DMM Noise comparison testing project

[rigrunner]

Converting to millivolts is easy enough. You want the data like this?

Elapsed Seconds,Millivolt,Raw
1,0.0000, 0.0000000E+00
2,0.0000, 0.0000000E00
3,0.0000, 0.0000000E+00
4,0.0010, 1.0000000E-06
5,0.0000, 0.0000000E+00
6,0.0000, 0.0000000E+00
7,0.0000, 0.0000000E+00
8,-0.0010,-1.0000000E-06
9,0.0010, 1.0000000E-06
10,0.0000, 0.0000000E+00
11,0.0000, 0.0000000E+00

[3roomlab]

errm not the 1 and 0 log, i think it can only be useful reference for DMM-ers if it show the rest of the digits. if you do a STDEV calculation on the 1 and 0, it is not going to be very useful. i think you yourself need a way to know your own STDEV too? no? i hope some HP veteran can tell you which SCPI command.

on my K2015, in 100mV, i estimate the resolution is probably down to 0.5nV. so i recall/read/fetch all the reading to 1234nV, in your case, you should probably try to read the full 0.00000123456V. so you are missing the "23456". you really need those numbers if you want to help yourself to some STDEV math,esp on a spreadsheet. but i think some HP users do the entire math inside the MCU. i have no idea what code they use

[rigrunner]

I can't get any better resolution in the 3V range.

The 30mV range logged this kind of data:

-450.00000E-09
-420.00000E-09
 540.00000E-09
 240.00000E-09
-340.00000E-09
 130.00000E-09
-210.00000E-09
-190.00000E-09
-40.000000E-09
-60.000000E-09
-160.00000E-09
 0.0000000E+00
 90.000000E-09
-210.00000E-09
-210.00000E-09

[3roomlab]

yea thats the right data  .
if you want to make it easy to read, just multiply by 1000,000,000. so you read it directly as -450nV, -420nV etc as integers. but otherwise i think this data is good for TiN's collection
if you want to DIY your own STDEV in spreadsheet, have a read in opencalc webpage on STDEV, AVERAGE, for peak-peak, read use of MIN(range), MAX(range)

[rigrunner]

I uploaded the data for the 3V, 300mV and 30mV data a while ago. 

30mV log gives :

stdev   232.8749695261nV
max   860nV
min   -960nV
avg   -30.9714889124nV

[TiN]

Got around and made graphs of dr.diesel's K7510 vs mine K2002 with few opamps vs sick evil 3458A..

No magic, folks. While Keithley 2040 (erm, 7510 ) is bit better on low-range, difference on higher ranges vanishes at instant..





At base range, they all very close:







3458A with it's averaging on NPLC > 10 still kills them all with ease.

Neeeext....

Thanks dr.diesel for stepping up and actually running test for us. Perhaps you can share your script, so I can add into article for other busy 7510 owners, so we can get more cross-relation.

P.S. Let's give little guy some credit too, 7510 at 4K$ price point being 7.5D meter still good instrument, comparing to 2002 (6K$, 8.5D) and 3458A (10K$, 8.5D).

[dr.diesel]

3458A with it's averaging on NPLC > 10 still kills them all with ease.

Thanks TiN!

We could simulate the NPLC > 10 and continue the fun on both the 7510 and 2002. 

Perhaps you can share your script

My program was integrated with another logging suit specifically written for my equipment and needs, so sharing won't help anyone.   

P.S. Let's give little guy some credit too

I have no buyers pride, but I think the 7510 has much value at $3700.  At the 20v range and below where many/most of us will be using it it really shines, and can compete with the mighty 2002/3458A.  Add in the beautiful screen and other included options, it's win win IMHO.  And I still bet a K8510 is in the works!

I may also redo some of the sets in the 7510 sweet spot and include NPLC 5:

[TiN]

2002 have native NPLC up to 50. Not that much use of it though.....

As of 8510, I'm sceptical. But would be happy to be wrong.

7510 has much value at $3700.

It does indeed. As many already stated, and I agree, 3458A being more a metrology tool, rather than general-purpose bench DMM.

[Dr. Frank]

Hi,

I completed the noise diagram for my HP3458A, from the fastet rate at 4 1/2 digits, NPLC 0.00007 = APER 1.4µs, to NPLC 1000.
Everything is done at maximum speed, e.g. 100kHz @ 1.4µs, and with AZERO OFF, apart from NPLC 1000, which uses AZERO, due to the long sampling time.
I omitted again 100V and 1kV as these should be practically identical to 1V and 10V, respectively.
 
The measurements are done at 50Hz line frequency, so compared to 60Hz , my curves look a little bit better at the same NPLC number, and NPLC >= 1.

