Rigol DHO804 Test and Compare Thread

[Mechatrommer]

When I see screenshots showing 62.5MSa/s, 625000 points, and RBW 100 (assuming that 100Hz is not really the -3dB RBW, but actually the frequency bin spacing), then I get the impression that the FFT size is the full record length of 625000 points, which would also imply that this FFT implementation can handle sizes that are not a power of two. Of course, there has to be a limit to the maximum FFT size, so the question remains what happens if the acquired records are longer than the limit. What subset of samples is selected? Start of the record? Center? Or are long records even split into multiple chunks of the maximum FFT size?

If we distrust the numbers displayed on the screen, ist it possible to export the FFT math trace to a text file? If yes, how many points does it export if full span (0...fs/2) is selected? The number of saved points and the difference between adjacent frequencies should give an indication for the actual FFT size and frequency bin spacing (unless only decimated screen data are saved - that's useless, of course).

you people maybe right about bin size vs RBW, but that probably due to some advanced feature of FFT that may not accessible to normal people, such as spectrum leakage and varying window function effect on actual spectrum power, until you people can provide more reliable materials to read, i think rigol refering RBW as bin spacing as defined by link i provided. if rigol implementing your peole's definition of RBW, then there is no way of knowing how many points. but then... what practical benefit to know how many points? other than swinging dick contest this brand is lower points than the other. just read the goat damned FFT/SA the like old people did it... no need number of pts or question this brand is more reliable than others... ymmv.

(attached image is for illustration or amusement only, dont be too serious! life is short. name erased to avoid swinging dick contest, source: https://www.youtube.com/watch?app=desktop&v=S8jrpCoZyx8&t=1953s) i cannot erase voice sorry

[Martin72]

If we distrust the numbers displayed on the screen, ist it possible to export the FFT math trace to a text file?

Directly from the FFT menu not, only the peak/marker table.
But I'll try another thing, maybe that works.

Unfortunately this (save/recall-->save-->from memory, after this from screen) did not work, everything is recorded and the maths display is probably not included.
What I also noticed is that if you call up such system functions while maths (FFT) is active, the scope starts to sweat a bit and is paralysed by the processing of the inputs.

[rf-loop]

When I see screenshots showing 62.5MSa/s, 625000 points, and RBW 100 (assuming that 100Hz is not really the -3dB RBW, but actually the frequency bin spacing), then I get the impression that the FFT size is the full record length of 625000 points, which would also imply that this FFT implementation can handle sizes that are not a power of two. Of course, there has to be a limit to the maximum FFT size, so the question remains what happens if the acquired records are longer than the limit. What subset of samples is selected? Start of the record? Center? Or are long records even split into multiple chunks of the maximum FFT size?

If we distrust the numbers displayed on the screen, ist it possible to export the FFT math trace to a text file? If yes, how many points does it export if full span (0...fs/2) is selected? The number of saved points and the difference between adjacent frequencies should give an indication for the actual FFT size and frequency bin spacing (unless only decimated screen data are saved - that's useless, of course).

you people maybe right about bin size vs RBW, but that probably due to some advanced feature of FFT that may not accessible to normal people, such as spectrum leakage and varying window function effect on actual spectrum power, until you people can provide more reliable materials to read, i think rigol refering RBW as bin spacing as defined by link i provided. if rigol implementing your peole's definition of RBW, then there is no way of knowing how many points. but then... what practical benefit to know how many points? other than swinging dick contest this brand is lower points than the other. just read the goat damned FFT/SA the like old people did it... no need number of pts or question this brand is more reliable than others... ymmv.

(attached image is for illustration or amusement only, dont be too serious! life is short. name erased to avoid swinging dick contest, source: https://www.youtube.com/watch?app=desktop&v=S8jrpCoZyx8&t=1953s) i cannot erase voice sorry

Not perfectly to you even when I quote you.

1st.

Bins are intervals between samples in frequency domain.
Frequency domain include fact that there need be 2 samples.
All know that if we have example 1GSa/s (sampling frequency 1GHz)  FFT full scale in not  1GHz. It is 0-500MHz...   2 samples  means frequency interval in frequency domain... 1 Bin, one Δf.  One FFT bin need two sample (in time domain)
(example for carify: 1GSa/s and 524288pts Δf=1.9073kHz. Siglent display Δf 1.91kHz) (If you use FFT full span frequency  for Δf (for Bin) you need use FFTpts/2)

Spectrum analyzers, old analog... normally RBW mean filter 3dB width.
 
This RBW is so commonly used like it in many places just for meaning filter width as it is used tens of years in many instruments. It is confusing if some use RBW as Bin frequency interval. Better way imho, is to tell Δf and example Siglent have selected this way.

But as can see in my previous one image some Siglent oscilloscope may tell both,  Δf  and RBW  and my opinion is that this is good way to avoid confusion. In this case RBW change when user change FFT window, because RBW_3dB really change depending filter parameters. It looks like also one filter type, example  Hanning may have bit different filter factor in different implements depending some deeper details. Example Siglent hanning RBW3dB is not just 1.44 x Δf (bin).

