Low frequency Noise of Zero Drift Amplifiers

[chuckb]

It appears some choppers do not completely eliminate the basic flicker noise.

We started discussing the low frequency performance of Zero Drift Amplifiers in the "Re: DIY low frequency noise meter and some measurement result of voltage references" topic a month ago.

My interest started with some information on the Analog Devices web site about the ADA4522 AutoZero Amplifier.
https://ez.analog.com/amplifiers/operational-amplifiers/f/q-a/102479/ada4522-1-f-noise

This other paper by David Hoyland (2016) documents the flicker noise performance of several other chopper opamps below 0.1Hz. It also has the current noise spectrums.
https://dcc.ligo.org/public/0126/T1600206/001/Opamp%20Noise%20Test%20Results.pdf

Then Echo88 found this additional paper

Maybe this doc is relevant in this case: http://www.sensorsportal.com/HTML/DIGEST/january_2011/P_745.pdf 

The results from the above papers and my testing are summarized below. I have attached the pdf results from my testing of six different amplifiers. Later posts will have more details of the testing and the results. For my testing of the ADA4528-2, ADA4522-1 and the OPA189 I ended up repeating all the tests with different amplifiers. The results of the second round of tests were within 1dB of the first.

So far -
flicker noise free down to 0.0014Hz
ADA4528 (5v)
OPA189 (36V)

Low Frequency Flicker noise corner, less than 0.1Hz
CS3002
ADA4522
OPA188
OPA180
LTC2057
LTC2058

My impression is some amplifier architectures do not have sufficient gain below 0.1Hz to fully correct for the flicker noise. "Well how good do we need to make it? Well, we need to have good performance for this 0.1Hz to 10Hz spec..."

This extra flicker noise is not a concern when buffering a Zener based voltage reference. The Zener noise is typically ten times the op amp noise.

However if you need stable or Very low noise performance below 0.1Hz don't assume just any chopper will remove the semiconductor flicker noise.

[chuckb]

I tested the amplifiers in a shielded test setup with the op-amps amplifying their own noise 10,000x. My first round of testing just used 1,000x gain. This was not enough gain to be above the HP35665A Digital Spectrum Analyzer (DSA) noise floor. With the higher gain the lowest noise op-amp is 4x above the DSA noise floor.
Gathering the complete noise spectrum of each op-amp took 4 different setups and roughly 30 hours. At high frequencies hundreds of sweeps could be averaged in a short time. At the lowest frequencies it took over an hour per sweep. It usually required 20-30 sweeps to reduce the trace noise to an acceptable level. The DSA results were manually transferred to a spreadsheet.

The output DC voltage was also recorded with a HP34420A 7.5 digit Nano-voltmeter. This data was captured and stored on a PC for later analysis. I usually had a 10 hour data record of samples every 3.3 seconds. This data produced a time record and a Modified Allan Deviation plot.

Attached are details of the testing procedure and the summaries for the first three op-amps.

[chuckb]

I have two strong broadcast transmitters within a few miles of my location. I have learned that I will have issues unless I am proactive with EMI minimization.

My son-in-law gave me 50 of these, very nice, 500g, enclosures. I dedicated one to power input filtering, one to the amplifiers under test and one to DC output isolation. Then I bolted them all together. See attached photos.

I used parts that I had on hand and I have not seen any EMI issues so it looks like the design is good enough.

[chuckb]

One last post for the night.

The stability for the air temperature in this part of the lab is usually 23 +-0.2 deg C. Humidity is held to 35-40% RH.

The air temperature directly next to the heavy metal enclosures was monitored with a HP2804 Quartz Thermometer. The temperature data was recorded with 0.001 deg C resolution. The temperature over 7 hours is included in the time record of the attached pdf. This temperature plot starts a little warm because I had just setup and handled the equipment.

I have not directly measured it but I'm sure the internal box temperature has less variation than the air temperature.

[MatteoX]

I've noticed that you have replaced the floppy drive on the HP35665A with what seems to be usb with some LED display. I am curious what did you use. I have recently got one and was thinking of modifying it.

[chuckb]

I've notied you have replaced the floppy drive on the HP35665A with what seems to be usb with some LED display. I am curious what did you use. I have recently got one and was thinking of modifying it.

I used the SFR1M44-U100 Enhanced version 3.5" 1.44MB USB SSD FLOPPY DRIVE EMULATOR GOTEK.
For my analyzer I have to use the high data rate 500kbps and format the USB to 720k (if I remember correctly). The USB needs to be in the drive at boot up and only one 720k drive is available. No matter how big the USB drive is. The DSA reads and writes to the USB and the PC reads and writes to the USB also.

