resistor in the 1V range. The linearity of the meter shouldn't matter because of the very close values. Even with a 6.5 digit voltmeter I'm able to measure the 1V steps to 0.1ppm of FS (10V).
In the next step I would use the DUT in 10V range an measure across some resistor combination (easiest way 0 to R1, 0 to R2 and so on). In the end I throw all the data in a table. One column contains the sum of the resistor voltages (R1, R1+R2 and so on) and the next column contains the DUT measurements. Afterwards, I apply a linear regression to the data to get rid of offset and gain errors. The deviation from this linear regression should show the INL. Is this correct?

Is a battery powered reference for the resistor chain really that important? Or is it possible to use the floating output of the calibrator (Fluke 5440B)?

[Andreas]

Andreas, I read somewhere you are using a resistor chain for measuring the linearity of your LTC2400 setup. Is it possible to measure linearity down to 0.1ppm FS with such a setup?

Hello,

that depends mostly on noise and short term stability of the setup.
For the LTC2400 I get down to around 1 ppm FS.

With best regards

Andreas

[e61_phil]

Hello,

that depends mostly on noise and short term stability of the setup.
For the LTC2400 I get down to around 1 ppm FS.

With best regards

Andreas

Hi,

I would like to measure 1:10 ratios within 1ppm. Therefore, I think 0,1ppm INL of FS is needed.

I have an old ESI SR1010-10 transfer standard. Unfortunately I have not the shorting bars. Therefore, it is a bit useless. My idea is to salvage the very nice 10R resistors for other things and use the housing to install some other resistors. I have some russian hermetic sealed 1k14 specified with 5ppm/K but they seem to be better. I think 1k is a good value. High enough to have an acceptable current at 10V and low enough to avoid loading errors. Or is that a bad idea?

Is there anything I should have in mind before soldering in a SR1010? Any toxic solder or something like that?

Best
Philipp

[Kleinstein]

....
If I understand it correct one could build a chain of similiar resistors. 10 times 1k for example. I would use resistors with small tolerances and (more important) low TC (some S102 perhaps). In the next step I would measure the voltage on every resistor in the 1V range. The linearity of the meter shouldn't matter because of the very close values. Even with a 6.5 digit voltmeter I'm able to measure the 1V steps to 0.1ppm of FS (10V).
In the next step I would use the DUT in 10V range an measure across some resistor combination (easiest way 0 to R1, 0 to R2 and so on). In the end I throw all the data in a table. One column contains the sum of the resistor voltages (R1, R1+R2 and so on) and the next column contains the DUT measurements. Afterwards, I apply a linear regression to the data to get rid of offset and gain errors. The deviation from this linear regression should show the INL. Is this correct?

Is a battery powered reference for the resistor chain really that important? Or is it possible to use the floating output of the calibrator (Fluke 5440B)?

The way described here is different from what I thought, but it might be a very viable way offering quite a few test points in a single run. If the measurement is reasonable fast and repeated, there is no need for very good stability of the resistors and stability of the references is less critical too.

Manually changing cables could be rather slow or introduce extra thermal EMF. Something like a board with lots of places for two jumpers might be an intermediate level solution - still manual switch over, but simple and low thermal EMF and reasonable fast to operate.  With a set of 8 resistors something like 30 measurements for one run. This won't be practical in the 8 Digit mode of the Solartron, but could be in the 7 digit mode with external averaging (to reduce the time lost during switching).
 
If the floating output of the calibrator is good enough, depends on the degree of isolation the source and the meter offer. It might be well good enough, especially if the resistor values are not that high (e.g. 1 K each).

[e61_phil]

I can imagine, one could use a HP 3488A togehter with two multiplexer modules (to control HI and LO separately). With such a setup it is possible to switch HI and LO independent from each other to every node of the resistor chain. If the calibrator is also controlled via GPIB it should be possible to cancel out most of the thermal EMF errors. But first I will try it manually.

[guenthert]

I still don't quite grok how you get those 8.5 digits (or where they are coming from).  Reading my HP 34401A while measuring an AD588 reference using default configuration (10NPLC) the last two digits are always 0 (hence 7.5 digits, not necessarily 20 million discrete values):
+1.00014400E+01,+1.00014500E+01,+1.00014400E+01,+1.00014500E+01,+1.00014500E+01,+1.00014500E+01,+1.00014400E+01,+1.00014400E+01,+1.00014500E+01,+1.00014400E+01,+1.00014400E+01,+1.00014500E+01,+1.00014500E+01,+1.00014500E+01,+1.00014400E+01,+1.00014400E+01,+1.00014400E+01,+1.00014500E+01,+1.00014400E+01,+1.00014500E+01,+1.00014400E+01,+1.00014500E+01,+1.00014400E+01,+1.00014400E+01,+1.00014400E+01

For comparison, I measured the same reference at the same time (*) using a Datron-Wavetek 1271:
HP34401A 10 NPLC mean of 25 rdgs: 10.0014444 variance: 2.46399999981e-11
DW1271 fast mode mean of 25 rdgs: 10.001369164 variance: 2.50304000168e-13

Using NPLC of 100, still the last two digits are 0 [edit: this might have been due to an error in programming the DUT]:
 +1.00013300E+01,+1.00013200E+01,+1.00013300E+01,+1.00013200E+01,+1.00013400E+01,+1.00013400E+01,+1.00013400E+01,+1.00013300E+01,+1.00013300E+01,+1.00013400E+01,+1.00013300E+01,+1.00013200E+01,+1.00013400E+01,+1.00013300E+01,+1.00013200E+01,+1.00013500E+01,+1.00013300E+01,+1.00013400E+01,+1.00013300E+01,+1.00013300E+01,+1.00013300E+01,+1.00013300E+01,+1.00013300E+01,+1.00013400E+01,+1.00013300E+01

HP34401A 100NPLC mean of 25 rdgs: 10.001332 variance: 5.60000000063e-11
DW1271 normal mode mean of 25 rdgs: 10.001369256 variance: 1.54464000023e-13

*) Even in 'fast' mode, the 1271 is much slower than the 34401A.  The test series only started at about the same time (block read to memory on both).

[e61_phil]

I still don't quite grok how you get those 8.5 digits (or where they are coming from). 

You will only get 8.5Digits with 100NPLC and if the voltage is below 10V. If your voltage is above 10V you can add an offset:

HP34401A.write("CALCulate:NULL:OFFSet 1")

I think (not know) the 34401A ADC is only capable of 10NPLC with 7.5 digits. If you switch the unit to 100NPLC it will take 10 times a measurement with 10NPLC (like the 3456A) and average them. This leads to an extra digit. I think this digit is cut above 10V due to the internal math capabilities. Further, I think these internal math capabilities are the reason for some missing codes at 8.5digit.

I made a histogram with 10NPLC with 100nV spacing like in the histogram before (see attachment). I can't see any missing code at 7.5digits. I think you will get real 7.5 digits from the 34401A and it is possible to average 10 of these measurements to increase the resolution. I did another measurement with 10x 10NPLC averaged (averaging is done on my pc) and there is no missing code in 8.5 digits (that's clear )


The difference in the standard deviation between the 34401A and the 1271 is perhaps due to the better reference (and ADC)?

[splin]

If I understand it correct one could build a chain of similiar resistors. 10 times 1k for example. I would use resistors with small tolerances and (more important) low TC (some S102 perhaps). In the next step I would measure the voltage on every resistor in the 1V range. The linearity of the meter shouldn't matter because of the very close values. Even with a 6.5 digit voltmeter I'm able to measure the 1V steps to 0.1ppm of FS (10V).
In the next step I would use the DUT in 10V range an measure across some resistor combination (easiest way 0 to R1, 0 to R2 and so on). In the end I throw all the data in a table. One column contains the sum of the resistor voltages (R1, R1+R2 and so on) and the next column contains the DUT measurements. Afterwards, I apply a linear regression to the data to get rid of offset and gain errors. The deviation from this linear regression should show the INL. Is this correct?