My opinion is that if use RBW  as RBW_Bin it need also tell somehow for avoid confusion.

Least I have default in my mind after tens of years using many kind of instruments that when I see frequency domain and there RBW I will automatically asume it is RBW_3dB until specially noted exception. Because this have been so common practice. And it is same also in these SA what have full digital IF  and  example FFT (more or less wide realtime or step swept FFT) instead of analog filters. Newer seen there RBW is FFT bin).


One maestro for this mess is National (NI) when they simply say: "The resolution bandwidth (RBW) determines the fast-Fourier transform (FFT) bin size, or the smallest frequency that can be resolved."


In this Keysight page there is some good example about different FFT window and RBW_3dB and RBW_Noise

GW Instek is one example about good information in this matter.

MDO2000A data sheet have nice detail. 

FFT Windows,  FFT Factor:
Hanning 1.44 ; Rectangular 0.89 ; Hamming 1.30 ; Blackman 1.68

[Martin72]

Yeah, I was just playing around, this time with my flir i60...
...The scope after about 2 hours of operation.

[Martin72]

FFT again, going through several times/div.

Timebase Samplerate Memory FFT-SR RBW

10µs        1.25G        125Kpt   1.25G 10K
20µs        1.25G        250Kpt   1.25G 5K
50µs        1.25G        625Kpt   1.25G 2K
100µs      625M         625Kpt    625M 1K
200µs     312.5M       625Kpt  312.5M 500
500µs   156.25M   781.25Kpt 156.4M 199.9
1ms         62.5M        625Kpt   62.5M 100
2ms       31.25M        625Kpt  31.25M 50
And so on.
I don't understand the values at 500µs, these "crooked" values in the FFT, nor the memory, which is 625Kpt before and after, but 781.25kpt in this time base.
Rigol specifies the maximum number of FFT points as "up to 1Mpt".
But how could you find out how many points are currently being used if this value is not displayed?
BTW,
Not only is the RBW not updated at times, the graphically displayed span is not updated either.
Both can be refreshed by moving the vertical offset.

[ebastler]

But how could you find out how many points are currently being used if this value is not displayed?

Seems like the number of data points listed in your "memory" column is also used for the FFT. With that assumption, the listed "RBW" (which is really delta_f) is consistent with the sample rate and number of data points given, delta_f = samplerate / memory.

I don't understand the values at 500µs, these "crooked" values in the FFT, nor the memory, which is 625Kpt before and after, but 781.25kpt in this time base.

Bit strange indeed. It looks like they wanted to maintain the systematic 500/200/100 cadence for delta_f. If delta_f = 200 Hz and sampling rate = 156.25 MHz are given, you land at that odd number of data points.

[gf]

FFT again, going through several times/div.

Have you ever seen a FFT sample rate different from the acquisition sample rate?
What happens if you increase the memory depth beyond 1M points (i.e. beyond the maximum FFT size)?

[Martin72]

Right here and now, just shooting from the hip:
500µs Auto Memory, 500µs 10Mpts.
Note the FFT sampling rate.

[ebastler]

Right here and now, just shooting from the hip:
500µs Auto Memory, 500µs 10Mpts.
Note the FFT sampling rate.

The 156.4 can't be right; as gf said, how could it be different from the acquisition rate? Must be a rounding error, which also results in the awkward 199.9 bandwidth figure. We have seen in other places (axis scales and table entries) that Rigol struggles with proper rounding...

[gf]

10Mpts.png

Seems that it down-samples^*) if the record length is larger than the maximum FFT size.
And "RBW: 200" suggests that we got 625000 FFT points @125MSa/s.

So eventually it seems to do both here:
1) 10x down-sampling, which reduces 10M points to 1M points
2) Truncate the 1M points to 625k FFT points

^*) And since the noise fllor does not roll off sharply a little bit below Nyquist (62.5 MHz), it's likely down-sampling w/o decimation filtering.

[TurboTom]

Wouldn't worry too much about this since the most relevant window functions von Hann, Blackman-Harris and rectangular are available and appear to work correctly.

I would. A flattop window is important to get good amplitude accuracy for frequencies that are not integer multiples of the bin spacing. The maximum amplitude error of the flattop window ¹⁾ is only 0.016 dB. Hann, Blackman and various others still have significant scalopping loss for frequencies that fall in the middle between two bins. Blackman-Harris has also ~0.83 dB. And a rectangular window leaks as hell if the frequency is not an integer multiple of the bin spacing.

¹⁾ I consider the Matlab variant here, but there exist several other variants too

Whatsoever, it shouldn't be too difficult to correct this: In the scope's directory /rigol/resource/window/ there are all the window functions as binary arrays of 2^20 single precision (float32, little endian) numbers stored. Obviously, the file "flatTop_1048576.hex" is defective since it starts and ends with some value of -0.853 while it should be something very close to 0.

But since it isn't too complicated to generate a correct file, those who need the proper flattop window shape could easily modify it and replace the file. Nevertheless, this should have been Rigol's job. 

[Martin72]

Nevertheless, this should have been Rigol's job.

