Capability as spectrum analyzer

Dear Cleverscopians,

I'd appreciate to get pointed out, if there is a flaw in my reasoning or in the way I use Cleverscope. I was measuring the low frequency noise floor of the instrument, but actually it is the white noise floor I'm interested in. I'm enclosing the measured spectrum when the input is 50-ohm terminated. It shows about -125dB-volts when RBW=0.61Hz, i.e. the trace indicates a noise floor of ~700nV/rtHz.

Now, the +/-20mV full range and a 14-bit ADC would imply 2.5uV quantization step. The ADC seems to be still sampling at the full 100MSa/s (which is good), so I'd expect the quantization noise to be distributed evenly over the 50 MHz Nyquist band. The result is 0.34 nV/rtHz, lower than the 1 nV/rtHz Johnson noise of my termination. I know that the above calculation is *way* optimistic: the ENOB of you converter is probably much less than 14, there are prefactors in the calculation which I have ignored and so on. Still, I would have expected much lower noise floor than the 700 nV/rtHz.

In case someone is interested, I also enclose another plot which shows the dynamic range explicitly: here I'm applying almost full-scale 1kHz, which resolves quite nicely.

As a comparison, there is also the noise floor of the Agilent 89640B on its most sensitive -30dBm range. The plot only shows the 1/f part of the spectrum, but the white part settles somewhere around -155dBm at RBW=1Hz, in other words the noise floor is roughly 4 nV/rtHz, probably limited by the LNA. The -30dBm full scale, i.e. 7 mVrms or +/-10 mV is in the same ballpark as Cleverscope's +/-20mV. Although the 89640B probably cost 50x the Cleverscope on its day, its digitizer is comparable at 12 bits and 95 MSa/s.

What gives? I'm not complaining: the Cleverscope is intended primarily as an oscilloscope; it just would be great if you could use it to substiture a 10x more expensive spectrum analyzer.

I have three hypotheses:
(i) the Cleverscope's LNA is just so noisy and it dominates.
(ii) Even distribution of the quantization noise over the Nyquist band (effective bandwidth limitation from the full 100 MSa/s or BW=50MHz to BW=10kHz indicated in the plot) would requre dithering, which is not implemented.
(iii) I have made a blunder of some sort.

Any insight?

-------------------------------------------
By the way: I had noticed that the device averages spectra in the frequency domain, but did not realize it has been accidental (I have always happened to have trig = AUTO) until I saw the thread here which says averaging is trig mode dependent.

Dear Cleverscopians, I'd appreciate to get pointed out, if there is a flaw in my reasoning or in the way I use Cleverscope. I was measuring the low frequency noise floor of the instrument, but actually it is the white noise floor I'm interested in. I'm enclosing the measured spectrum when the input is 50-ohm terminated. It shows about -125dB-volts when RBW=0.61Hz, i.e. the trace indicates a noise floor of ~700nV/rtHz. Now, the +/-20mV full range and a 14-bit ADC would imply 2.5uV quantization step. The ADC seems to be still sampling at the full 100MSa/s (which is good), so I'd expect the quantization noise to be distributed evenly over the 50 MHz Nyquist band. The result is 0.34 nV/rtHz, lower than the 1 nV/rtHz Johnson noise of my termination. I know that the above calculation is *way* optimistic: the ENOB of you converter is probably much less than 14, there are prefactors in the calculation which I have ignored and so on. Still, I would have expected much lower noise floor than the 700 nV/rtHz. In case someone is interested, I also enclose another plot which shows the dynamic range explicitly: here I'm applying almost full-scale 1kHz, which resolves quite nicely. As a comparison, there is also the noise floor of the Agilent 89640B on its most sensitive -30dBm range. The plot only shows the 1/f part of the spectrum, but the white part settles somewhere around -155dBm at RBW=1Hz, in other words the noise floor is roughly 4 nV/rtHz, probably limited by the LNA. The -30dBm full scale, i.e. 7 mVrms or +/-10 mV is in the same ballpark as Cleverscope's +/-20mV. Although the 89640B probably cost 50x the Cleverscope on its day, its digitizer is comparable at 12 bits and 95 MSa/s. What gives? I'm not complaining: the Cleverscope is intended primarily as an oscilloscope; it just would be great if you could use it to substiture a 10x more expensive spectrum analyzer. I have three hypotheses: (i) the Cleverscope's LNA is just so noisy and it dominates. (ii) Even distribution of the quantization noise over the Nyquist band (effective bandwidth limitation from the full 100 MSa/s or BW=50MHz to BW=10kHz indicated in the plot) would requre dithering, which is not implemented. (iii) I have made a blunder of some sort. Any insight? ------------------------------------------- By the way: I had noticed that the device averages spectra in the frequency domain, but did not realize it has been accidental (I have always happened to have trig = AUTO) until I saw the thread here which says averaging is trig mode dependent.