This is what HP had to say about linearity testing of the 3458A, from page 14 of the April 1989 HP Journal:

LinearityLinearity
High-resolution linearity was one of the major challenges of the ADC design. The autocalibration technique requires an integral linearity of 0.1 ppm and an differential linearity of 0.02 ppm. One of the more significant problems was verifying the integral linearity. The most linear commercially available device we could find was a Kelvin-Varley divider, and its best specification was 0.1 ppm of input.
Fig. 11 compares this with the ADC's requirements, showing that it is not adequate.

Using low-thermal-EMF switches, any even-ordered deviations from an ideal straight line can be detected by doing a turnover test. A turnover test consists of three steps: (1) measure and remove any offset, (2) measure a voltage, and (3) switch the polarity of the voltage (i.e., turn the voltage over) and remeasure it. Any even-order errors will produce a difference in the magnitude of the two nonzero voltages measured. Measurements of this type can be made to within 0.01 ppm of a 10V signal.

This left us with only the odd order errors to detect. Fortunately, the U.S. National Bureau of Standards had developed a Josephson junction array capable of generating voltages from -10V to + 10V

I'm not sure that's very helpful though unless you know a lot about the relative levels of odd and even order non-linearities in the meter.

Since you want to do 10:1 transfers, then one idea I had is to connect 20 reasonably well matched resistors in a ring with pins at every connection. Connect one point to +Vref and the opposite connection to ground such that you have two parallel strings of 10 resistors. Measure the voltage at the top of the two grounded resistors.

Now move the Vref and ground connections one step around the ring and repeat until you have 40 measurements of 1/10 of Vref; averaging them should give a pretty good result - both by averaging out the resistor value mismatches and by reversing many of the thermal EMFs arising at the connections which may all have small temperature variations. Each resistor is connected +/- when 'travelling up' the divider chain, then reverses when coming down the other side.

As to how close the resistors need to match, and how well all the EMFs can be cancelled to achieve .1ppm is left as an excercise for the reader 

Many other arrangements are possible including using 40 resistors, or re-ordering/reversing sections of the ring and repeating. The ring should allow a complete set of tests, rotating every resistor into the low and high sides, to be done quickly enough to minimise thermal EMF problems due to temperature changes.

To that end, the 20 resistor connections could all be brought to an in-line 2x22 way header. A 2x3 pin socket would connect Vref, ground and the two adjacent resistors to the meter (via a switch) The socket would be advanced one pin at a time along the header to make all 40 readings quickly and easily. Rotary or electronic switches could also be used. Thermal EMFs arising at each pin header/switch contact might not all cancel though, so it would be good to ensure that are all kept at the same temperature as much as possible.

You can of course take readings at other points to measure linearity at any or all 1/10 intervals.

Having establised the linearity at exactly 10% full scale, you can then use a low level triangle/sawtooth/sine wave with a 10% FS offset to measure the differential linearity of the steps near to the 10% point and thus calculate the INL over that range. The linearity of the waveform has to be better than your required accuracy, hence the necessity for a small amplitude signal. It also has to be small to allow the total measurent to be done in a reasonable time.

I hope that made sense. Do let us know how you get on.

[guenthert]

I still don't quite grok how you get those 8.5 digits (or where they are coming from). 

