using audio ADC for instrumentation use and probing noise floor of amplifiers

[loop123]

I have 2 questions.

1. What is the negative effects of using an audio ADC specifically the E1DA Cosmos ADC in sampling voltages for instrumentation use? I know it blocks low frequency like 10Hz but how is its frequency response? Can it somehow distort the voltage sampling of microvolt? The E1DA Cosmos ADC is not for general audio use like listening to music. It's built only to measure noises of audio amplifiers so it doesn't have the normal front end audio shaping like done in other full fledge audio devices. But for the chip used. How is its frequency response (see below)? 

https://e1dashz.wixsite.com/index/cosmos-adc



If  you will say audio ADC is not good for sampling voltages, then what is the best ADC for instrumentation use with noise floor in the -128dB (A). see below.

2. The E1DA Cosmos ADC is very sought after in the audio world because it has a -128 dB (A) noise floor (or 6nV-7nV/Sqrt (Hz)). Since most amplifiers installed in complete equipment have total noise of 10nV/Sqrt (Hz) or higher. Can it literally be used to measure the amplitude of noise floor of other amplifiers (with no input)?  I have tested them and don't know how accurate it is:

The following is the noise floor of the E1DA Cosmos ADC. I tested the Cosmos using latest REW RTA without any input connected.





It has such low noise of -128.2db (A) (in A-weighted value) (between 6 and 7nV/Sqrt (Hz) voltage noise).  Can you find other ADC with such good noise? Again it's marketed for people to measure noises of audio amplifiers and not for music use. The reason it's not for music use is because they minimize all components to avoid noise.

I used the E1DA to measure an amplifier with a LF412 conditioning chip connected to the AMP01 amplifier (with no input) with gain set for 10X gain and bandwidth of 100Hz. This is the noise floor as measured by the E1DA.



The following is the noise with 10X gain and 1000Hz bandwidth chosen in the AMP01 amplifier.  I chose 10X gain because it is the minimum, there is no 0 gain in the switch.



How accurate are the tests?

Can you use the E1DA to test all existing amplifier equipments?  Will it accurately measure the noise floor of them like the above? In audio tests. They have standard bandwidth of 20,000 Hz. But if you will vary it to different bandwidths. Can the E1DA measure each bandwidth accurately or is it biased to measure only 20,000Hz audio bandwidth, and what is the technical reasoning if this is true? (Is it?)

Thank you.

[MasterT]

FYI:


 SAR is preferable IMHO, since it allows to split conversion and data transfer phases.
 DS (SD) ADC may pickup less noise from data bus only if oversampling is internaly implemented, thats seems not a case with audio adc. 

[Kleinstein]

The main limitations with audio ADCs are the low frequency cut off, possible extra LF noise (seems to be not the problem here) and often a not that stable gain (e.g. gain TC could be > 10 ppm/K). For the noise tests on amplifiers the ADC is not that critical at all. In the usual test the Amplfier to test provides quite some gain. If needed an additional amplifier is between the DUT and ADC. This way one can measure amplifier noise (referrred to the input) much smaller than the ADC noise. The gain of the amplifier(s) effectively attenuates the ADC noise.

It depends on the exact ADC / audio card how bad the LF cut of is. The noise curve alone may not show it. One could compensate some of the drop off, but only a limited amount.


A SD ADC also has it's pros: it has much of the AA filtering included and is thus easier to use. The higher speed of an SAR type ADC would not be really useful as much of the extra BW would be lost to the AA filter transition range. So the sampling rates are not directly comparable.

The software the the audio ADC should also be able to measure noise with different bandwidth or calculate spectral noise density values, as shown in the last 3 graphs. The accuracy should be good enough - with noise one rarely cares so much about high resolution or high accuracy. So the gain drift is not an issue here.

[WatchfulEye]

Subject to the limitations given by Kleinstein above, there is no reason why an ADC shouldn't be a particular problem as long as you are only interested in noise in the audio band (50-20 kHz).

