Lowest noise OP amps for low frequency, low level AC coupled signals.

[chuckb]

To provide comparison, I made a new plot with a line I called the "0110 Axel line".
It's a spline on log-log diagram, with just three points, preliminary values at
 
100R 4nVrms
100K 10nVrms
1G 500nVrms

The 0110 Axel could be a baseline to relate to and plot results against. Op-amp results below the line would be Breaking News, or errors 

Should the 0110 Axel line be raised by by a factor of 3.16 because you are looking at the noise over a 10Hz bandwidth?

The X axis of the graph is labeled "Input Impedance R1 ohm". Is this assuming a source impedance of 0 ohms?
If the source had a 1k impedance the noise in a 10 Hz bandwidth would be 12nV rms.
If the source had a 100k impedance the noise in a 10 Hz bandwidth would be 120nV rms.

Linear Tech sorted their opamps based on input Z in the attached Design Note. The graph is based on a 1Hz bandwidth. You are looking for noise in a 10Hz BW so your results will be 3.16 higher that the Design note values.
Good luck!

[David Hess]

Added OPA27 which has input bias cancellation and it performs very well here! Interesting, didn't know about this....

OPA1611 was mentioned, I had it in there originally but removed it as it wasn't top notch in _this_ application. However it seems one of the very best in terms of THD. I also saw somewhere that it's virtually identical to OPA211. Anyway, I put the OPA1611 back in th plot.

Beware that early input bias current cancellation circuits caused distortion in audio applications.  As I recall, this applies to the OP-27 and related parts and was one of the things that Linear Technology corrected in their LT1027.  The Burr-Brown OPA27 dates back to at least 1984 so very likely suffers from this but I would need to see the detailed schematic to know.

To get noise data down to 0.1 Hz one might want to take something like 10 times the minimum 10 second interval to average over some random parts like popcorn noise.

Be careful doing this if you want to make comparisons to datasheet specifications for noise.  Limiting the measurement to 10 seconds as is commonly done adds another pole into the frequency response.  Making the measurement over a longer time period results in a different filter shape factor complicating comparisons with other sources.

The better way today is more like an FFT to get the noise spectrum.

This is what I would prefer down to 0.01 or 0.001 Hz which requires a 100 or 1000 second measurement and there is no reason to use AC coupling in the test circuit then.

On the other hand, the standard 0.1 to 10 Hz peak-to-peak or rms noise measurement with the proper filter shape is much easier to produce and more directly comparable to the datasheet specification.  Note that operational amplifiers are not tested in production for this specification except in special cases like the customer paying a premium for it.  Instead, a spot noise measurement at a higher frequency is made which correlates well enough with low frequency noise.

The 0.1Hz lower limit of noise testing seems too high a frequency, somehow...  one gets the sense that this limit was chosen because it was possible to measure with reasonable equipment, rather than what we really need to know.

For example, if you have an instrument that you want to stay stable for hours or days on end, you really need to know the noise performance down to the milliHertz or even microHertz range.

AC coupled applications tend not to extend below 0.1Hz.

I have measured few wet-slug electrolytic caps, same as Jim used in his preamp.
They shown leakage <4 nA at 10V bias. Only problem is the cost.

On a parallell note, any leakage through the input cap would destroy the results. For high impedance inputs, tiny cap values require v-e-r-y low leakage while for the lowest noise amps, low input impedance requires large caps. Which E-lyths are good enough?

Some low leakage high voltage aluminum electrolytic capacitors can achieve similar performance but must be tested for burst noise.

For shunt capacitors in a filter, two capacitors can be used in series with the middle node bootstrapped with 10 times the input capacitance.  This diverts the leakage to the output and minimizes the voltage across the top capacitor lowering its leakage.  This is a common configuration for filters on precision references where excessive leakage would compromise DC accuracy.

In some applications, a DC accurate filter using an AC coupled operation amplifier is used.