That's exactly what it should be.
If you're still wondering how easy it is to hack the new generation scopes....
These are evaluation boards with a screen and a basic software framework and if something is wrong, you can fix it yourself.

[TurboTom]

I just generated a new Flat Top window file as per the formula described here. The result looks much better than the original but there's still some side lobes emerging from the noise. I cannot tell for sure how flat the top actually is, but if someone wants to try, here's the download link. 



Edit: Sorry I had to update the download link since I found a small error in my math... Now it's corrected and the FFT with the FlatTop window looks really nice 
See attached screenshot: UR - FlatTop; LL - Blackman-Harris; LR - von Hann

P.S. I could just barely control myself not to use an AM signal. 

Edit: Download link updated. Alternativeley, you can download a windows EXE file that will generate the window file here.

[gf]

Here is the frequency response of the spectrum analysis filter corresponding to Matlab's flattopwin().
The flatness in the range -0.5...+0.5 is important.
Stopband rejection is > 68 dB. I'm surprised that you seem to get more

[ebastler]

I just generated a new Flat Top window file as per the formula described here. The result looks much better than the original but there's still some side lobes emerging from the noise. I cannot tell for sure how flat the top actually is, but if someone wants to try, here's the download link. 

Very nice work, both finding the file and fixing it -- thank you!

Is it possible that this messed-up Flat Top window has been a Rigol tradition forever? My old DS1054Z shows very similar "leaky" spectral lines when using the Flat Top window. No easy way to swap out the window function on that scope, I'm afraid...

[Martin72]

Oops, forget to switch off my DHO804, now it´s (nearly)24h burn-in proofed...

Earlier I took the 1khz reference signal as the source and went through all the windows.
Then I did the same with the Siglent SDS1104X-E, but of course I checked the 1khz reference again to see if there were any differences.
Here are the pictures.

[Martin72]

Continued, rigol got triangle, siglent not.

[Martin72]

So, the scope is now switched off, this time I haven't forgotten it.
Finally, three more things:
One picture where the scope "forgot" to change the scale.
Then the thing with the pop-up menu regarding Peaksearch, am I the only one who thinks the respective arrow direction is wrong?
I would do it the other way around.
And it's no longer a secret that I think the visual display of the tables is great.
But...
Too much space is taken up, e.g. the index column could be much narrower, and there is no need for 3 or even more decimal places for the other values.
So you would have a little more space for the signal, which brings me back to saying that the window function is great if you could influence it more in terms of position and size.

[ebastler]

The FFT results look pretty nasty in comparison. Is there a general problem, maybe with the non-power-of-two number of data points?

I share your feelig that too much space is wasted for frames, decorations, long numbers etc., give the small size of the display. I have posted a small rant in the bugs & wishlist thread a couple of weeks ago (or did I just mean to do that?).

The upward/downward arrow is not meant to indicate an action, but a current status of the collapsible window, I guess. I have seen this elsewhere quite often (on websites and also in Windows or MS Office?). But I guess you could argue either way. Just like I prefer to change the "touchpad scrolling direction" on every laptop I use for more than a few minutes...

[Martin72]

See attached screenshot: UR - FlatTop; LL - Blackman-Harris; LR - von Hann

Wait a second...von Hann....2GSa/s....What scopemodel do you have...

[TurboTom]

See attached screenshot: UR - FlatTop; LL - Blackman-Harris; LR - von Hann

Wait a second...von Hann....2GSa/s....What scopemodel do you have...

Of course, it's a DHO1k, but the flattop window is equally broken as on DHO800 / 900 instruments. Moreover, I simply don't call the "Hanning" window function this way since the man behind this mathematics was the Austrian meteorologist Julius von Hann, so the window function should be correctly called "von Hann" or "Hann" window. Maybe I'm pathetic but IMO we don't need to "americanize" every scientist's name... 

[gf]

Earlier I took the 1khz reference signal as the source and went through all the windows.

Looks broken on your Rigol. And not just Flattop, the others look weird, too.

In order to show maximum leakage, don't choose exactly 1kHz, but rather 997.5 Hz in conjunction with your Rigol settings, and 998.97 Hz with your Siglent settings. IOW, us a frequency f = (N+0.5)*delta_f, where N is an integer. Calculate with enough decimal places, i.e. delta_f = 4.76837 and not just 4.77 for your Siglent settings.

[Martin72]

This was taken from the calibrate signalsource of the scopes, but yes, I could use my generator for this.
Tomorrow.

[2N3055]

See attached screenshot: UR - FlatTop; LL - Blackman-Harris; LR - von Hann

Wait a second...von Hann....2GSa/s....What scopemodel do you have...

Of course, it's a DHO1k, but the flattop window is equally broken as on DHO800 / 900 instruments. Moreover, I simply don't call the "Hanning" window function this way since the man behind this mathematics was the Austrian meteorologist Julius von Hann, so the window function should be correctly called "von Hann" or "Hann" window. Maybe I'm pathetic but IMO we don't need to "americanize" every scientist's name... 

It's not even "americanizing".. It is just wrong.  "Von Hann" or "Hann" is the name.

[Mechatrommer]

Maybe I'm pathetic but IMO we don't need to "americanize" every scientist's name...