You will only get 8.5Digits with 100NPLC and if the voltage is below 10V. If your voltage is above 10V you can add an offset:

HP34401A.write("CALCulate:NULL:OFFSet 1")

I tried that (together with "CALC:FUNC NULL; STATE ON") which set the offset, but doesn't give me any more digits:
--8<--
+9.00132200E+00,+9.00133400E+00,+9.00134300E+00,+9.00132600E+00,+9.00134500E+00,+9.00133400E+00,+9.00132600E+00,+9.00133400E+00,+9.00133600E+00,+9.00133600E+00,+9.00135100E+00,+9.00133100E+00,+9.00132500E+00,+9.00133400E+00,+9.00134500E+00,+9.00133100E+00,+9.00133500E+00,+9.00135300E+00,+9.00133800E+00,+9.00133900E+00,+9.00133000E+00,+9.00133400E+00,+9.00134000E+00,+9.00133000E+00,+9.00132600E+00
-->8--
Perhaps there is a difference in firmware?  Mine reports itself as
--8<--
HEWLETT-PACKARD,34401A,0,5-1-1

1991.0
-->8--
(*IDN? followed by SYSTEM:VERSION)

[guenthert]

The difference in the standard deviation between the 34401A and the 1271 is perhaps due to the better reference (and ADC)?

Heaven knows.  I don't think they are easily comparable (one being a fine bench meter, the other meant for ATE with left-over parts of the metrology-grade 1281).

The 1271 is actually not sooo much slower in normal mode using 8.5 digits (~10s for one rdg) than the 34401A in 100 NPLC, nor in fast-mode using 7.5 digits, than the 34401A using 10 NPLC.  It (at least mine) is still more precise then:
--8<--
HP34401A 10NPLC avg. of 25 rdgs: 9.0013308 8.22720000002e-10
DW1271 'fast' mode, 7.5 digits avg. of 25 rdgs: 10.00137676 var: 2.34240000072e-12
-->8--

Given that both instruments are more than 25 years old, here's a big thank you to the engineers and other folks who made them! 

[e61_phil]

I tried that (together with "CALC:FUNC NULL; STATE ON") which set the offset, but doesn't give me any more digits:
--8<--
+9.00132200E+00,+9.00133400E+00,+9.00134300E+00,+9.00132600E+00,+9.00134500E+00,+9.00133400E+00,+9.00132600E+00,+9.00133400E+00,+9.00133600E+00,+9.00133600E+00,+9.00135100E+00,+9.00133100E+00,+9.00132500E+00,+9.00133400E+00,+9.00134500E+00,+9.00133100E+00,+9.00133500E+00,+9.00135300E+00,+9.00133800E+00,+9.00133900E+00,+9.00133000E+00,+9.00133400E+00,+9.00134000E+00,+9.00133000E+00,+9.00132600E+00
-->8--
Perhaps there is a difference in firmware?  Mine reports itself as
--8<--
HEWLETT-PACKARD,34401A,0,5-1-1

1991.0
-->8--
(*IDN? followed by SYSTEM:VERSION)

I have two 34401A. One with Agilent Brand (but still the old buttons) and the other one is HP branded.

I run this script to compare with your readings:

Code: [Select]

print("======== Agilent 34401A ========")
A34401A.write("*RST")
A34401A.write("SENSe:VOLTage:DC:RANGe 10")
A34401A.write("SENSe:VOLTage:DC:NPLC 100")
A34401A.write("INPut:IMPedance:AUTO ON")
print( A34401A.query("*IDN?") )
print( A34401A.query("SYSTEM:VERSION?") )
print( A34401A.query("READ?") )
print( A34401A.query("READ?") )
print( A34401A.query("READ?") )

print("======== HP 34401A ========")
HP34401A.write("*RST")
HP34401A.write("SENSe:VOLTage:DC:RANGe 10")
HP34401A.write("SENSe:VOLTage:DC:NPLC 100")
HP34401A.write("INPut:IMPedance:AUTO ON")
print( HP34401A.query("*IDN?") )
print( HP34401A.query("SYSTEM:VERSION?") )
print( HP34401A.query("READ?") )
print( HP34401A.query("READ?") )
print( HP34401A.query("READ?") )

The result is:

Code: [Select]