I am thinking of setting this up again. would use
- Mechanical relay MUX
- A fA leakage input amp buffer (LMC660?) for the DMM, repeatedly checked vs a reference.
- A temperature controlled box.

The LMC6081 which is the lowest noise and highest precision of that series was the natural choice years ago but there may be some better ones now.  The LMC6001 is a low bias current tested LMC6081 but practically all of the LMC6081s are 10s of femtoamps at worst and testing them is pretty easy with the operational amplifier configured as an integrator to integrate its own input bias current.

[David Hess]

For the precision parts the bias cancellation is kind of a requirement. The extra circuit is not a big deal anymore - so hard to find some.
Even the low noise audio amplifiers (e.g. OPA1611) tend to include bias cancellation as otherwise the bias would be rather large.
Too keep the current noise reasonable it may take super beta transistors, and these usually require a kind of bootstrapped cascode to use them. So not that attractive in discrete form.

Integrated Linear Systems sells a super beta NPN pair but good luck finding it; they have one of the worst web sites ever.

Input bias cancellation is kind of weird.  On one hand, it makes the need for balanced source resistance unnecessary so there is no added noise from the balancing resistance.  On the other hand, correlated input bias current cancellation *increases* noise with unbalance source impedance.  *runs in circles waving arms wildly*

It looks like even in the low frequency range good JFETs are better than BJTs, when it comes to high impedance - so there is likely not that much market for discrete BJT solutions and such BJT pairs. So if discrete it's more like JFETs anyway.

The low noise JFETs have higher input capacitance which adds its own complications.  Still, they are awfully good now.

With lot's of effort a discrete build chopper amplifier may also be an option, as it can use a JFET based amplifier instead of the usually CMOS based in AZ OP chips. So no more need for a high chopper frequency to get low noise, especially if a low BW is OK.

Low noise low drift discrete JFET input stages are still feasible albeit at a high cost.  A more economical option is to chopper stabilize a low noise JFET input stage, discrete or integrated, with a parallel CMOS chopper stabilized amplifier and get the best of both including suppressed 1/f noise and suppressed charge injection.  The only drawback is again high JFET input capacitance although I think this can be suppressed with a cascode.

[SilverSolder]

The 0.1Hz lower limit of noise testing seems too high a frequency, somehow...  one gets the sense that this limit was chosen because it was possible to measure with reasonable equipment, rather than what we really need to know.

For example, if you have an instrument that you want to stay stable for hours or days on end, you really need to know the noise performance down to the milliHertz or even microHertz range.

AC coupled applications tend not to extend below 0.1Hz.

I was thinking DC applications -  and acknowledging that due to 1/f phenomena,  DC is the new AC!

[Kleinstein]

...
Be careful doing this if you want to make comparisons to datasheet specifications for noise.  Limiting the measurement to 10 seconds as is commonly done adds another pole into the frequency response.  Making the measurement over a longer time period results in a different filter shape factor complicating comparisons with other sources.
...
On the other hand, the standard 0.1 to 10 Hz peak-to-peak or rms noise measurement with the proper filter shape is much easier to produce and more directly comparable to the datasheet specification.

The 0.1 Hz to 10 Hz range is kind of used as a standard, but AFAIK there is no real standard on what type of filter to use exactly. Defining an effective noise bandwidth is not accurate if there are different amounts of 1/f noise.  AFIAK the common choice is a 2 nd order HP and LP each. However the AC coupling and similar might add some modifications to the ideal filter curve, especially at the lower end, even if that extra cross over is below 0.1 Hz. 

The 10 seconds window is another option for the low end and would especially well fit modern implementations. Longer data sets could still be looked at with 10 s sections at a time.  Especially for the direct peak to peak reading it would take many 10 second intervals and than some average of the peak to peak values from different intervals. The peak to peak readings show quite some scattering, when just looking at single intervals.

So there might be a value to using the spice models instead of datasheet values that might use slightly differing definitions of the 0.1-10 Hz band.

Another way to show low frequency noise, especially well below 0.1 Hz is using Allan deviation plots. Here the upper end (corresponds to lowest frequencies) is scattering by definition. So the good data only extend to something like 1/5 the measurement time.

[MiDi]

A large part of the LNA noise floor is also contributed to (the AC-part of) leakage currents of the input coupling capacitor.
(if not using foil capacitors with thousands of uF).

How did you measure that?
My LNA did not show any significant difference between input shorted and 33mF@8V ~30nA leakage (10min):

On a parallell note, any leakage through the input cap would destroy the results. For high impedance inputs, tiny cap values require v-e-r-y low leakage while for the lowest noise amps, low input impedance requires large caps. Which E-lyths
are good enough?

I once stared doing measurements on various caps but results where not very clear due to
- leakage through the voltmeter, even it has 1 Gohm input, that was way too much.
- temperature variations on the cap made results very shaky.

I am thinking of setting this up again. would use
- Mechanical relay MUX
- A fA leakage input amp buffer (LMC660?) for the DMM, repeatedly checked vs a reference.
- A temperature controlled box.

Attaching a pic of a good 200uF polypro. It's BIG.

A two stage design rejects the leakage induced offset.
Take a look at the thread diy-low-frenquency-noise-meter.
I measured two Panasonic 2.23.3mF M Typ A <4nA@10V.

I have measured few wet-slug electrolytic caps, same as Jim used in his preamp.
They shown leakage <4 nA at 10V bias. Only problem is the cost.

Wet-slugs might be better at high temperatures, for normal lab environment selected el-caps can get in the same range.
For selecting in the nA range there is no need for special equipment, a dmm with 1-10MOhm input resistance in series with cap will do in conjunction with reasonable linear psu.
It takes a couple of days to soak/settle.

Further details e.g. in Thread measuring-capacitor-leakage

[David Hess]

The 0.1 Hz to 10 Hz range is kind of used as a standard, but AFAIK there is no real standard on what type of filter to use exactly. Defining an effective noise bandwidth is not accurate if there are different amounts of 1/f noise.  AFIAK the common choice is a 2 nd order HP and LP each. However the AC coupling and similar might add some modifications to the ideal filter curve, especially at the lower end, even if that extra cross over is below 0.1 Hz.

That is one reason I really liked Linear Technology.  They were very specific about how they measured 0.1 to 10Hz noise in their datasheets so customers could reproduce their results.  Back when I was messing with LT1128s and LT1150s, I got results exactly in line with what their datasheets said.  It was actually weird when the math, theory, measurements, and datasheet specifications all matched up.  It was even weirder being able to measure a 50 ohm change in source resistance to better than 2% by noise alone.

The 10 seconds window is another option for the low end and would especially well fit modern implementations. Longer data sets could still be looked at with 10 s sections at a time.  Especially for the direct peak to peak reading it would take many 10 second intervals and than some average of the peak to peak values from different intervals. The peak to peak readings show quite some scattering, when just looking at single intervals.

If I was automating noise measurements, I would certainly do that.  My own experience is that if the peak-to-peak measurement over 10 seconds is inconsistent, then something is wrong.  It is worth doing the RMS measurement as well in parallel as a consistency check.

So there might be a value to using the spice models instead of datasheet values that might use slightly differing definitions of the 0.1-10 Hz band.

As pointed out earlier, I would not blindly trust the SPICE model noise results.  I would only use them for tuning after having verified them in the real world.

Another way to show low frequency noise, especially well below 0.1 Hz is using Allan deviation plots. Here the upper end (corresponds to lowest frequencies) is scattering by definition. So the good data only extend to something like 1/5 the measurement time.

But my math already uses integrated noise density and that is what is in the datasheets.

[David Hess]

The 0.1Hz lower limit of noise testing seems too high a frequency, somehow...  one gets the sense that this limit was chosen because it was possible to measure with reasonable equipment, rather than what we really need to know.

For example, if you have an instrument that you want to stay stable for hours or days on end, you really need to know the noise performance down to the milliHertz or even microHertz range.

AC coupled applications tend not to extend below 0.1Hz.

I was thinking DC applications -  and acknowledging that due to 1/f phenomena,  DC is the new AC!

Some operational amplifiers, especially modern fast CMOS devices, have incredibly high 1/f noise but it tends not to be a problem for low frequency precision devices above 0.1Hz.  Chopper stabilized devices dominate below 0.1Hz which is why their informal noise specification goes down to 0.01Hz.  From 0.1 to 10Hz, non-chopper devices are competitive if not better.

Newer chopper stabilized operational amplifiers are better or say they are but I have not used them.

[Andreas]

How did you measure that?
My LNA did not show any significant difference between input shorted and 33mF@8V ~30nA leakage (10min):

Hello,

I have that usually when I did forget to put a bias voltage on the input of the LNA before storing the LNA.
So I have to bias the input capacitor for ~2 days before I can make useful measurements.
Otherwise the noise floor is increased above 200uVpp (low frequent) where it should be around 120uVpp.

at ~8V there are less problems with noise floor. But if I increase to ~13V bias (10*NiMH) I have even large excursions of the noise floor.
But I also have to admit that one of both input capacitors is rated only for 16V. (so somewhat at the limit).

with best regards

Andreas

[janaf]

To provide comparison, I made a new plot with a line I called the "0110 Axel line".
It's a spline on log-log diagram, with just three points, preliminary values at
 
100R 4nVrms
100K 10nVrms
1G 500nVrms

The 0110 Axel could be a baseline to relate to and plot results against. Op-amp results below the line would be Breaking News, or errors 

Should the 0110 Axel line be raised by by a factor of 3.16 because you are looking at the noise over a 10Hz bandwidth?

The X axis of the graph is labeled "Input Impedance R1 ohm". Is this assuming a source impedance of 0 ohms?
If the source had a 1k impedance the noise in a 10 Hz bandwidth would be 12nV rms.
If the source had a 100k impedance the noise in a 10 Hz bandwidth would be 120nV rms.

Linear Tech sorted their opamps based on input Z in the attached Design Note. The graph is based on a 1Hz bandwidth. You are looking for noise in a 10Hz BW so your results will be 3.16 higher that the Design note values.
Good luck!

The results are TOTAL noise 0.1-10Hz, not per sqrtHz. This because the 1/F noise at (low frequency) has nothing like a constant per sqrthz noise, as it usually has at higher frequency. And yes, it's calculated for zero source impedance as that would be application specific.

[MiDi]

 

Hi Andreas,
I did not used my LNA for couple of weeks - or more like months and left it input shorted.
When measuring my LTZ1000 first time, I let the input soak for ~15min on 9V battery and ~15min with 7V output, so first measurement was taken after ~30min - see results in ltz thread.
10V was measured afterwards and before every measurement series I did quick sanity check on noisefloor with input shorted - with matching results to earlier measurements.

Maybe the 0815 25V cap used is pretty good in terms of settling and leakage

[janaf]

To provide comparison, I made a new plot with a line I called the "0110 Axel line".
It's a spline on log-log diagram, with just three points, preliminary values at
 
100R 4nVrms
100K 10nVrms
1G 500nVrms

The 0110 Axel could be a baseline to relate to and plot results against. Op-amp results below the line would be Breaking News, or errors 

The X axis of the graph is labeled "Input Impedance R1 ohm". Is this assuming a source impedance of 0 ohms?