I had to modify the 34 pin ribbon cable to prevent the 5vdc to the drive from shorting to ground. The analyzer was not damaged but the first cable was fried. HP used non standard pinouts to the drives. My HP3589A (same vintage) uses yet another different pinout.

[David Hess]

I am suspicious of the results because a baffle was not used.  This looks a lot like the different results we got with the same part number at different times because the lead frame material changed resulting in very different thermocouple effects with the same assembly.

[Echo88]

Indeed, maybe slight thermal changes within the circuit masquerades as 1/f-noise. Do you have resistors with very low TCR (PTF56 might be cost-effective here) and do you have access to a good heating-controller to stabilize the temp of the setup?
User blackdog posted a few good heater-design-suggestions here somewhere.
Maybe the OP-Amps that appear to have 1/f-noise just have a higher quiescent current, therefore get hotter and lead to more air convection, which results in thermal change? Havent checked the datasheets yet, but might do that tomorrow.

Nonetheless: Thanks for this measurement series. Its impressive how much energy you put in this interesting project. 

[chuckb]

I am suspicious of the results because a baffle was not used.  This looks a lot like the different results we got with the same part number at different times because the lead frame material changed resulting in very different thermocouple effects with the same assembly.

Is there a baffling technique that you would recommend? Is there a particular IC that you think I should retest?

I have not looked at all the chips but most list a maximum of 15nV / deg C for temp co. Of course if they used the wrong material in the lead frame then they could have any temp co. If there was a 2 deg C peak to peak temperature change, then you could have 30nVpp noise from thermocouple effects. This would convert to around 5nV rms at a low frequency. I am seeing less than 1 deg C change over a test series.

The gain setting resistors are 0805 SMD parts with 10ppm / deg C. With the offset voltages I have and the limited temperature change, I have computed less than 1nV of noise from the gain resistors.

Once upon a time I opened an old Zener reference, in a North Hills CC source, and it was full of sand. That's one way to kill convection currents and add thermal mass. However, I would rather not do that here.

[chuckb]

Nonetheless: Thanks for this measurement series. Its impressive how much energy you put in this interesting project. 

I'm retired now so I look for interesting projects to stay busy. Another fellow I worked with retired at the same time and he comes by 20-30 hours a week to help. I would have a lot more unfinished projects if it wasn't for Jim.

[chuckb]

Here are the results for the LTC2057 and the LTC2058. I updated the OPA188 file to add a 5 minute average of the data. This helps when looking for correlations with temperature.

I added a noise density chart with all the op amps. It's cluttered but all the info is in one spot. Also I attached the schematic of the test setup.

There are a few more choppers to test then I will do a good Bipolar and a good JFET.

Sometime later I will update this data with low frequency current noise. The PCB for this has been ordered.

[David Hess]

I am suspicious of the results because a baffle was not used.  This looks a lot like the different results we got with the same part number at different times because the lead frame material changed resulting in very different thermocouple effects with the same assembly.

Is there a baffling technique that you would recommend?

Any minimum volume low profile enclosure over the operational amplifier and its feedback network works.  I usually end up using card stock but fishpaper or a wax paper Dixie cup cut down to size works well also.  The problem is convective air currents.  I used to print out a 1:1 folding diagram with slots and tabs on card stock using an impact printer to make perfect little baffles.

I suspect the major improvement comes from preventing turbulent flow of the air as it rises.

Is there a particular IC that you think I should retest?

Any of the ones which showed high flicker noise.  Pick one and test it again with and without a baffle.

I have not looked at all the chips but most list a maximum of 15nV / deg C for temp co. Of course if they used the wrong material in the lead frame then they could have any temp co.

That specification excludes thermocouple effects and represents the potential performance under ideal conditions.  It does not include the thermocouple junctions between the leads and circuit.

If there was a 2 deg C peak to peak temperature change, then you could have 30nVpp noise from thermocouple effects. This would convert to around 5nV rms at a low frequency. I am seeing less than 1 deg C change over a test series.

External thermocouple effects will always ruin the low frequency noise performance of even a low noise precision bipolar part (0.1uV/C) with its higher albeit low 1/f noise never mind a chopper stabilized part.

The gain setting resistors are 0805 SMD parts with 10ppm / deg C. With the offset voltages I have and the limited temperature change, I have computed less than 1nV of noise from the gain resistors.

Resistor selection is of course important but you probably cannot go wrong with precision metal film or wire wound parts unless they are defective which sometimes happens.  A thermally symmetric layout with phantom parts to equalize the number of junctions in series with both inputs can help.