There is a potential problem with using analysis software with an ADC which is not properly calibrated (or one or both are incorrectly configured), in that there may be systematic errors. For example, if your ADC has a +/- 5V full scale range, but your analysis software assumes +/-1 V FS, then there will be a systematic error with measured voltages being off by a factor of 5. For example, in the screenshots provided, the software is configured for a 1 V rms sine wave being full-scale, but this may not be a range supported by the E1DA unit. It is said to have a 1.7 V measurement range (among several other selectable ranges), but I don't know whether that is +/- 1.7V or 1.7V rms. This is important if you want numbers out. This particular ADC seems to have an active community on audio forums, and the designer of it seems to be happy to answer questions directly, so you should be able to find out how to set up your software and ADC from people experienced in its use.

Of course, if you are just looking for qualitative information (e.g. amplifier A is better or worse than B, or setting C is better than setting D), then this is less important.

Measuring amplifier noise floor is often not particularly demanding - because you are normally interested in "input referred" noise - and the amplifier will amplify its own noise. This means you can measure the output noise, and then divide the measurements by the amplifier gain. (Don't forget that amplifier noise floor can be sensitive to amplifier settings - especially gain, noise floor may be much higher at low gain).

[loop123]

Something puzzles me. The noise density in the nV/Sqrt (Hz) should not change with the bandwidth because the AMP01 has constant 5nV/Sqrt (Hz) from 12 Hz to 10kHz as shown below. But why does my REW RTA showed about 100nV/Sqrt (Hz) at 100Hz and about 1uV/Sqrt (H) at 1000Hz? Shouldn't nV/Sqrt (Hz) be constant since it is the noise density where you multiply by Sqrt (Bandwidth) to get the nV rms?



The input range I chose is 1.7Vrms. But isn't it 1 rms = 6.6 p-p so it's 1.7Vrms x 6.6 = 11.22 Volts P-P.

My Audacity has 1 to -1 for exactly 5V p - p.  Why is 1 to -1 not 11.22 Volts p-p? Btw.. when you use multimeter or dc output, is it in rms or peak-peak usually?

[TimFox]

In that graph, for f > 10 Hz, the noise spectrum is flat or "white", which means the noise voltage in a given bandwidth is proportional to the square root of the bandwidth.
Below 10 Hz, you are in "pink" or "1/f" part of the spectrum, where the "excess noise" found in all semiconductor and similar devices increases with decreasing frequency, so that the total noise over bands that include the low frequency region increases over the "white" voltage.
The "corner frequency", below which pink noise rears its ugly head over white noise, depends on the details of the active devices used in the circuit and requires that the circuit be powered externally.

[loop123]

In that graph, for f > 10 Hz, the noise spectrum is flat or "white", which means the noise voltage in a given bandwidth is proportional to the square root of the bandwidth.
Below 10 Hz, you are in "pink" or "1/f" part of the spectrum, where the "excess noise" found in all semiconductor and similar devices increases with decreasing frequency, so that the total noise over bands that include the low frequency region increases over the "white" voltage.
The "corner frequency", below which pink noise rears its ugly head over white noise, depends on the details of the active devices used in the circuit and requires that the circuit be powered externally.

I know but 5nV/Sqrt (Hz) should already described the noise voltage. In the last 2 images in my original message. The noise voltage at 100Hz vs 1000Hz bandwidth should have similar nV/Sqrt(Hz)  but it varies from 100nV/Sqrt (Hz) to 1uV/Sqrt (Hz), why? what components can make the noise density  change?

[floobydust]

Many audio ADC/DAC IC's incorporate "auto-mute" to achieve their magical S/N ratios. Some manufacturers are honest and give both specs, others just assume you're in the know.
You have to disable it to not get fooled or misled.

[loop123]

To rephrase above. When the bandwidth is increased, do you also need to change the noise density of the AMP01 from say 5nV/Sqrt(Hz) to 7nV/Sqrt(Hz)? Isn't it its graph already has the 5nV/Sqrt(Hz) from 10Hz to 10kHz bandwidth already? And to get the noise in rms, you multiple them by Sqrt (Bandwidth).

I mentioned this because someone in the audio tech asked me if my noise density really vary with bandwidth. Because he said noise density doesn't have to change with bandwidth.

In my actual tests with E1DA. The noise density measured changed or vary with bandwidth. Why?

In the last 2 images in my original message (also reproduced below). The noise voltage at 100Hz vs 1000Hz bandwidth vary instead of having constant nV/Sqrt(Hz). That is. It varies from 100nV/Sqrt (Hz) for 100Hz bandwidth to 1uV/Sqrt (Hz) for 1000Hz bandwidth. What components can make the noise density  change?

I'll illustrate by actual example of my circuits. The formula for noise already contain the equation to get noise in rms. For the AMP01, noise is 5nV/Sqrt (Hz), for the LF412 that conditions the signal prior to the AMP01. The noise is 25nV/Sqrt (Hz). So let's say the total noise is 30nV/Sqrt (Hz) (ignoring the contribution of other components like resistors).

For bandwidth of 100Hz.

Noise is 30nV/Sqrt (Hz) x Sqrt (100 Hz) = 30nV x 10 = 300nV rms

For bandwidth of 1000Hz.

Noise is 30nV/Sqrt (Hz) x Sqrt (1000Hz) = 30nv x 31.62 = 948nV rms

Let's say you use gain of 10 and run the amplifier without input at 100Hz and 1000Hz and the noise floor measured by the E1DA. (Here I multiply the gain to the noise density, correct?)

at 100Hz, 10 Gain. Why is the noise density not 30nV/Sqrt (Hz) x 10 gain = 300nV/Sqrt (Hz) but only 100nV/Sqrt (Hz) like the following actual noise measured by the E1DA?



at 1000Hz, 10 Gain. Why is the noise density not 30nVSqrt (Hz) x 10 gain= 300nV/Sqrt (Hz) but 1uV/Sqrt(Hz) like the following actual noise measured by the E1DA? 




Kleinstein, WatchfulEye.. you may be the only ones familiar with my setups so please comment as you know the major noise contributions or dynamics of them. Thanks.

[WatchfulEye]

I know but 5nV/Sqrt (Hz) should already described the noise voltage. In the last 2 images in my original message. The noise voltage at 100Hz vs 1000Hz bandwidth should have similar nV/Sqrt(Hz)  but it varies from 100nV/Sqrt (Hz) to 1uV/Sqrt (Hz), why? what components can make the noise density  change?

You keep mentioning this 5nV/Sqrt(Hz) noise spec, but this is not relevant. That is only the noise of one of the stages in your amplifier, the programmable gain amplifier stage, and even then only in optimal circumstances.

Going from your previous threads, your amplifier has several stages:
1. An input buffer and offset/CMRR trim stage. This stage has at least 40 nV/Sqrt Hz noise, and is the most importnat source of noise in most circumstances.
2. The programmable gain stage (Amp01). At medium and high gains (approx G>=1000), this gives optimal noise performance, but this noise is negligible compared to the noise of the input stage. At low gains, then the noise performance is absolutely terrible, at 300-500 nV/sqrt Hz, but likely to be irrelevant because low gains are used for large signals.
3. Finally, there are 2 additional gain and filtering stages, used for bandwidth selection and increasing the overall amplifier gain up to a max of 50,000.

If you want to measure the noise floor of your amp, it is best to do so when it is in the exact operating mode you wish to use.

Your noise spectrum figures are difficult to read (you have chosen sub-optimal settings for the spectrum acquisition). However, in both the 100 Hz and 1 kHz bandwidth cases, taking a representative estimate of noise at 50 Hz - your plots read approx 2.5 uV/sqrt Hz - after correcting for gain (1.7 V FS sine rms), and taking into account the amplifer gain of 10, this gives an overall input referred noise of approx 400 nV/sqrt Hz. This is about what I would expect due to the contibution of all the amplifier stages at this low gain.

I'd suggest when doing a spectrum measurement, you use a modest number of points 32-64k, select "forever" averaging, and collect a large number of averages (hundreds or thousands). This makes it much easier to read - see my random example (not measuring anything in particular).

Once you are happy with the acquisition setup, then you can test the amplifer at the gain you intend to use.

[loop123]

Watchfuleye, what software is the RTA in your last message?

If it's the REW RTA.. how did you run it without the hardware?  Did you use the 1000 Hz noise file?




the above was the RTA setting I used. Why can't you use FFT length of 4M instead of your 64k? or Averages of Exponential 0.97 instead of Forever?

You mean it's possible the E1DA has read them wrong reporting the 100nV/Sqrt (Hz) at 100Hz and 1uV/Sqrt (Hz) at 1000Hz?  that was why you couldn't figure how they were derived?  My original message was also about whether the E1DA can read noise floor accurate. It can't?

Edit: Or did you mean I must change the 1V rms to 1.7V rms in the RTA and the output will be accurate? Or must it be 5V peak to peak (changed to rms)?

[WatchfulEye]

Watchfuleye, what software is the RTA in your last message?

If it's the REW RTA.. how did you run it without the hardware?  Did you use the 1000 Hz noise file?

(Attachment Link)

It's the same software - REW. I just used my audio interface hardware.

the above was the RTA setting I used. Why can't you use FFT length of 4M instead of your 64k? or Averages of Exponential 0.97 instead of Forever?

Longer FFT gives more frequency resolution but takes longer - You don't need the frequency resolution or low frequencies.
Because noise is noisy, you need good amplitude resolution to see the patterns. This needs a lot of averages. The default exponential moving average in the REW software is not enough.

You mean it's possible the E1DA has read them wrong reporting the 100nV/Sqrt (Hz) at 100Hz and 1uV/Sqrt (Hz) at 1000Hz?  that was why you couldn't figure how they were derived?  My original message was also about whether the E1DA can read noise floor accurate. It can't?

Edit: Or did you mean I must change the 1V rms to 1.7V rms in the RTA and the output will be accurate? Or must it be 5V peak to peak (changed to rms)?

If you set the software scale to 1.7 V rms, I am confident that you will be getting accurate readings of output referred noise. You will need to divide by amplifier gain to get the input referred noise.
It will just be more useful if you have a clean plot - and this is obtained most quickly with short FFTs and lots of averages.

[dobsonr741]

Given you only inspect the noise floor, and not a signal vs. the noise it’s a problem in the software domain, not an ADC part selection. It allows you to mismatch the dynamic range of the ADC and the DUT.

[loop123]

Watchfuleye, what software is the RTA in your last message?

If it's the REW RTA.. how did you run it without the hardware?  Did you use the 1000 Hz noise file?

(Attachment Link)

It's the same software - REW. I just used my audio interface hardware.

the above was the RTA setting I used. Why can't you use FFT length of 4M instead of your 64k? or Averages of Exponential 0.97 instead of Forever?

Longer FFT gives more frequency resolution but takes longer - You don't need the frequency resolution or low frequencies.
Because noise is noisy, you need good amplitude resolution to see the patterns. This needs a lot of averages. The default exponential moving average in the REW software is not enough.

You mean it's possible the E1DA has read them wrong reporting the 100nV/Sqrt (Hz) at 100Hz and 1uV/Sqrt (Hz) at 1000Hz?  that was why you couldn't figure how they were derived?  My original message was also about whether the E1DA can read noise floor accurate. It can't?

Edit: Or did you mean I must change the 1V rms to 1.7V rms in the RTA and the output will be accurate? Or must it be 5V peak to peak (changed to rms)?

If you set the software scale to 1.7 V rms, I am confident that you will be getting accurate readings of output referred noise. You will need to divide by amplifier gain to get the input referred noise.
It will just be more useful if you have a clean plot - and this is obtained most quickly with short FFTs and lots of averages.

These are taken with your suggested 1.7Vrms,  64k and forever.

100Hz bandwidth, 10X gain



1000Hz bandwidth, 10X gain



Are you supposed to look at the peak or black lines? What do they mean? In the 1000Hz bandwidth case, the peak is 3.91uV/Sqrt (Hz), divided it by 10X gain is 0.391uV/Sqrt (Hz) or 391nV/Sqrt (Hz). Is this noise only for 1000Hz?  But then is it not you multiply 391nV/Sqrt (Hz) x Sqrt (Bandwidth) to get noise Vrms. So 391nV/Sqrt (Hz) x Sqrt (1000) = 391nv x 31.62 = 12364nV ??

In the case of the 100Hz bandwidth, the peak is 469.7nV/Sqrt (Hz) / 10 gain = 46.97nV/Sqrt (Hz)  Should you multiply this by Sqrt (100Hz bandwidth) to get nV rms?

But then, how do you derive the base formula for nV/Sqrt (Hz) where you can just multiply it by Sqrt (100Hz) or Sqrt (1000Hz) to get Vrms?

[WatchfulEye]

Are you supposed to look at the peak or black lines? What do they mean? In the 1000Hz bandwidth case, the peak is 3.91uV/Sqrt (Hz), divided it by 10X gain is 0.391uV/Sqrt (Hz) or 391nV/Sqrt (Hz). Is this noise only for 1000Hz?  But then is it not you multiply 391nV/Sqrt (Hz) x Sqrt (Bandwidth) to get noise Vrms. So 391nV/Sqrt (Hz) x Sqrt (1000) = 391nv x 31.62 = 12364nV ??

Ignore the peak line. Only the black "mean" line is relevant. The black line is the output-referred noise density, it tells you the amount of noise contained with a portion of the frequency spectrum.

If you have a frequency range of interest, then to compute the rms noise, requires integration. Where the noise density is constant (white noise) over the frequency band of interest, this can be simplified to N * Sqrt (BW) where BW is the bandwidth, and N is the representative noise density within the band. If the noise density is not constant, then numerical integration is required.

In the case of the 100Hz bandwidth, the peak is 469.7nV/Sqrt (Hz) / 10 gain = 46.97nV/Sqrt (Hz)  Should you multiply this by Sqrt (100Hz bandwidth) to get nV rms?

The peak measure is not relevant. Also, the 469.7 nV/Sqrt(Hz) measure appears to have been taken at 1000 Hz, which is outside of the band you are interested in and not representative of the band you are interested in, so this measure is irrelevant.

If all you are interested in is noise rms, then you don't need the noise density. You just need to filter your signal, and measure the amplitude rms directly. You already have an amplifier with built in filters. Conveniently, the REW software will measure and display rms voltage at the top right of the display.

So, by way of example, looking at your 1000 Hz filtered recording analysis. The overall noise amplitude at amplifier output is 74.5 uV, giving an input referred noise of 7.45 uV rms.

You mentioned in another thread, that a test signal you want to measure has an amplitude of 3.4 uV rms (10 uV p-p sine wave). As you can, see the expected signal to noise ratio is less than 0.5, meaning that your signal is unlikely to be recovered in a useful manner.

Your amplifier and recording setup will likely have much better performance with the amplifier set to higher gain. You should try repeating the measurements with the amplifier configured more appropriately for measuring signals of such a low amplitude (try a gain of 10k).

[loop123]

Are you supposed to look at the peak or black lines? What do they mean? In the 1000Hz bandwidth case, the peak is 3.91uV/Sqrt (Hz), divided it by 10X gain is 0.391uV/Sqrt (Hz) or 391nV/Sqrt (Hz). Is this noise only for 1000Hz?  But then is it not you multiply 391nV/Sqrt (Hz) x Sqrt (Bandwidth) to get noise Vrms. So 391nV/Sqrt (Hz) x Sqrt (1000) = 391nv x 31.62 = 12364nV ??

Ignore the peak line. Only the black "mean" line is relevant. The black line is the output-referred noise density, it tells you the amount of noise contained with a portion of the frequency spectrum.

If you have a frequency range of interest, then to compute the rms noise, requires integration. Where the noise density is constant (white noise) over the frequency band of interest, this can be simplified to N * Sqrt (BW) where BW is the bandwidth, and N is the representative noise density within the band. If the noise density is not constant, then numerical integration is required.

In the case of the 100Hz bandwidth, the peak is 469.7nV/Sqrt (Hz) / 10 gain = 46.97nV/Sqrt (Hz)  Should you multiply this by Sqrt (100Hz bandwidth) to get nV rms?

The peak measure is not relevant. Also, the 469.7 nV/Sqrt(Hz) measure appears to have been taken at 1000 Hz, which is outside of the band you are interested in and not representative of the band you are interested in, so this measure is irrelevant.

If all you are interested in is noise rms, then you don't need the noise density. You just need to filter your signal, and measure the amplitude rms directly. You already have an amplifier with built in filters. Conveniently, the REW software will measure and display rms voltage at the top right of the display.

So, by way of example, looking at your 1000 Hz filtered recording analysis. The overall noise amplitude at amplifier output is 74.5 uV, giving an input referred noise of 7.45 uV rms.

You mentioned in another thread, that a test signal you want to measure has an amplitude of 3.4 uV rms (10 uV p-p sine wave). As you can, see the expected signal to noise ratio is less than 0.5, meaning that your signal is unlikely to be recovered in a useful manner.

Your amplifier and recording setup will likely have much better performance with the amplifier set to higher gain. You should try repeating the measurements with the amplifier configured more appropriately for measuring signals of such a low amplitude (try a gain of 10k).

the following is taken with 10k gain (in both 1000Hz bandwidth selected)



so the Vrms of 36.28mV /10000 gain = 0.000003628 or 3.628uV rms noise?

and the noise density is 459.3uV/sqrt (Hz) /10000 gain = 0.0000000459 or 46nV/Sqrt (Hz)

for the following taken with maximum 50k gain



noise Vrms = 2.123V / 50k = 0.00004246  or 42.46uV rms... does it make sense? remember for my previous 10uV 50Hz signal you analyzed; the sine wave could still be resolved. Here the noise is much larger than it.

for noise density:  280.5uV/sqrt (Hz) / 50k gain = 5.61nV/Sqrt (Hz). 

Should the noise density be about 45nV/Sqrt (Hz)?   (5nV/Sqrt (Hz) for AMP01,  25nV/Sqrt (Hz) for LF412, 15nV/Sqrt (Hz) for other components like resistors in the circuit) Why 5.61nV/Sqrt(Hz)?

Also what are those peaks at 2kHz and so on not there at 10k gain?

[WatchfulEye]

noise Vrms = 2.123V / 50k = 0.00004246  or 42.46uV rms... does it make sense? remember for my previous 10uV 50Hz signal you analyzed; the sine wave could still be resolved. Here the noise is much larger than it.

Something is wrong at 50k gain. The ADC is clipping, leading to severe signal distortion (You can see the "max sample 0 db" which indicates that the ADC has clipped). There is probably a DC offset somewhere. When doing these noise measurements, you are shorting the amplifier inputs to each other and to common, aren't you? If the inputs are properly shorted, you will need to adjust the DC offset on the amplifier to bring it back into range.

At these high gains, you may need to adjust the DC offset from time to time.

Should the noise density be about 45nV/Sqrt (Hz)?   (5nV/Sqrt (Hz) for AMP01,  25nV/Sqrt (Hz) for LF412, 15nV/Sqrt (Hz) for other components like resistors in the circuit) Why 5.61nV/Sqrt(Hz)?

That sounds about right - there are 2x LF412, 2x 5k resistors, and the rest of the amp. So, from datasheet figures that adds up to somewhere in the region of 40 nV/Sqrt Hz at 1 kHz.   

[loop123]

noise Vrms = 2.123V / 50k = 0.00004246  or 42.46uV rms... does it make sense? remember for my previous 10uV 50Hz signal you analyzed; the sine wave could still be resolved. Here the noise is much larger than it.

Something is wrong at 50k gain. The ADC is clipping, leading to severe signal distortion (You can see the "max sample 0 db" which indicates that the ADC has clipped). There is probably a DC offset somewhere. When doing these noise measurements, you are shorting the amplifier inputs to each other and to common, aren't you? If the inputs are properly shorted, you will need to adjust the DC offset on the amplifier to bring it back into range.

At these high gains, you may need to adjust the DC offset from time to time.

Should the noise density be about 45nV/Sqrt (Hz)?   (5nV/Sqrt (Hz) for AMP01,  25nV/Sqrt (Hz) for LF412, 15nV/Sqrt (Hz) for other components like resistors in the circuit) Why 5.61nV/Sqrt(Hz)?

That sounds about right - there are 2x LF412, 2x 5k resistors, and the rest of the amp. So, from datasheet figures that adds up to somewhere in the region of 40 nV/Sqrt Hz at 1 kHz.

The inputs were not shorted but all floating. you mean I should short the +in, -in and ground together?

For the 10k gain. is the 46nV/Sqrt(Hz) (from 459.3uV/sqrt (Hz) /10000 gain) only for the 1kHz bandwidth selected or is it the formula for all bandwidth? I mean you multiply it by Sqrt (Hz bandwidth) to get the Vrms. but if the noise density is different for each bandwidth. how can you multiply each by Sqrt ( Hz bandwidth)? do you get what im saying here? I asked this several times pls elaborate on this. Many tnx.

[loop123]

noise Vrms = 2.123V / 50k = 0.00004246  or 42.46uV rms... does it make sense? remember for my previous 10uV 50Hz signal you analyzed; the sine wave could still be resolved. Here the noise is much larger than it.

Something is wrong at 50k gain. The ADC is clipping, leading to severe signal distortion (You can see the "max sample 0 db" which indicates that the ADC has clipped). There is probably a DC offset somewhere. When doing these noise measurements, you are shorting the amplifier inputs to each other and to common, aren't you? If the inputs are properly shorted, you will need to adjust the DC offset on the amplifier to bring it back into range.

At these high gains, you may need to adjust the DC offset from time to time.

Should the noise density be about 45nV/Sqrt (Hz)?   (5nV/Sqrt (Hz) for AMP01,  25nV/Sqrt (Hz) for LF412, 15nV/Sqrt (Hz) for other components like resistors in the circuit) Why 5.61nV/Sqrt(Hz)?

That sounds about right - there are 2x LF412, 2x 5k resistors, and the rest of the amp. So, from datasheet figures that adds up to somewhere in the region of 40 nV/Sqrt Hz at 1 kHz.

The inputs were not shorted but all floating. you mean I should short the +in, -in and ground together?

For the 10k gain. is the 46nV/Sqrt(Hz) (from 459.3uV/sqrt (Hz) /10000 gain) only for the 1kHz bandwidth selected or is it the formula for all bandwidth? I mean you multiply it by Sqrt (Hz bandwidth) to get the Vrms. but if the noise density is different for each bandwidth. how can you multiply each by Sqrt ( Hz bandwidth)? do you get what im saying here? I asked this several times pls elaborate on this. Many tnx.

Here I tied up the +IN, -IN and ground as you suggested. set to 50000gain, 1000Hz bandwidth, and not clipping.

integrated noise is 40.82mV / 50000 gain = 816nV rms or 0.816uV rms   It tallies to the 10uV 50Hz noise  you saw?

But for the noise density of 877uV/Sqrt (Hz) (in the RTA above)  divided by 50000 gain = 17.54nV/Sqrt(Hz)

Why is it smaller than the 45nV/Sqrt(Hz) that the components actually produced?             

[Kleinstein]

The data with the shorted inputs looks already much better.
The point at 1000 Hz is aready at a point where the response / gain goes down. A more reasonable frequency to look at is 200 Hz or so, so well within the flat region.
The noise there is higher by maybe a factor 2 or a little less. This may still be a bit less than the estimated 40 or 45 nV/sqrt(Hz) for the front end, but no longer much. The stimates for the noise are a bit crude, as the actual noise performance of the LF412 may scatter - some can be better and some can be worse.

[jonpaul]

read LT and ADI and BB notes and apps on noise in opamps.

See Jim Williams and Bob Pease.

j

[gf]

Noise density is a function of frequency. If you want to specify it as a single number for a particular frequency range, then this number can only be the average density for that range. Only for white noise, this number is the same for any frequency range. I don't know the ENBW of the involved filter. If I assume 1kHz, then I'd roughly estimate an average density of 40.82mv/(sqrt(1000Hz)) = 1.29mV/(sqrt(Hz) for the region below 1kHz. Referred to the input, with 50000x gain, that would be ~26nV/sqrt(Hz) then. However, for an exact calculation of the average density in a particular frequency range, it would really be necessary to integrate power density over this range. [ Btw, note that you have to integrate power density over frequency, i.e. you integrate V²/Hz, not V/(sqrt(Hz)). If you add two noise sources, then their powers (or squared voltages) sum up, not their voltages. ]

[loop123]

In the above I used 20Hz to 1000Hz in the distortion setting but the noise density didn't change, so it still integrates from 20Hz to 20kHz. How do you make it integrate over 20Hz to 1000Hz only? Also using the 200Hz flat response portion, the noise density is 1.323mV/Sqrt (Hz) / 50000 = 26.4nV/Sqrt (Hz). Still too low considering AMP01 has 5nV/rtHz and LF412 has 25nv/rtHz noise. It's not possible for the LF412 to become half noise only. That will be beyond milspec for them.

Can you manually try to integrate power density over frequency in the range above? using the V^2/Hz instead of V/Sqrt (Hz)? Would the result come close to 45nV/Sqrt (Hz)? Please compute.

Also I wonder if the gain of 50000 is really accurate. How do you test it? For example. For the 1.7Vrms in the E1DA. What voltage must appear in the 1 to -1 of Audacity. Is it 1.7Vrms? Maybe I can check whether 50000 gain is accurate by amplifying  say 10uV x 50000 times to come up with 0.5V and see if it would tally with the 1.7Vrms setting? I can use peak to peak or rms in the 10uV no problem since I'm not sure if the 10uV in the Netech signal generator is peak-to-peak or rms.

[loop123]

(Attachment Link)

In the above I used 20Hz to 1000Hz in the distortion setting but the noise density didn't change, so it still integrates from 20Hz to 20kHz. How do you make it integrate over 20Hz to 1000Hz only? Also using the 200Hz flat response portion, the noise density is 1.323mV/Sqrt (Hz) / 50000 = 26.4nV/Sqrt (Hz). Still too low considering AMP01 has 5nV/rtHz and LF412 has 25nv/rtHz noise. It's not possible for the LF412 to become half noise only. That will be beyond milspec for them.

Can you manually try to integrate power density over frequency in the range above? using the V^2/Hz instead of V/Sqrt (Hz)? Would the result come close to 45nV/Sqrt (Hz)? Please compute.

Also I wonder if the gain of 50000 is really accurate. How do you test it? For example. For the 1.7Vrms in the E1DA. What voltage must appear in the 1 to -1 of Audacity. Is it 1.7Vrms? Maybe I can check whether 50000 gain is accurate by amplifying  say 10uV x 50000 times to come up with 0.5V and see if it would tally with the 1.7Vrms setting? I can use peak to peak or rms in the 10uV no problem since I'm not sure if the 10uV in the Netech signal generator is peak-to-peak or rms.

Here is another information. The above was taken with 10k gain (with same 1000Hz bandwidth chosen). You can easily calculate that the 271.8uV/sqrt (Hz) / 10000 gain = 27nV/Sqrt (Hz).  Still fall short of the expected 45nV/Sqrt (Hz) noise. Could it be the E1DA is not accurate at all? Or maybe the LF412 is not really used at all to condition signal to enter the AMP01? So the noises are due to the resistors and not really the LF412? WatchfulEye mentioned in another thread:

"The BF412 in the BMA-200 acts as a pre-amplifier with gain of 2 and very high input impedance. It also provides trimming for CMRR and DC offset.  Due to its near infinite input impedance and very low bias current, it has near zero current noise, so is ideal for buffering very high impedance signal sources.

The AMP01 is bipolar input, rather than FET input like the BF412. As such it has higher bias currents and higher current noise, which means that if the impedance of your signal is more than about 200 kOhms, it will likely give more noise than the BF412."

The input impedance is 10k-20k. My concern is. Could the AMP01 be used without the BF412 at all (why did he refer to the LF412 as BF412, are they identical?), like perhaps the AMP01 engineers have designed it such that you don't need another amplifier to require before it. So the LF412 (BF412) has other use?

[Kleinstein]

The 200 Hz noise may still be plausible with the LF412. If would not be surprized to find lower noise with newer version and 20 nV/SQRT would still be plausible. So an overall noise of 26 nV/SQRT(Hz) would not be such a surprize (the noise adds as squares and thus not much added from the smaller parts). There is however still some oddity: the noise seems to be flat at all the way to some 20 Hz and thus only relatively little 1/f noise. So there may be something wrong with the front end / amplfiier circuit or a rather good example of the LF412. Less 1/f noise is possibly as an exception from better purity - not very likely, but possible.  I have seen TL031 with surprisingly low 1/f noise.

I don't see why the expected noise should be at 45 nV/sqrt(Hz) - more like 30 nV/sqrt(Hz with a bit more 1/f noise than observed.

A BF412 would normally be a BJT transistor and not an OP-amp.
As part of the amplifier there is also the possibility that the actually used part is not a LF412, but a different lower noise part.