This is because there's no zero on a log scale, I stopped at 1. Where the noise line in these plots becomes horizontal, it's in reality equivalent to what you get if you short the non-inverting input to ground (with zero or 1 ohm). This is also simple to verify by measurement; Set the (non inverting) gain to some very high value, 1000 or more, short the non inverting input to ground and measure the noise on the output. The resistor values should be low, and can be as the output will only be some microvolt. That measured value, divided by the gain, should be equivalent to the datasheet rock bottom noise value for the device.

[Andreas]

Hello,

I fear I have to excange the input caps in my LNA if you can even short the input during storage.

Last time I had forgotten to bias the input capacitor (for several weeks) I had following results:

100nVpp noise floor at 7.6V NiMH
3.8uVpp noise floor at 11V NiMH
so completely unusable for a 10V reference.

I never measure noise floor with shorted input.

with best regards

Andreas

[janaf]

Needs less than 100 ohms input impedance and a 22000 uF input capacitor it maintains around 6nVrms in 0.1-10Hz. There are a few others like the LT1007 that would be better around 1kOhm.

Hmm,

the question for me is: how usefull are source impedances below 1K Ohm. (when building a LNA for noise measurements).
Typical Zeners (1N829) have 10-20 Ohms zener impedance.
The unbuffered LTZ1000 is also in the some Ohms range.

So if you load such a zener by 100 Ohms you already measure up to 20 % lower noise than actual available.

So typical I want to stay at 1KOhms as input impedance which is the domain of LT1007/LT1037/OP27/OP37.
A large part of the LNA noise floor is also contributed to (the AC-part of) leakage currents of the input coupling capacitor.
(if not using foil capacitors with thousands of uF).

with best regards

Andreas

Yes. Measuring on unbuffered zeners with 1K in would  . The graph was made to give a first view of what noise level can be achieved for a wanted input impedance (and thereby also what capacitor size is needed).

For a zener with something like 200nVrms 0.1-10Hz (1.2uV ptp) i'd want something like a factor 5 or 10 lower noise from the analyzer amp, ie 20-40nVrms. Depending on if the DUT is buffered, like the LTZ1000 standard circuit, or un-buffered, one can see what opamp solutions are possible to use. A quick look at the graph will tell that at 20nVrms(0.1-10), the best devices would require some 100k input or less, while if we accept 40nVrms(0.1-10) almost 1Mohm would be usable the best suited opamps... One can easily see on these graphs why they where struggling to measure the 2.5V LT6655, which has around 600nVptp/100nVrms(0.1-10), with some decent safety margin.

[janaf]

Sorry if I wasn't clear. I meant shorting inputs on the DUT, DC coupled inputs to analyzer, not trashing your months aged bias capacitor...

[jmelson]

I made this graph for some of the lowest noise OP amps available for

- Low frequency i.e 0.1-10Hz

About 10 years ago I had an urgent need for a low 1/f noise op-amp for a voltage regulator circuit.  I found the AD706 which had a guaranteed spec on .001Hz noise, individually tested.  It really solved the problem in a circuit that was VERY sensitive to steps in the power supply.  We could VERY easily see the result of tiny, micorvolt-level steps in the supply upsetting the circuit.  Putting the AD706 in made the jumps disappear.

Jon

[Andreas]

Sorry if I wasn't clear. I meant shorting inputs on the DUT, DC coupled inputs to analyzer, not trashing your months aged bias capacitor...

If you mean my answer: this was related to the post of MiDi:

10V was measured afterwards and before every measurement series I did quick sanity check on noisefloor with input shorted - with matching results to earlier measurements.

@MiDi: I was a bit confused that on your scope readings the RMS-noise is in the same ball park as the PP-Noise.
I usually use the "AC-RMS" on my scope to remove the scope offset from the RMS calculation.

with best regards

Andreas

[David Hess]

About 10 years ago I had an urgent need for a low 1/f noise op-amp for a voltage regulator circuit.  I found the AD706 which had a guaranteed spec on .001Hz noise, individually tested.  It really solved the problem in a circuit that was VERY sensitive to steps in the power supply.  We could VERY easily see the result of tiny, micorvolt-level steps in the supply upsetting the circuit.  Putting the AD706 in made the jumps disappear.