======== Agilent 34401A ========
HEWLETT-PACKARD,34401A,0,10-5-2
1991.0
+8.99989590E+00
+8.99989670E+00
+8.99989640E+00

======== HP 34401A ========
HEWLETT-PACKARD,34401A,0,10-5-2
1991.0
+9.00005570E+00
+9.00005600E+00
+9.00005650E+00

Both units show 8.5 digits (but it seems to be the same firmware)

[guenthert]

Hey, thanks for your patience there with me.  I tried that (plus the 1V offset) and it still yields only 7.5digits on my HP 34401A.  As if there were different firmware versions claiming to be 1991.0

+9.00141400E+00
+9.00141300E+00
+9.00141500E+00
+9.00141500E+00
+9.00141400E+00
+9.00141400E+00
+9.00141500E+00
+9.00141600E+00

Ah, hmmh, GPIB programming ... I just noticed that the order of the statements matters, e.g. "CONF:VOLT:DC" will reset any math operation and "SENSE:VOLT:RESOLUTION" will implicitly set NPLC.

I can now also confirm that the HP 34401A doesn't use fixed-point or BCD arithmetic (as it arguably should), but some sort of float.  Specifying a low resolution as a test, I found:
--8<--
+9.00089900E+00
+9.00190000E+00
+9.00140000E+00
+9.00089900E+00
+9.00140000E+00
+9.00140000E+00
+9.00140000E+00
+9.00250000E+00
-->8--

And 10uV in the 10V range seems to be the best it can do, "SENSE:VOLT:DC:RESOLUTION 1e-6" yields '532,"Cannot achieve requested resolution"'

Earlier there was an issue with setting integration time.  Now I seem to have got to work and I get:
+1.00014240E+01,+1.00014230E+01,+1.00014220E+01,+1.00014220E+01,+1.00014240E+01,+1.00014230E+01,+1.00014230E+01,+1.00014220E+01,+1.00014230E+01,+1.00014220E+01,+1.00014230E+01,+1.00014230E+01,+1.00014220E+01,+1.00014230E+01,+1.00014220E+01,+1.00014230E+01,+1.00014230E+01,+1.00014230E+01,+1.00014230E+01,+1.00014230E+01,+1.00014230E+01,+1.00014240E+01,+1.00014230E+01,+1.00014240E+01,+1.00014240E+01

HP34401A 100NPLC, avg. of 25: 10.00142296 var: 4.38400000162e-13
DW1271 'normal' mode, 8.5digits, avg. of 25: 10.001364592 var: 2.51136000008e-13

So, some progress 

[e61_phil]

@guenthert: Thanks for sharing the data. It is very intersting to see almost the same variance on the 34401A and the 1271.

In the meantime I tried some linearity measurements. Unfortunately, my only resistor chain is a SR1010-10 with 10Ohm/step. Not quite the best values for such a test, but I gave it a try.
I used my Knick JS 3010 calibrator to source 100mA through 10 of the SR1010 resistors (not very nice due to selfheating and so on). After half an hour of settling time I measured the voltages on the 10 resistors 3 times with my Agilent 34401A. The second and third run were very well within 1ppm (avg below 0,5ppm).

After that I measured the sum of the resistor voltages in the 10V Range of the Agilent 34401A and also with my HP34401A.

The picture 34401A_linearity.png shows a measurement against the 3458A at work.
The picture 34401A_lin2.png shows the result of my little experiment.

I don't care about the sign, because I'm not sure in which way I calculated the deviation with the 3458A. It looks quite good I think.

I also made a spot check and compared (offset compensated) the ratio between 1V and 10V of the SR1010 chain with the measured 1V and 10V ratios:

Agilent 34401A 1,14ppm deviation from the SR1010
HP 34401A 3,17ppm deviation from the SR1010

So the Agilent is quite close to the 0,1ppm FS. In the next days I will test the Solartron 7081 (wasn't warmed up today and I'm running out of time).