Once upon a time I opened an old Zener reference, in a North Hills CC source, and it was full of sand. That's one way to kill convection currents and add thermal mass. However, I would rather not do that here.

Sand would sure work and it would dampen vibration.  Potting in gel or epoxy also works.  I am not so sure about immersion in oil but I know that has been done also.  An air baffle is the simple if perhaps lower performance way and allows an easy comparison of performance with and without the baffle.  In production make the baffle conductive for free shielding at the expense of more common mode capacitance.

[chuckb]

Thanks for the guidance. I have some thin, flexible GAP PAD that I will cut and place over the sensitive parts.

[RandallMcRee]

I suggest testing the OPA140. I have used it for low-pass filters and it does seem to have good low-noise properties. The datasheet says

Very Low Offset Drift: 1 μV/°C Maximum
• Very Low Offset: 120 μV
• Low Input Bias Current: 10 pA Maximum
• Very Low 1/f Noise: 250 nVPP, 0.1 Hz to 10 Hz
• Low Noise: 5.1 nV/√Hz,

[RandallMcRee]

Chuck,
I am puzzled when comparing the graph of the ad4522-1 to the opa188. The opa188 seems to have sample measurements less than 1uV but greater than 950e-9 volts? The 4522-1 has samples around 500e-9. Yet both have 0.1 -10Hz noise in the 0.1-0.2uV range according to the datasheets. Can you shed some light on this? (Graphs from post #1 and #3). What is being measured in those graphs?

[chuckb]

The opamps are configured for a gain of 10,000. The output voltage passes through a 230 Hz single pole filter before leaving the enclosures. This DC voltage (around 10mV) is measured by an HP34420A 7.5 digit meter. The meter does not have any filters enabled. It's configured for 100 NPLC and offset compensation AutoZero is on. The meter integrates the input voltage for 1.67 seconds. Then it measures it's internal offset and zeros the meter. The meter supplies a reading every 3.3 seconds. The software takes that reading divides it by 10,000 and stores it in a spreadsheet along with time and temperature.

I don't know how the Multislope Integrator responds to higher frequency noise. This meter is taking a sample every 1.67 seconds and the response probably rolls off quickly above that frequency.

[Kleinstein]

The multi-slope, like other simple integrating converters should give an sin(x)/x type frequency response. So higher frequencies should be reasonably well suppressed, especially the 60 Hz and related as they fall to zeros of the filter.

The absolute position should be just the offset of the OPs - not directly connected to noise. So the 2 OPs would be at around 1 and 0.5 µV offset.

[David Hess]

I don't know how the Multislope Integrator responds to higher frequency noise. This meter is taking a sample every 1.67 seconds and the response probably rolls off quickly above that frequency.

Everything you need to know is here:

It's configured for 100 NPLC and offset compensation is on.

You are in the US so 60Hz power means the integration time is 1.67 seconds.  That produces a sin(x)/x frequency response with a first null at 0.6Hz and a -3dB bandwidth of like 0.29Hz (1) although I am not sure what significance that has with a non-linear roll-off.

Obviously this will not work for measuring noise up to 10Hz or even 1Hz.  When I have done this in the past, I had to settle for higher noise and lower resolution but faster measurements and then amplify before measurement accordingly because I was interested in noise up to 10Hz but I got excellent results.

(1) The 0.35 relationship between transition time and bandwidth becomes 0.442 or 0.468 depending on your perspective.

[chuckb]

I performed a test on all three channels to test for convection air currents interacting with dissimilar metal junctions on the PCB. I tested the LTC2057, LTC2058 and the OPA188. All three had previously shown low frequency flicker noise.

I cut some small sections of Berquist Gap Pad HC 5.0 that is 1mm thick. One thin slice covered the gain setting resistors. Then I placed two move layers over the resistors and the opamp covering as much adjacent space as I could. This material is slightly tacky and very compliant. It’s like a wet blanket. If there was any convection current happening before I estimate it is over 10 times less now.

With this modification the parts were retested. The low frequency noise of the OPA188 was within +- 0.5 dB of the previous open air test (inside the metal box). The LTC parts were monitored with the HP34420A and the peak to peak noise along with the Allan Deviation were within 10% of previous open air test.

With the Gap Pad, the LTC2058 offset shifted from -240nV to -160nV. I don't know why it shifted. Maybe it was from the temperature change. The LTC2057 offset voltage shifted less than 20nV. I did not check the OPA188 for an offset voltage shift.

Conclusion - The low frequency flicker noise that I am measuring is not caused by convection currents in the air interacting with dissimilar metal thermocouple junctions. That type of noise absolutely happens in other installations, it is just not happening here in this small, low power density enclosure.

The ADA4522 was tested in open air with +-2.5v and +-5v power. The noise amplitude and spectrum was the same as the internal power dissipation, and chip temperature rise, doubled. If there were convection currents the noise level would have changed.

Next testing will be for some conventional opamps just for comparison. Then I will build up a new pcb to start current noise measurements.

[chuckb]

I tested three normal amplifiers for low frequency noise. This was to provide a baseline for comparison to the Zero Drift amplifiers.
All three amplifiers were tested at +-5Vdc and a Gain of 10,000. I used the same 10 ohm and 100k ohm feedback resistors to set the gain as I used on previous tests.

The OPA227 (CORRECTED, WAS OPA277) bipolar amplifier was tested. It's low frequency flicker voltage almost reached the performance of the ADA4522 chopper!

I also tested two JFET amplifiers, the ADA4622 and the ADA4625. The ADA4625 has a very good 3nV / rt Hz white noise floor and very good flicker noise performance for a a JFET.


I made a new PCB for current noise testing last week. It came in Friday and Jim built it up for me. The first board was 2 layer with no flood fill of copper. This new PCB board has the same basic layout but I made it 4 layer with grounded copper flooding all the unused spaces. We will see if it makes a noticeable difference when I rerun a few of the voltage noise tests. In actual real world applications the Zero Drift amplifiers may not be in a heavy metal enclosure like I have and they may be subjected to faster temperature changes. A PCB with more copper to equalize the temperature may be useful for an environment with a faster ambient temperature change.

The current noise testing PCB started with a 1Meg ohm resistor for the current sensing resistor. This quickly proved to be too large, the op amp output was on the 15v rail. I changed it to a 100k resistor. That seems to work. I elected to only measure the current noise in the positive input of the opamp. When other people (and manufacturers) test the current noise they test both inputs at once. For me, 90% of my circuits will have a low impedance in the negative input and a higher impedance in the positive lead so that is what I tested. For this test my negative input is connected to the 10 ohm feedback resistor. Does anyone see a problem with this approach?

I have provisions to test the impact of 20pf and 2000pf across the 100k resistor. Reference the attached schematic.

When I first powered up the pcb it worked fine up to +-10V then the negative rail started drawing 1 amp (PS limit) and the voltage fell to 3V. Later I could take it up to 13V before it started drawing 1 amp. Well, when you put 5 solid tantalum capacitors in backwards on the negative rail, they will do that. I need to fix my engineering drawing.

The low frequency current noise testing has started and the results will be provided soon.

[Kleinstein]

I see no problem with testing only the current at the non inverting input.  The current noise can be correlated between the inputs. So for a full test one would need a test with resistors at both inputs too.

Because of the correlation in current noise the test with resistors at both inputs at the same time can not be replaced with tests with the resistors one at a time.  It's kind of measuring different things.

I also prefer the current noise for just 1 input over the date for the uncorrelated part only.

Wether the 100 K resistor is right for the test likely depends on the OP to test - some might need an even lower resistance and others might want the 1 M back. For the low voltage noise types the 100 K could well be sufficient. Low bias AZ OPs (e.g. AD8551) might want a higher resistance.

[chuckb]

Current noise testing of Zero Drift amplifiers continues. I will be traveling so this will not get updated for awhile.

Attached are some low frequency current noise spectrums for the ADA4528, the OPA189, the ADA4522 and the LTC2058. These were all with no decoupling capacitor across the 100k current sensing resistor. The spectrums are about 3 dB better with a decoupling capacitor. Charts to follow when I have time.

This is the current noise of just the non-inverting input. The inverting input has 10 ohms to ground. With the 100k current sense resistor, the current noise spectrum is at least 10dB above the voltage noise spectrum.

I started current noise testing with a 100k resistor as the current shunt. The next step was to add some filter capacitance across the resistor and note the change. The change was larger than I expected. Take the ADA4528 for example. The offset voltage was -1uV. When I added just a 100k resistor the offset raised to -20uV. Well that's the bias current... Then I added a 2nF COG cap across the 100k and the offset dropped to -6.5uV.

My current noise results for the ADA4528 do not match Hoyland's. Hoyland has a flicker noise rise starting at 0.1 Hz. In my testing the rise is very slight at the lowest frequencies. I will have to look into that. Maybe it's because I'm just looking at the current on the non-inverting input.

[EmmanuelFaure]

Nice PCBs! Where do you get them manufactured?

[chuckb]

ExpressPCB has a quick turn small sized 3 board deal for about $100. You use their pcb software and you don't have to mess with gerbers, drill files etc.

[chuckb]

Thanks for the work combining those graphs. Making a combined chart like that is on my to-do list.