Siglent SVA1015X and SVA1032X 1.5, 3.2GHz Spectrum & Vector Network Analyzers

[S13]

:'(
I have measured S21 on full span, 1601 points, default configuration (15 dB internal attenuator) and input disconnected, in order to measure noise with my upgraded SSA3021X+.
Two plots, single (yellow) and with 100 averages (pink).
The averaging is clearly trace based, because the AVG counter increases one trace at a time. The averaging gives a negligible reduction of noise, not the 20 dB that one expected if it were vector based on each point. Even the trace noise shows very little improvement. Most improvement happens during the first 5-10 averages.

Why is your average not "averaging"?
Here is my example:



I dont know what setting you are using that causes the letter [M] to be displayed behind the "TR1 S21" text?

[hendorog]

:'(
I have measured S21 on full span, 1601 points, default configuration (15 dB internal attenuator) and input disconnected, in order to measure noise with my upgraded SSA3021X+.
Two plots, single (yellow) and with 100 averages (pink).
The averaging is clearly trace based, because the AVG counter increases one trace at a time. The averaging gives a negligible reduction of noise, not the 20 dB that one expected if it were vector based on each point. Even the trace noise shows very little improvement. Most improvement happens during the first 5-10 averages.

I appreciate the comparison. It's indeed disappointing and it seems there is potential for improvement. I wonder if it is a bug or a "feature"?

The improvement during the first few averages is expected due to the exponential averaging equation.
The current average counter is used in the averaging algo, not the overall averaging factor, so as the trace count increases the effect of each trace gets less.
This is the standard algo and is used by Keysight and likely all other manufacturers.

In terms of vector or scalar averaging, my money is on vector averaging. My SSA X-R shows about a 20dB improvement with 100 averages. (Sorry prob not the right thread for this device)
(I am running older FW - 6.2R5)
Point averaging in my mind is IFBW in a VNA. These devices do not have an adjustable IFBW, so any point averaging is fixed.

However, I may have discovered a trick to work around the limitations in the hardware using some custom software.

Increase the point count to max, averaging as before, and then apply _vector_ smoothing.
The idea is that it is kind of a "poor mans IFBW" setting, as the smoothing is operating on complex data points which are very close in frequency. similar enough to IFBW which is averaging complex points at a single frequency.
Smoothing is much faster than averaging as it is calculated on the trace data just before it is displayed. The max point sweep is slower though so it ends up the same as 10 averages time wise.

The picture shows what that looks like. Red limit line is -100dB in all cases and scale is the same as previous posters but just shifted up a bit due to the lower floor. (ref -60, 10/ but ref pos 7)

Top (Yellow) is 1001 points, dim memory trace had no avg and bright yellow trace has 100avg.
Middle (Cyan) is 10001 points, 100 avg
Bottom (Pink) is 10001 points, 100 avg + 1% smoothing (100 points smoothing)

The smoothing is just maths applied to the averaged trace data, I can click smoothing on and off in the bottom two traces and it instantly switches. This is not averaging the smoothed traces.

Interestingly at the high frequency end, the smoothing behaves just like scalar smoothing would on a VNA - a thin line around the middle of the noisy trace. But at the lower end it has a disproportionate effect.
I think on these devices, that the noise floor is not actually all pure noise. Which is where this technique seems to shine.

Obviously smoothing will change the shape of sharp features, so needs to used with care.

[tautech]

Anyone notice the interesting display layout hendorog shared above ?
AFAIK its online debut !

It's a universal VNA application SW package that he's designed and been developing for a bit that gather the raw data from the VNA and also lets you take full control of the VNA.....much like the webserver but far far more powerful.

I get to beta test it some and check its online updates are downloading as they should......

[RoV]

The averaging is clearly trace based, because the AVG counter increases one trace at a time. The averaging gives a negligible reduction of noise, not the 20 dB that one expected if it were vector based on each point. Even the trace noise shows very little improvement. Most improvement happens during the first 5-10 averages.

Why is your average not "averaging"?

I dont know what setting you are using that causes the letter [M] to be displayed behind the "TR1 S21" text?

The M stands for memory, the yellow trace was stored.
I have made other tests. My SVA does weird things when averaging background noise in VNA mode, I suspect this is a bug of last FW update SVA1000X_V3.2.2.6.3R16_EN, or perhaps they tried an averaging algorithm optimized to smooth traces not purely made of noise.
According to the number of points and to the frequency band selected, it sometimes averages like your picture and sometimes not at all. In some cases it even reduces the background noise!

[gf]

Interestingly at the high frequency end, the smoothing behaves just like scalar smoothing would on a VNA - a thin line around the middle of the noisy trace. But at the lower end it has a disproportionate effect. I think on these devices, that the noise floor is not actually all pure noise.

Yes, it's unlikely that the noise from the separately measured points is correlated.
It could probably be a higher residual isolation error at high frequencies?

Otherwise, your results look very plausible
I just wonder how representative they are for SVA1000X, since X-R is a different class.

[tautech]

Interestingly at the high frequency end, the smoothing behaves just like scalar smoothing would on a VNA - a thin line around the middle of the noisy trace. But at the lower end it has a disproportionate effect. I think on these devices, that the noise floor is not actually all pure noise.

Yes, it's unlikely that the noise from the separately measured points is correlated.
It could probably be a higher residual isolation error at high frequencies?

Otherwise, your results look very plausible
I just wonder how representative they are for SVA1000X, since X-R is a different class.

SVA1075X and SSA3075X-R have very similar specs.

When we drop down to the 3.2 GHz models the differences become more apparent.
https://int.siglent.com/u_file/download/25_11_27/SVA1000X_DataSheet_DS0701X_E05G_.pdf
https://int.siglent.com/u_file/download/25_11_27/SSA3000X-R_DataSheet_DS0703R_E02G_.pdf

[willie.from.texas]

Had to help move a boat from the Florida Keys to the Bahamas so I've been off-line for the last week. What a great discussion! Thought I would share my final results. I spent some time doing a tune on the duplexer using the VNA.  There was plenty of performance using the VNA for tuning single cavities but the nulls for the dual cavities were quite noisy. I completed the final tuning for the nulls using the spectrum analyzer. Thought I would share my results. I also averaged 64 traces together, which which took significantly longer than the SA mode, and still contained a significant amount of noise. The averaged VNA measurements were not useful for final tuning.
Thanks again for all the comments!