On 04/20/2020 12:33 PM, Lukas Haase wrote:
Hi Marcus,
On 04/19/2020 09:26 PM, Lukas Haase via USRP-users wrote:
Hi,
https://kb.ettus.com/UBX#Noise_Figure lists 2-3dB noise figure,
UBX_Data_Sheet.pdf 2-4dB for UBX-160 (at 915 MHz).
I connect a 50 Ohm load to RX and create a simple gnuradio application with
USRP Source that calculates the RMS value of the sampled data (and plots the
noise in time domain).
samp_rate = 5 MHz.
RX Gain = 0 dB.
The value is 0.000113 RMS.
The time domain waveforms look a bit like quantization noise.
I repeat the experiment, this time with RX Gain value set to 37.5dB.
The value is 0.000841 RMS.
According to
https://files.ettus.com/performance_data/ubx/UBX-without-UHD-corrections.pdf,
NF=2-3dB for 37.5dB gain and ~23dB for 0dB gain. The high gain value coincides
with the values from the datasheet above.
Question 1: Is the reason that for 0dB it looks like quantization noise, that
the noise is smaller than LSB so I am effecively seeing the ADC noise?
Question 2: Is the ADC quantixation noise the main reason for the difference in
NF? (Note: I am familiar with Friis' equation)
Question 3: The noise difference for both cases is
20*log10(0.000841/0.000113)=17.43dB. According to the measurement PDF it should
be ca. 23-3=20dB. The difference is ~2-4dB. Is this just measurement
uncertainty/part-to-part mismatch or is something wrong with my approach?
Question 4: I repeat the last experiment by setting the "Ch0 Bandwidth [Hz]" to
0 (default), 5e6 and 1e6. I would expect that my rms value decreases by a factor of
sqrt(2) for each halfing of bandwidth. However, the value always stays around 0.000841
RMS, regardless of the bandwidth value. Why?
Question 5: According to theory, my captured signal is -174 + NF + 10*log10(BW)
= -174+3+10*log10(5e6)=-104dBm. Is it correct that my 0.000841 corresponds to
-104dBm input power? (If not, why not?). Or differently, -104dBm input power
corresponds to 20*log10(0.000841 / 1) = -61.5dBFS ?
Thank you,
Lukas
There'll be quantization noise, certainly.
But RF systems with variable gain use variable attenuators, and that
attenuation value is directly added to the noise figure of the stage that
follows it.
Now the magnitude of this effect obviously depends on where that
attenuator is placed in the gain chain. A 40dB attenuator after a 20dB
gain means that the 40dB of hoise is only "washed out" by the 20dB of
the previous stage(s).
Yes, that makes sense.
According to the schematics (UBX160 for 10MHz-1.5 GHz), there is one LNA
(MGA-62563), one 6-bit attenuator and another gain (NBB-400). These three can
be lumped together having a programmable gain of 0...37.5dB and a certain NF
(as a function of gain).
It follows from Friis':
Ftot = 1 + (Flumped - 1) + (Fadc-1)/(Glumped)
I think what is in UBX-without-UHD-corrections.pdf would be Ftot (total
measured noise figure).
With 0dB RX gain: Ftot=Flumped+Fadc-1. Now since the noise looks like
quantization noise it looks like Ftot~Fadc=23dB. That would make sense
(generally ADCs have huge noise figures).
With 37.5dB gain: Ftot = 1 + (Flumped - 1) + (Fadc-1)/(5600) ~ 1 + (Flumped -
1) = Flumped = 2dB. Since the noise looks less like quantization noise it seems
that the actual thermal noise is amplified now beyond the ADC quantization
noise and I am actually seeing thermal noise (as opposed to the 0dB case). Note
that an ADC noise figure includes quantization noise!
My question was if this is correct or if I am missing something (Questions 1-3).
It is entirely normal and conventional to quite noise figures at
different gains settings precisely because of this architectural necessity.
Yes.
Using a purely-analytic and numerical approach to determining the
magnitude of an input signal, given and intervening chain of
RF bits and pieces, whose noise figure and gain have uncertainties,
is what we call a "state of sin". You MUST calibrate with a known
signal source over your expected run-time parameter space. It's the
only way. It's the way laboratory instruments are routinely calibrated.
Using "datasheet analysis" which is basically what you're doing
above, leads to results that cannot be trusted to any degree of certainty.
Unfortunately I do not have a signal generator at hand or anything with which I
could create a known input signal.
But that again was not really my question. If number are off by a few dB
*because* of uncertainty/part-to-part mismatch (as I wrote) that's OK.
It is precisely because of accumulated uncertainty that a "datasheet RF
chain analysis" for deriving precision estimates is doomed to failure.
A well-matched resistor at ambient produces -174dBm/Hz at 25C. At
higher gain settings, you should be able to see changes in the
temperature of the resistor based on this.
My question is if my approach/understanding is right.
In particular I do not understand Question 4 (why does noise not reduce if I
reduce bandwidth).
If you're varying *analog* bandwidth, rather than sampling rate, be
aware that UBX doesn't have variable analog bandwidth. It's always fixed.
Furthermore, I'd be interested if Question 5 is conceptually correct.
Conceptually, I don't see any problem with it, but it very-squarely
enters "consider a spherical cow" territory. You CANNOT use a purely
arithmetic analysis, due to uncertainties. I would, in fact,
encourage you to acquire a decent broad-band, calibrated, noise source for
you lab so that you can do Y-factor analysis, if for no other reason
than to satisfy yourself that the noise equations work.
Lukas
_______________________________________________
USRP-users mailing list
USRP-users@lists.ettus.com
http://lists.ettus.com/mailman/listinfo/usrp-users_lists.ettus.com