At NPLC 1000, the 10V range nearly reaches 9 digits.
 
Frank

[3roomlab]

hi dr frank, how do you use the APER parameter? or rather, how should 1 use APER? while i know its a measurement window, but thats about it

[TiN]

You can set it just like NPLC, using APER command instead. Maximum integration time is 1 sec.

When I get my meter working again, I'll test on different mains frequency, from 45Hz to 65Hz

[3roomlab]

what i dont understand is, if we command APER to capture a window of x samples, it does not depend on cycles like NPLC ?

edit** if based on Dr Franks plot, when we set NPLC 1, i am guessing we should also set APER = 0.02, yes ?
if so, what if we set NPLC =1, but we set APER as something else, 0.5? or its going to introduce more noise?

[Dr. Frank]

what i dont understand is, if we command APER to capture a window of x samples, it does not depend on cycles like NPLC ?

edit** if based on Dr Franks plot, when we set NPLC 1, i am guessing we should also set APER = 0.02, yes ?
if so, what if we set NPLC =1, but we set APER as something else, 0.5? or its going to introduce more noise?

APER and NPLC are equivalent commands. You can either specify the one or the other, following the fixed relation APER * Line_Freq = NPLC.
Therefore you cannot set APER and NPLC  independently. Setting one variable changes the other one automatically.

The APER parameter equals the A/D integration time. Therefore, everything below NPLC = 1 or APER 0.02 does not suppress line frequency, giving higher noise.
Maximum effective integration time is 200ms, or NPLC 10 = APER 0.2. All longer settings are done by averaging multiple NPLC 10s.

It is also important, that the sampling rate is always slower than the setting by NPLC or APER, due to some dead time between samples, plus time for other measurement phases like AZERO, or OCOMP with OHM measurements.

I configured for fastest possible acquisition by PRESET DIG and TIMER event triggering.

APER TIMER Rate
1.4µs  10µs  100kHz
10µs   20µs   50kHz
20µs   30µs   33kHz
200µs 250µs  4kHz
2ms    2.5ms 400Hz

As fast data transfer directly over GPIB is tricky, I instead used the internal 148kB memory, using SINT, DINT format, and downloaded data after complete acquisition.

For NPLC 1000, I had to use AZERO ON, as otherwise the 3458A already drifted a lot, giving erroneous / instable readings, as probably can be seen in TiNs measurements.
As one sample takes 40 seconds, I limited this measurement to 15 samples / 10 minutes.

Frank

[3roomlab]

thanks TiN n Frank. that tip on AZERO also confirmed my assumption about its use. although in my case, it seem to add a tad of noise.

i tried experimenting with cycling AZERO (ie:1 AZERO per 100samples), but it seem to add some non-linearity to long term reading as well. so in my conclusion, i think i will just leave it 100% on. i got this cycling idea from reading the web, but i think it is only usable for certain machines

as i am using RS232, 1 sample of 100NPLC takes approx 6.6seconds to log to PC. NPLC1000 will take about 70 seconds (3600 samples = 7days/70days ! shiet ! )

non linearity is the wrong word

[Dr. Frank]

thanks TiN n Frank. that tip on AZERO also confirmed my assumption about its use. although in my case, it seem to add a tad of noise.

i tried experimenting with cycling AZERO (ie:1 AZERO per 100samples), but it seem to add some non-linearity to long term reading as well. so in my conclusion, i think i will just leave it 100% on. i got this cycling idea from reading the web, but i think it is only usable for certain machines

as i am using RS232, 1 sample of 100NPLC takes approx 6.6seconds to log to PC. NPLC1000 will take about 70 seconds (3600 samples = 7days/70days ! shiet ! )

You may use AZERO ONCE to let precede one zero reading before the complete acquisition.
For longer sampling, like NPLC 1000, your 3458A will drift too much, anyhow.
That's not non-linearity, that's simply an offset, what you'll see.

It makes no sense to take so many samples for long integration times, as the standard deviation (per sample) is already extremely low.
You would erroneously measure mid-to-long term drifts, like LTZ and gain resistor drifts, instead of noise.

For these short apertures, I acquired 32768  or 65536 samples within 1 second, which makes sense, to get stable stdev values.

Frank


 

[3roomlab]

65k samples a sec. wow i dont think the K2015 is possible to do it that way. or maybe i dont know how
but in any case, it has been very interesting journey playing with the K2015 THD (nope i dont have a 3458a, thats TiN)

which means, with long windows of captures, i have included all kinds of environmental noise into my data

[plesa]

Keithley 2000/2001/2002 can do 2000 samples per second. I expect same spec for 2015.
HP 3458A can make 100k, and for instance 34411A 50k samples per second.

[3roomlab]

At NPLC 1000, the 10V range nearly reaches 9 digits.
 
Frank

btw, how do we gauge usability of a last digit. is it using noise STDEV? say 10v range 12,345,678.9uV, what should STDEV look like so that the 9th is considered a usable reading?

and also what is the diff between RMS noise and STDEV?

[Dr. Frank]

btw, how do we gauge usability of a last digit. is it using noise STDEV? say 10v range 12,345,678.9uV, what should STDEV look like so that the 9th is considered a usable reading?

and also what is the diff between RMS noise and STDEV?

RMS noise equals STDEV. You can recognize that directly from comparing the STDEV and the RMS formulae.
For the 3458A, averaging gives converging results, obviously, whereas the FLUKE 7510 may be worse at higher averages (it diverges).
I am not able to identify the reason for that, but it has something to do with the character of the internal noise sources.

Provided that the mid- to long-term stability (determined by variations of temperature, reference voltage and  gain resistors over the measurement interval) are small enough, you may theoretically achieve 9 usable digits when averaging sufficiently.

The variance (or uncertainty) of a set of N measurements is given by stdv/ sqrt(N).

For the 10V range, you may average over about 50 measurements of NPLC 100 to have 1e-9 variance of the result.
4 averaged measurements of NPLC 1000 also give 1e-9.

I doubt that anything below 1e-8 is useful.
At first, the mentioned environmental parameters are worse, and second, the linearity of the A/D is 2e-8 'only'.

Therefore, any higher resolution is useless.

For classic analog references and DMM, an instability of 1e-8 is the absolute limit.

Below that, you need totally different working principles.

Together with our National Standards Office, PTB, a company named Supracon developed an AC  quantum voltmeter, having about 1e-10 stability, see here:
http://www.supracon.com/en/ac_quantum_voltmeter.html

Frank

[TiN]

That's if DC offset on RMS is removed. STDev removes DC component by definition.

Also on 3458A APER does actually allow integration times longer than 10 NPLC, up to 1 second, while higher NPLC are averaged in digital domain. At least that's what in manual.

All my measurement are with autozero, unless clearly stated otherwise, so issue visible on my data is due other factors, not due autozero function. One of those was mentioned earlier in repair thread, reference output on A9 jumps ~1ppm in random matter, upsetting samples on long speeds. That is likely due to small airflow leaking into LTZ1000 pins.

I did run tests before with NPLC vs APER speeds, results are very close. Will dig data out and post later for clarity.

Have to get ADC fixed first before redoing any good measurements.

[Kleinstein]

The autozero has two functions: the obvious is to counteract drift. The second is that with autozero the 1/f noise is reduced, just like in chopper amplifiers. So when 1/f noise takes over it is better to use autozero and digital averaging than relying on A/D conversions with long integration times.

To be fair in comparision one should also note the minimum time between readings. For the user this is the more important factor than the actual time used for the integration phase. The fraction of time actually used for measurement is also important, is this sets the noise bandwidth for noise from the signal source. So the extra time spend in autozero and similar phases is bad in both ways.

[3roomlab]

i cant remember the ic pin of this trace (iirc its from U153), i caught this 2 pics while mucking around previously, they are while NPLC 1 /NPLC 10 are running. im not sure if each pulse represent the capturing of 1 sample but i assume it is, and it is pretty slow, nearly 500ms for NPLC10, and for some strange reason, NPLC10 at 100mV is faster than NPLC10 at 10v

[Dr. Frank]

Also on 3458A APER does actually allow integration times longer than 10 NPLC, up to 1 second, while higher NPLC are averaged in digital domain. At least that's what in manual.

TiN, you're fully correct, it's explained in the notes on pages 59/60 in the manual, and I just checked that on the OCOMP ohm function, where you can directly see the difference in the timing, when digitizing the voltage across the resistor.

That I did not know before, but it's pretty interesting for some reason, although maximum A/D conversion is limited to 1sec only.

Frank

[TiN]

Here's data, just to back that discussion up:

3458A 10 and 1 NPLCs, RES 2W, OCOMP ON



3458A 0.6 and 1 sec APERs, RES 2W, OCOMP ON. Note 400ms timescale, instead of 100ms



Keithley 2001 10 NPLC, AZER,LSYNC OFF and AZER,LSYNC ON, RES 2W, OCOMP ON

[Kleinstein]

So in 10 NPLC mode the HP3458 need a little over 20 cycles and the Keithly2001 about 50 cycles. Thats quite a difference not to call it cheating, with even the shorter phase 20 cycles long.

Is there an extra setting that set the output rate of the Keitly to a fixed 1 Hz ?