A couple of times I have resorted to my own variation of the circuit shown below but using an OP27/LT1007 class of operational amplifier to make low noise regulator.  It is a little weird to have a regulator capable of putting out an amp of current with the noise of a reference and a load regulation of less than 10 microvolts from 0 to 1 amp.

[iMo]

OPA388 (chopper)?

..ultra-low noise, fast-settling, zero-drift, zero-crossover devices that provide rail-to-rail input and output operation. These features and excellent ac performance, combined with only 0.25 µV of offset and 0.005 µV/°C of drift over temperature..

[Andreas]

OPA388 (chopper)?

..ultra-low noise, fast-settling, zero-drift, zero-crossover devices that provide rail-to-rail input and output operation. These features and excellent ac performance, combined with only 0.25 µV of offset and 0.005 µV/°C of drift over temperature..

From the web-page:

Single supply: 2.5 V to 5.5 V

so as most choppers not so interesting for a 10V system.

with best regards

Andreas

[iMo]

OK, >>5V Vcc requirement limits the choices.

PS: I would be "extremely careful" when comparing opamp's "special" params like noise, TC of their offsets/drifts, etc. based on SPICE simulations..

Most (or, better, I would say "all") opamp's models use only a few "behavioral voltage and current sources" (behavioral source is a source where its value is given as a "function of any parameters/variables used in the circuit") inside the model.

They do not use the actual opamp's internal wiring/schematics with the specific models for each transistor/diode/R/C used on the Si chip/die..

[Kleinstein]

The OPA388 is just another 5 V supply chopper. So limited in use if higher voltage is needed. Depending on the circuit is may still be useful.

For the input AC coupling I see mainly 2 options:
1) a foil cap that is available / practical to some 100 µF. This requires an OP that works well with high impedance.
This essentially only leaves good  JFET (e.g. OPA140, ADA4625,..) or chopper based OPs (e.g. OPA189, ADA4522,...). One may be able to use a few in parallel.

2) a relatively large selected electrolytic cap. The input impedance kind of sets an upper limit to the use full capacitance: The larger the capacitor, the longer it takes for settling, as the resistor ground is kind of limited to some 10 K as an lower bound. DA related settling inside the cap can also be a problem - though not much effected by the size. This would alow for BJT based OPs. So far the good candidates are LT1007 (or LT1037) and AD8675. The LT6018, LT1028 or similar would need unrealistic size caps.  I kind of doubt the OPA27 curve shown - the datasheet suggests more noise, more like the LT1007.

In some cases there could be an alternative to AC coupling at the input: a separate DC removal loop could be used. This would also need an OP reading an RC element, but not connected to the input. So there could be a little more freedom to use a lower impedance / very large cap.

Just comparing the spice models can be tricky, as the models may not be accurate.   As slight advantage they get a round possibly different definitions of the 0.1-10 Hz interval and if lucky also handle to correlation of current noise between the inputs. But this about the only advantage.  Especially low frequency noise can be different between samples / batches.

[SilverSolder]

a separate DC removal loop

Every time I start thinking down that path, I get a nagging feeling that a DC servo will introduce its own noise, which may or may not be possible to subtract from the signal we are interested in?

I am increasingly convinced that DC doesn't actually exist...   it is an ideal only, a simplification of the complexities of very low frequency AC!

[Kleinstein]

At least for the circuits I know, the DC servo will introduce some noise. Unless the DC to compensate is rather small the noise can end up essentially as high as the input amplifier and needs an amplifier to work with a rather similar RC combination it tries to replace. If the DC voltage to compensate is only small (e.g. only a few10  mV of the offset of an FET amplifier) the noise due to the DC loop can be rather low. It still may need quite some power to avoid resistor noise.