The Forum software mangled my pictures together. So, here is the Cleverscope noise floor.

The Forum software mangled my pictures together. So, here is the Cleverscope noise floor.

... here is the almost-fullscale sine, registerd by the Cleverscope.

... here is the almost-fullscale sine, registerd by the Cleverscope.

...and here is the Agilent 89640 noise floor.

BES: just out of interest - are all those spurs below 1 kHz (of about -140 dBV) from the spec an, or your signal?

...and here is the Agilent 89640 noise floor. BES: just out of interest - are all those spurs below 1 kHz (of about -140 dBV) from the spec an, or your signal?

Hello ttemsk,
Thanks for posting these images. You are right, it would be nice if the spectrum performance were as good as an instrument 10x more expensive. Darn.
The reason it is not as good is that the wide band noise of our 120 MHz BW front end is aliased back into the Frequency span you have chosen. This suggests a way to reduce the noise floor - operate with as wide a frequency span as you can manage with the available frequency resolutions. You were using about 1 Hz frequency resolution. So I set the Frequency Span to 1 MHZ, and used a resolution of 1.95 Hz (it was only a 4 M sample unit). Make sure the number of frames is 2.

The second thing you can do if you don't need the bandwidth is to use filtering to reduce the components aliased back into the pass band. I used Settings/Acquisition settings, and turned on the moving average filter, with a 1.28 us time constant. I enabled the Moving average LED's. Click the Filter button on the control panel (though with the most sensitive range we turn on filtering by default).

Attached is a graph showing the results with averaging in the frequency domain - Auto captured - as you have observed. We see the 5 kHz noise floor down around -130 dBV this way - about 10 dB improvement over what you got.

If you can average coherently - using Triggered - in the time domain, you can do better yet. Attached is a time averaged plot, with about 6 dB improvement to -136 dBV. This represents about 158 nV in 2 Hz, or 112 nV/rt Hz. Still not ideal, but better. The Agilent spec an has the benefit of a tracking filter that means it only sees the RBW bandwidth when digitizing. In our next product , we will have a bigger FPGA, and we will do the same.

I hope this explains it.

Hello ttemsk, Thanks for posting these images. You are right, it would be nice if the spectrum performance were as good as an instrument 10x more expensive. Darn. The reason it is not as good is that the wide band noise of our 120 MHz BW front end is aliased back into the Frequency span you have chosen. This suggests a way to reduce the noise floor - operate with as wide a frequency span as you can manage with the available frequency resolutions. You were using about 1 Hz frequency resolution. So I set the Frequency Span to 1 MHZ, and used a resolution of 1.95 Hz (it was only a 4 M sample unit). Make sure the number of frames is 2. The second thing you can do if you don't need the bandwidth is to use filtering to reduce the components aliased back into the pass band. I used Settings/Acquisition settings, and turned on the moving average filter, with a 1.28 us time constant. I enabled the Moving average LED's. Click the Filter button on the control panel (though with the most sensitive range we turn on filtering by default). Attached is a graph showing the results with averaging in the frequency domain - Auto captured - as you have observed. We see the 5 kHz noise floor down around -130 dBV this way - about 10 dB improvement over what you got. If you can average coherently - using Triggered - in the time domain, you can do better yet. Attached is a time averaged plot, with about 6 dB improvement to -136 dBV. This represents about 158 nV in 2 Hz, or 112 nV/rt Hz. Still not ideal, but better. The Agilent spec an has the benefit of a tracking filter that means it only sees the RBW bandwidth when digitizing. In our next product , we will have a bigger FPGA, and we will do the same. I hope this explains it.

The time averaged plot:

The time averaged plot: