Trying to display bandwith via Math on Siglent SDS2k+/2kHD/800X HD

[gf]

For a single impulse (differentiated edge) inside the FFT window, definitively use a rectangular window function.

If multiple periods of the pulse train are within the FFT window, you still need to use a non-rectangular windowing function. And you need to make sure that you have at least (say) 10 periods of the pulse train inside the FFT window to keep the amplitude error of the resulting comb teeth within the scalloping loss of the window function. A rectangular window is not suitable in this case. See attachment.

EDIT: With a huge number of FFT points, the leakage of the rectangular window no longer has such a strong effect as demonstrated, but the scalloping loss remains, and it is still worse than with a flattop window.

[Martin72]

I found an application note by Lecroy that is enlightening for me, at least on the first three pages.
I'm slowly understanding more and more, it takes a while, but still.

[gf]

SA44_Pulsetrain_1MHz_W1ns_RT500ps

@Performa01, I was thinking about that again.

The spectrum of rectangular impulses is not flat, but is expected to have a sinc-shaped roll-off. And the tapering of the edges introduces additional roll-off. So they are not well suited for an application that requires a flat stimulus spectrum (unless they are so narrow that the roll-off can be neglected -- but then they can no longer be generated by the AWG).

What do you think about using a train of sinc-like pulses instead of rectangular pulses? They are bandwidth limited (so they can be reproduced by an AWG based on Shannon-Nyquist reconstruction from samples), and they still have a flat spectrum up to a certain frequency (with a sharp roll-off beyond this corner).

Plots of two example pulses are attached, designed for a 950-1000MHz transition band and for a 1000-1250MHz transition band. According to the specs, I assume that the SDG7000A should be able to reproduce these waveforms well. [By the way, does the SDG7000A promise 0.3dB amplitude flatness up to 1GHz?]

[Performa01]

What do you think about using a train of sinc-like pulses instead of rectangular pulses? They are bandwidth limited (so they can be reproduced by an AWG based on Shannon-Nyquist reconstruction from samples), and they still have a flat spectrum up to a certain frequency (with a sharp roll-off beyond this corner).

It certainly is a good idea and I have done a few experiments with this, as described below.

Plots of two example pulses are attached, designed for a 950-1000MHz transition band and for a 1000-1250MHz transition band. According to the specs, I assume that the SDG7000A should be able to reproduce these waveforms well. [By the way, does the SDG7000A promise 0.3dB amplitude flatness up to 1GHz?]

Thanks a lot for your effort again! The SDG7102A “promises” +/- 0.3 dB flatness up to 1 GHz, you can find an actual frequency response measurement (SDG7102A_1000MHz_-10dBm) attached.

This is for sine waves though; we cannot expect the spectrum of a pulse to be that accurate.

I had a quick go at it and found a pre-defined Sinc pulse amongst the built-in arbitrary waveforms. This is 32768 points though and the pulse width is some microseconds. Of course, I can downsample it, yet the shortest possible length is 64 points. At 2.5 GSa/s, the resulting frequency was much too high at ~39 MHz, yet I still did a first check with the SA:


SDG7102A_AWG_Sinc_64pts

This is the corresponding peak table:

1,39.071413 MHz,-15.08 dBm,0.000000 Hz,0.00 dB
2,78.128540 MHz,-15.39 dBm,39.057127 MHz,-0.32 dB
3,117.185668 MHz,-15.49 dBm,78.114255 MHz,-0.42 dB
4,156.271366 MHz,-15.38 dBm,117.199953 MHz,-0.30 dB
5,195.299922 MHz,-15.42 dBm,156.228509 MHz,-0.35 dB
6,234.357050 MHz,-15.49 dBm,195.285637 MHz,-0.42 dB
7,273.442748 MHz,-14.99 dBm,234.371335 MHz,0.09 dB
8,312.499876 MHz,-15.38 dBm,273.428463 MHz,-0.31 dB
9,351.557003 MHz,-14.72 dBm,312.485590 MHz,0.36 dB
10,390.642702 MHz,-14.80 dBm,351.571289 MHz,0.28 dB
11,429.699829 MHz,-14.60 dBm,390.628416 MHz,0.47 dB
12,468.728385 MHz,-14.61 dBm,429.656972 MHz,0.47 dB
13,507.814084 MHz,-14.80 dBm,468.742671 MHz,0.28 dB
14,546.871211 MHz,-15.11 dBm,507.799798 MHz,-0.04 dB
15,585.928338 MHz,-15.06 dBm,546.856925 MHz,0.02 dB
16,625.014037 MHz,-16.12 dBm,585.942624 MHz,-1.05 dB
17,664.071164 MHz,-15.14 dBm,624.999751 MHz,-0.07 dB
18,703.128292 MHz,-16.35 dBm,664.056879 MHz,-1.28 dB
19,742.185419 MHz,-15.12 dBm,703.114006 MHz,-0.04 dB
20,781.271118 MHz,-22.15 dBm,742.199705 MHz,-7.08 dB

It can be seen that the harmonics should be accurate to 0.5 dB up to at least 586 MHz. Generally, the pretty accurate odd harmonics up to 742 MHz are quite striking, just like the fact that even the 64 points sinc-pulse at 2.5 GSa/s can still produce such strong harmonics up to 742 MHz.

Of course, there is no point in trying to display bandwidth with a high frequency like 39 MHz. Even though it would have been tempting to generate my own version of a Sinc pulse, maybe getting the real pulse a little faster with still enough “padding” samples to achieve a reasonably low repetition frequency when played back, I’ve rather utilized the sequence-function of the AWG, by just adding 2436 samples of a zero signal to the 64 points Sinc.

With this, the time domain signal looks like this:


SDS6204_Pro_H12_Sinc_Pulse

The pulse width is 772 ps and the rise time should be around 360 ps (assuming ~230 ps rise time of the SDS6204).

With the padding, the pulse is now played back at a rate of 1 MHz at 2.5 GSa/s.

The result on the SDS800X HD is not too bad – certainly much better than all the previous attempts to measure its bandwidth using a pulse train.


SDS824X HD_FFT_Sinc_1MHz

This looks about right, including the cable reflections because of less than perfect impedance matching.

[Martin72]

The SDG7102A “promises” +/- 0.3 dB flatness up to 1 GHz

According to your picture, this is once again a typical siglent understatement.
Nice generator btw....

[gf]

The SDG7102A “promises” +/- 0.3 dB flatness up to 1 GHz, you can find an actual frequency response measurement (SDG7102A_1000MHz_-10dBm) attached. This is for sine waves though; we cannot expect the spectrum of a pulse to be that accurate.

Why? if it can reproduce sine waves up to 1GHz, then I would expect that it can reproduce any bandwidth limited signal up to 1GHz. Why is a sine wave special in this regard? Would be interesting, though, if AWG vs. AFG makes a difference.

Even though it would have been tempting to generate my own version of a Sinc pulse, maybe getting the real pulse a little faster

I tried to make custom .csv files. The data are flat (0.1%) up to 950 MHz. One is for AWG mode, with 2500 samples for 1:1 output @2.5GSa/s, and the second one with 32k AFG samples which is supposed to be sent at a repeat rate of 1MHz, and which needs to be resampled by the DDS (unavoidable if the AFG wants a wavetable with exactly 32k samples for a period). I just don't know if I got the CSV format correct.

...with still enough “padding” samples to achieve a reasonably low repetition frequency when played back

I also got only 183 AWG samples for the FIR, the rest up to 2500 samples are zero-padding. AFG has more samples, due to the higher effective sample rate of 32768MSa/s for the wavetable.

Octave code:

Code: [Select]


pkg load signal

function generate(sample_rate, transition_band, filename)

  [n,Wn,beta,ftype] = kaiserord(transition_band,[1 0],[0.001 0.001],sample_rate)
  impulse = fir1(n,Wn,ftype,kaiser(n+1,beta),'noscale');
  impulse *= 1/max(impulse);
 
  % zero-pad to get 1MHz repeat rate
  samples = [ impulse zeros(1,sample_rate-length(impulse)) ];

  % quantize
  samples = floor(samples*8192+0.5)/8192;

  % time axis
  t = [0:length(samples)-1]/sample_rate/1e6;

  % write CSV file
  f = fopen(filename, "w");
  fprintf(f,"data length,%d\n", length(samples))
  fprintf(f,"frequency,1.0E+6\n")
  fprintf(f,"amp,2\n")
  fprintf(f,"offset,0\n")
  fprintf(f,"phase,0\n")
  fprintf(f,"\n\n\n\n\n\n\n")
  fprintf(f,"xpos,value\n")
  fprintf(f,"%.8E,%15.13f\n",[t; samples])
  fclose(f);

  figure
  plot(1e9*t(1:length(impulse)), impulse);
  grid on; xlabel("ns")

  figure
  [H,f] = freqz(impulse/sum(impulse),1,100000,sample_rate);
  plot(f,(abs(H))); grid on; xlabel("Mhz")

endfunction

close all
transition_band = [950 1000] % MHz
generate(2500,transition_band,"AWG.csv")
generate(32768,transition_band,"AFG.csv")



[Performa01]

The SDG7102A “promises” +/- 0.3 dB flatness up to 1 GHz, you can find an actual frequency response measurement (SDG7102A_1000MHz_-10dBm) attached. This is for sine waves though; we cannot expect the spectrum of a pulse to be that accurate.

Why? if it can reproduce sine waves up to 1GHz, then I would expect that it can reproduce any bandwidth limited signal up to 1GHz. Why is a sine wave special in this regard? Would be interesting, though, if AWG vs. AFG makes a difference.

I agree with that, yet my statement was meant more general. The much more popular DDS-technique (which Siglent’s AFG-mode effectively is) is well known for its problems with pulse reproduction, particularly jitter. It should be obvious that a jittery pulse will waste energy in sidebands, hence its harmonics cannot be accurate.

Being aware of this, I’ve used Siglent’s AWG-mode and the tools offered by the instrument (decimating, padding) to tweak the internally stored arbitrary waveform (Sinc).

I tried to make custom .csv files. The data are flat (0.1%) up to 950 MHz. One is for AWG mode, with 2500 samples for 1:1 output @2.5GSa/s, and the second one with 32k AFG samples which is supposed to be sent at a repeat rate of 1MHz, and which needs to be resampled by the DDS (unavoidable if the AFG wants a wavetable with exactly 32k samples for a period). I just don't know if I got the CSV format correct.

Excellent work, thank you very much! Both files worked beautifully right away – but then again, what else to expect from someone who figured out Siglent’s binary format without further assistance – just with the not so fool-proof documentaton?

I tried the AFG version of the Sinc pulse first and I think it is not much different from the internal one.

As expected, your AWG version of the Sinc pulse is quite a bit better than the decimated internal version, just look at the parameters:


SDS6204_Pro_H12_PR-Sinc_Gf-AWG

The pulse width is even narrower at just 650 ps now, and the rise time is less than 250 ps!  I think this is going to be at the absolute limit of what to expect from an SDG7000A output.

Of course, I was curious how well the theory meets practice this time and did a FFT with the SDS6204, to check the frequency response of the generated pulse.


SDS6204_Pro_H12_FFT_Sinc_Gf-AWG

The frequency response pretty much meets the expectations. Of course, we don’t get 0.1% accuracy even for DC, let alone up to 1 GHz, yet max. 0.31 dB deviation up to 500 MHz and <0.7 dB up to 1 GHz is not bad, taking the combined tolerances of both the AWG and (particularly) the DSO into account.

We actually see a steep cut-off beyond 950 MHz, but the noise floor is already pretty close at about 25 dB below. This is just because of the very low energy in the wanted signal. We need to acknowledge that this is a 650 ps wide 1 volts high pulse occurring once every microsecond, hence the duty cycle is just 0.065%.

[gf]

The much more popular DDS-technique (which Siglent’s AFG-mode effectively is) is well known for its problems with pulse reproduction, particularly jitter.

The one sample period edge jitter of some DDS generators (when they generate pulses or square wave) is usually an aliasing artifact. It happens if the samples in the wavetable are not properly bandwidth-limited for the down-sampling done by the DDS. It does not happen with properly bandwidth-limited samples. You cannot fill the wavetable with [ -1 -1 ... -1 -1   1 1 ... 1 1 ] and expect a jitter-free resampling at any frequency, but you must pre-filter the wavetable according to the desired waveform repeat rate and DDS sample rate. Then this jitter is gone.

The phase truncation errors of a traditional DDS are a different issue, but they are usually much lower, and they can be reduced to a negligible amount by either using a huge wavetable (say > 1Mpts), or by interpolationg the wavetable instead of truncating the phase to the nearest wavetable entry. With a 32k wavetable, linear interpolation is quite sufficient. And I would expect a €€€€€ AWG to do that. So DDS does not need to be bad per se. But TrueArb (or however it is called by different manufacturers) offers more flexibility, of course.

I tried the AFG version of the Sinc pulse first and I think it is not much different from the internal one.

You did play it with 1MHz, right? Up to which frequency was it spectrally flat? If it starts rolling off below 950MHz, then the generator was obviously rather "conservative" when it did pre-filter the wavetable.

SDS6204_Pro_H12_PR-Sinc_Gf-AWG

The pulse width is even narrower at just 650 ps now, and the rise time is less than 250 ps!  I think this is going to be at the absolute limit of what to expect from an SDG7000A output.

We could go up to say 1200MHz with a steep transition band to 1250MHz. Then you could see the generator's intrinsic roll-off up to 1200 MHz.

For measuring a scope's frequency response (which is the topic of this thread), it's IMO better to keep the impulse bandwidth below the half sample rate of the scope in order to avoid aliasing (just in case). That's why I did choose 950MHz in this example, leaving 50MHz for the transition band, and virtually no power beyond 1GHz.

Btw, I find it interesting that the impulse is no longer symmetric. So the phase response already suffers somewhat at 950MHz. I guess that's mostly due to the analog reconstruction filter.

Of course, I was curious how well the theory meets practice this time and did a FFT with the SDS6204, to check the frequency response of the generated pulse.

SDS6204_Pro_H12_FFT_Sinc_Gf-AWG

The frequency response pretty much meets the expectations.
...
Of course, we don’t get 0.1% accuracy even for DC, let alone up to 1 GHz, yet max. 0.31 dB deviation up to 500 MHz and <0.7 dB up to 1 GHz is not bad, taking the combined tolerances of both the AWG and (particularly) the DSO into account.

The performance of this gear is amazing

With 0.1% I just meant that the impulse was designed for 0.1% (0.01dB) passband ripple and -60dB stopband ripple. The samples are indeed within these limit. I did not expect the analog reproduction to be within this limit as well. The impulse is, btw, not an exact sinc, but it is truncated and windowed to achieve the given ripple and transition band width with a finite number of samples.

We actually see a steep cut-off beyond 950 MHz, but the noise floor is already pretty close at about 25 dB below. This is just because of the very low energy in the wanted signal. We need to acknowledge that this is a 650 ps wide 1 volts high pulse occurring once every microsecond, hence the duty cycle is just 0.065%.

Sure, peak-to-average power ratio (PAPR) is almost 2500. You could of course use a larger FFT size to lower the noise floor, but then you'll see the individual comb teeth again. You can't have everything in life

[Martin72]

 

[Performa01]

The much more popular DDS-technique (which Siglent’s AFG-mode effectively is) is well known for its problems with pulse reproduction, particularly jitter.

… So DDS does not need to be bad per se. But TrueArb (or however it is called by different manufacturers) offers more flexibility, of course.

Well, yes, I’ve seen early DDS-generators, with 10 bits resolution, 10 MHz bandwidth and max. 1024 Samples memory for the ARB-mode, which didn’t sound too optimistic with regard to the predefined waveforms either. That was also the time when embedded systems used some 8051/2 MCU derivates at best, whose performance couldn’t even match a today’s Arduino.

These things have burned into my brain and of course I’ve not seen any serious problems of this kind with the Siglent SDG6052X and SDS7102A that I have here. Yet I’m still a bit wary whenever I’m using an AWG in DDS-mode close to its limits.

I tried the AFG version of the Sinc pulse first and I think it is not much different from the internal one.

You did play it with 1MHz, right? Up to which frequency was it spectrally flat? If it starts rolling off below 950MHz, then the generator was obviously rather "conservative" when it did pre-filter the wavetable.

Sorry, I’ve been sloppy once again. I just loaded the AFG-version while the generator still was in AWG-mode and got the usual ~76.3 kHz. For the reasons given above, I didn’t look more closely but hurried to get the AWG-version of the Sinc pulse instead. Yes, there is indeed a huge difference, as I will show later.

We could go up to say 1200MHz with a steep transition band to 1250MHz. Then you could see the generator's intrinsic roll-off up to 1200 MHz.

Maybe you feel like providing such a waveform, so we could have a go at it?

Btw, I find it interesting that the impulse is no longer symmetric. So the phase response already suffers somewhat at 950MHz. I guess that's mostly due to the analog reconstruction filter.

That was already the case with the decimated version of the internal waveform. But not with the unmodified original at 2.5 GSa/s, where I’ve now measured 400 ns pulse width and 216 ns rise time...

With 0.1% I just meant that the impulse was designed for 0.1% (0.01dB) passband ripple and -60dB stopband ripple. The samples are indeed within these limit. I did not expect the analog reproduction to be within this limit as well. The impulse is, btw, not an exact sinc, but it is truncated and windowed to achieve the given ripple and transition band width with a finite number of samples.

Yes, of course. And I forgot to mention another source of inaccuracies at higher frequencies – it is the impedance matching. The output impedance of the SDS7102A isn’t particularly close to a pure 50 Ω resistor, so there we could lose the occasional tenth of a dB at certain frequencies.

Speaking of accuracy, I’ve of course wondered if I could improve the result by providing a better match at the generator output. At first I just grabbed a cheap 6 dB BNC-inline attenuator and connected it directly at the generator output. This degraded the frequency response quite a bit – no wonder, these attenuators most likely have ½ W through hole resistors inside, so we cannot expect 0.1 dB accuracy up to 1 GHz.

Okay, I have much better attenuators, rated for up to 18 GHz, but of course they are SMA. So I’ve tried such an attenuator together with several SMA-cables, one of them the same quality as the previous BNC-cable, just 100 cm instead of 50. Unfortunately, all these experiments led to nothing. If anything, accuracy was at least 0.3 dB worse.

In the end, the original setup, i.e. a short (50 cm) low-loss Hyperflex 5 BNC-cable between generator and DSO worked best after all.

Sure, peak-to-average power ratio (PAPR) is almost 2500. You could of course use a larger FFT size to lower the noise floor, but then you'll see the individual comb teeth again. You can't have everything in life

How true! 😉


Now for a closer look at the AFG-variant of the Sinc pulse. First in the time domain to obtain the precise measurements:


SDS6204_Pro_H12_PR-Sinc_Gf-AFG

We can conclude that the measurements are essentially the same as the AWG-version. One would expect there are also no differences in the frequency domain, yet there were some, with the AFG-version of the Sinc pulse performing slightly worse by at least 0.1 dB.

I noticed that there were still minor fluctuations of the measurements in the marker table, so I decided to repeat the measurements with both versions of the pulse, this time with 1024x averaging to get really stable readings.

First the AFG-version:


SDS6204_Pro_H12_FFT_Sinc_Gf-AFG

This is up to 0.2 dB worse than the measurement in the previous posting with AWG-version and 64x averaging.

And now the AWG-version with 1024x averaging:


SDS6204_Pro_H12_FFT_Sinc_Gf-AWG

The results are even better than in the first run.

So we actually have deviations of ~0.4 dB up to 500 MHz and <0.85 dB up to 1 GHz with the AFG version, yet ~0.35 dB up to 500 MHz and <0.65 dB up to 1 GHz with the AWG version. I’m pleased with both versions! 😉

[gf]

We could go up to say 1200MHz with a steep transition band to 1250MHz. Then you could see the generator's intrinsic roll-off up to 1200 MHz.

Maybe you feel like providing such a waveform, so we could have a go at it?

Sure. See attachment.

So we actually have deviations of ~0.4 dB up to 500 MHz and <0.85 dB up to 1 GHz with the AFG version, yet ~0.35 dB up to 500 MHz and <0.65 dB up to 1 GHz with the AWG version. I’m pleased with both versions! 😉

So the AFG works nicely too

[Performa01]

Many thanks to member gf, who provides the arbitrary waveforms that give me the opportunity to get the most out of my Siglent gear. I was really curious what we’d get from the 1200 MHz version of the Sinc pulse. And yes, this is going to be spectacular, fasten your seat belts!

As always, I’ve analyzed it in the time domain at a fast time base first, to obtain the precise pulse parameters.


SDS6204_Pro_H12_PR-Sinc_Gf-AWG1200

The pulse width is narrower again and has now finally landed at just 500 ps, while the rise time is less than 200 ps. My calculation says ~188 ps, but this is rather uncertain territory, as the numbers are now even below the rise time of the DSO. Yes, we can still calculate the signal rise time from the measured value and the DSO’s own rise time – as I did, but we cannot absolutely rely on the specification and I suspect that the rise time of the SDS6204 might actually be a little faster than 230 ps. If that’s the case, the signal rise-time is correspondingly higher. For instance, if the DSO rise time actually is just 200 ps, then the generated Sinc pulse would have a rise time of 220 ps. Either way, we’re dealing with pretty fast edges here.

Now we expect a rather spectacular result for the frequency domain analysis, and I can state in advance that the result managed to surprise and impress me at the same time…


SDS6204_Pro_H12_FFT-Sinc_Gf-AWG1200

Above is the frequency response graph with measured values at strategic frequency values from 10 MHz to 1.2 GHz. And this is the summary of it:

Up to 200 MHz, the deviations are <0.1 dB.
Up to 1.12 GHz, the deviations are <0.3 dB.
At 1.2 GHz the deviation is about -3 dB, so that appears to be the actual bandwidth of the generator.

I can only repeat over and over again that we see the combined tolerances of AWG, cabling and DSO here.

[gf]

So it is approximately flat up to 1.1 GHz

Another aspect which limits the maximum "usable" frequency is the image rejection in the 2nd Nyquist band. We do not want frequencies beyond 1.25 GHz in the generated signal. Still they are present. The image of 1200MHz (at 1300MHz) is attenuated by ~15dB, and the image of 1100MHz (at 1400 MHz) is attenuated by ~25dB. But that's already the noise floor of your FFT, and that's also the point where your frequency scale ends. Could you connect it also to your SA and swep from say 1GHz to 2.5GHz, hoping that you can get a better dynamic range if you use a small enough RBW? At which frequencies beyond 1.3GHz do we get say -30dB, -40dB, -50dB, -60dB?

EDIT: And what's the attenuation at 1500MHz, since this is the image frequency of the nominal 1GHz bandwidth.

EDIT: And where does the spur at 1.25GHz come from? Is it from the siggen or from the scope? Iguess the latter. The SA will reveal.

[Performa01]

Could you connect it also to your SA and swep from say 1GHz to 2.5GHz, hoping that you can get a better dynamic range if you use a small enough RBW? At which frequencies beyond 1.3GHz do we get say -30dB, -40dB, -50dB, -60dB?

I don’t have a high-end spectrum analyzer and while my SA44 is clearly better than average at low frequencies and narrow spans, it doesn’t provide narrow RBW-settings at wider spans, and above 1.5 GHz its noise floor starts rising considerably. Consequently, I can only provide a scan over 1-2.5 GHz with 6.5 kHz RBW:


SA44_Sinc_Gf-AWG1200_2-1-2.5GHz

This first scan however already shows pretty much everything relevant. The signal doesn’t seem to get any lower than -45 dBc. I still feel more comfortable with the DSO, even though it has lots of spurs at these low levels. For a more detailed view, I divided the spectrum in several 500 MHz wide sections. First is the critical part around the Nyquist frequency of the AWG at 1.25 GHz:


SDS6204_Pro_H12_FFT-Sinc_Gf-AWG1200_1-1.5GHz

We see the first minimum at 1.25 GHz – unfortunately the DSO generates a spurious signal there, so we have to measure the level at a 1 MHz higher frequency and get -36.46 dBc. At 1.29 GHz, the maximum level beyond Nyquist is measured as -13.25 dBV. At a random frequency of 1.462 GHz, the level is -40.7 dBc and we won’t see much less as the initial spectrum has already hinted on.

Here is the frequency range around 1.5 GHz:


SDS6204_Pro_H12_FFT-Sinc_Gf-AWG1200_1.25-1.75GHz

There is absolutely nothing special to see, except for the pulse spectrum at levels between approximately -42 and -47 dBc extending to 2 GHz.

In the range from 1.75 – 2.25 GHz the spectrum decreases in level very slowly, yet not falling below -48 dBc.


SDS6204_Pro_H12_FFT-Sinc_Gf-AWG1200_1.75-2.25GHz

EDIT: And where does the spur at 1.25GHz come from? Is it from the siggen or from the scope? Iguess the latter. The SA will reveal.

Yes, even without SA I could verify that this stronger spur comes from the SDS6204. But here is the proof from the SA:


SA44_Sinc_Gf-AWG1200_1.5GHz

[gf]

Thanks for trying. My conslusions are:

1) We cannot go higher than 1000-1050 MHz if we want to avoid the "hill" between 1250 and 1000 1400 MHz (i.e. we should not see this hill with the previous 950 MHz impulse).

2) The stopband attenuation of the digital upsampling/reconstruction filter seems to be not more than about 40 dB.

The next point of interest would be the region around 3900 MHz (say +-300 MHz). That's where the analog reconstruction filter is mostly challenged.

I think you'll understand what I mean if you look at the attached figure, which shows what the (continuous) staircase output of an ideal DAC is expected to look like after 2x upsampling to 5 GSa/s with a digital reconstruction filter. This staircase is the signal going into the analog reconstruction filter.

And the second figure shows the spectrum of this staircase (and yes, it is not bandwidth-limited, but repeats periodically, decaying proportionally to 1/f). The "valley" up to 3600 MHz was already suppressed by the digital reconstruction filter (which is very helpful), but anything we see in this spectrum beyond 3600 MHz must still be eliminated (or sufficiently attenuated) by the analog filter. I have also drawn in orange what the frequency response of an example analog filter might look like.

EDIT:

Yes, even without SA I could verify that this stronger spur comes from the SDS6204. But here is the proof from the SA:

Looks like typical interleaving spurs (offset spurs). Calibration may help, but I guess that the resolution of the calibration is limited to integral ADC codes, while fractional resolution might be required to get rid of the spurs.

[Performa01]

Okay, we are slowly getting a bit off-topic here (discussing the output filtering in an SDG7000A AWG instead of displaying bandwidth on an entry level DSO), but then again, the problems that led to the actual topic will be solved at one point. In the meantime, we can look at noisy spectra, e.g. from 3.6 to 4.2 GHz for our trusty old Sync pulse, designed by gf:


SA44_Sinc_Gf-AWG1200_3.9GHz

As we can see, we can see nothing...