Siglent SDS2000X HD, 12bit, 2GSa/s high definition oscilloscope serie.

[Performa01]

And you are sure the difference isn't created by you taking the "fast track" in relation to the long method used by Martin?

Yes, absolutely sure. Even for a traditional swept spectrum analyzer, the required total sweep time for accurate amplitude measurements would be less than 4 milliseconds.

The formula for the sweep time is as follows:

SWT = k * Span / RBW²;

In our example: SWT = 3 * 1 GHz / (910 kHz)² = 3*10e9 / 828.1e9 = 3.623 ms;

K is a factor, depending on the filter characteristics, usually between 1 and 3 - and I’ve assumed the worst case here.

But our FFT on a DSO isn’t a swept analyzer, rather a realtime analyzer which requires much less time, let’s try to calculate this:

This is a 8k FFT, that means we have 4096 frequency bins, each of them ~244 kHz wide. Even more important, the resolution bandwidth (RBW) is even wider, ~910 kHz with the Flattop window, so we gain quite a safety margin if we just consider the frequency step instead of RBW.

With a 600 seconds sweep for 1 GHz, the signal stays within one frequency bin for 600 s * 244 kHz / 1 GHz = 146.4 ms – this is about 40 times longer than the entire sweep time required for a swept analyzer. In other words – it’s orders of magnitude slower than what is needed for accurate measurements.

I don’t know the exact speed of the FFT with these settings, but it certainly is less than 100 ms. That means, at 600 seconds sweep speed, we get at the very least one full FFT for the width of a single frequency bin. And this is exactly what we want to get a nice contiguous trace with only one single sweep.

I’ve also tried different sweep speeds and the net result has always been the same.

Or the gen quality?

That’s a valid question. Apart from the better matching because of the attenuator, I also had to raise the output level to +10 dBm. So maybe the output level accuracy suffers at the higher level, or has different flatness?

I’ve looked up my old measurements, which indicate an accuracy of +0.38 / -0.14 dBm at 0 dBm output and +0.34 / -0.15 dBm at +10 dBm for the frequency range from 100 kHz to 1 GHz (it would be vastly better if we stick to just 100 MHz). At 600 MHz, the amplitude error is about -0.1 dB for both level settings. So this is not the reason.


Thinking about the differences, there are only two things:

1.   The attenuator, where I’ve used a high quality 18 GHz Narda part to ensure best accuracy.
2.   The cable – now I needed SMA connectors so couldn’t use the Hyperflex 5 BNC-BCN cable from before.

I’ve grabbed the first random cable, which looked trustworthy – but a closer look now revealed that it’s just crappy RG58 C/U. So right now I’ve repeated that test with a proper Hyperflex 5 low loss cable, this time with SMA connectors and BNC adapters


SDG7102A_1000MHz_-10dBm-2
.
Yes, this makes quite some difference. The -3 dB bandwidth is only 570 MHz, but that is within the ballpark. Particularly the drop of only 1 dB at 500 MHz is a very welcome result.
The less than stellar flatness between 350 MHz and 450 MHz remains and is just a property of my instrument.

[tv84]

I’ve grabbed the first random cable, which looked trustworthy – but a closer look now revealed that it’s just crappy RG58 C/U. So right now I’ve repeated that test with a proper Hyperflex 5 low loss cable, this time with SMA connectors and BNC adapters

Damn, in that case I must rephrase:

Good professor + Good equipment + Good cables = Good result 

[Martin72]

Hi,
Earlier I reduced the dwell time from 1sec to 300ms - so one sweep took about an hour, yesterday it was three...
As you can see, it doesn't matter, it still looks very good.



Inspired by Perfoma01's cable change, I did the same and used a cheap noname RG58 from amazon with N to BNC adapter instead of the self-made Kusch Hyperflex5 with the solid connectors.



It's amazing what effect this has, the course is more "unsteady".
I still have enough Hyperflex5 cable here and will make various BNC cables from it.
And throw everything else in the bin.

[Performa01]

As I've demonstrated, you can reduce the dwell time even further - 50 ms should work just perfectly fine.

A small remark:

I chose a level of 0 dBm for this test so that we can use the absolute numbers directly (and don't need the expanded table with the difference values).

In your measurements, the reference is not 0 dBm but almost +0.4 dBm. That means, we need to subtract 0.4 dBm from all the subsequent measurements - and in case of the total bandwidth, all we know is that the drop is already -3.4 dB at 587 MHz, but we cannot know what the -3 dB bandwidth is.

I think once that's rectified, your -3 dB bandwidth will be significantly closer to mine


The even more important question is, where does that 0.4 dB error come from?
In my case, I just set the AWG to 0 dBm output (or +10 dBm with the 10 dB attenuator in place) and that's it. I get accurate measurements at 1 MHz and roughly the same at 10 MHz.
Maybe your signal generator is in need of a calibration; in that case you should either adjust the level (to -0.39 dBm) or use the full table with the difference measurements, where you adjust the frequency for the bandwidth measurement such that the difference is -3 dB.

[Martin72]

I think once that's rectified, your -3 dB bandwidth will be significantly closer to mine

I can probably live with that.
But the level thing is more important, I'll have a look at it, I hadn't noticed it before.

[Martin72]

Got it.
I'm posting it here because I figured it out with the help of HD.
I had first reduced the level from 0dBm to -0.55dBm to compensate for the excess gain.
That works, according to the table:



But why is the first marker "so high" and immediately afterwards the level plot is lower?
I was interested in this and resolved the spectrum further, initially up to 50MHz:



Aha....What is this ?
Resolved even further:



This is the problem, for whatever reason the generator is doing this(I will deal with this later).
And after the first 5Mhz it sticks exactly to the 0dBm.
Well, I think it is legitimate to set the first marker not to 1Mhz, but let's say 10Mhz.
My respect for the capabilities of the 2504HD has increased once again.

Martin

[kladit]

Here my contribution to the theme:


The SDG6054X can output frequencies up to 500 MHz maximal, so I can not serve with a -3dB frequency.
The output in sweep-mode seems to be limited to 0dBm,. so I used a 6 dB attenuator (mini ciruits HAT -6db+),
cable is Huber & Suhner Environflex 400, 75 cm long.
The output of the generator has to be corrected to get a reference level of -6 dBm.

[gf]

This is the problem, for whatever reason the generator is doing this(I will deal with this later).

Is it the generator? Can you check with an SA?

If a window function with a wider main lobe is used, then some leakage from DC and negative frequencies is to be expected in the first few FFT bins (maybe up to 1.5 MHz or so), but up 5 MHz seems a bit much to me.

[Martin72]

Can you check with an SA?

Not yet, SA is the next I´ll have in my "Lab".
But I had measured the amplitude with 0dBm set on the generator, at 1Mhz it is higher, at 6Mhz it is correct.

[Performa01]

Here my contribution to the theme:

Even though it doesn't matter in this particular case, you should never, ever under no circumstances use Peak Detect acquisition mode as the source for an FFT.

I had measured the amplitude with 0dBm set on the generator, at 1Mhz it is higher, at 6Mhz it is correct.

Here too: even though it doesn't matter in this particular case, always use "cycle stdev" for amplitude measurements, to get accurate results even when there is a small DC offset and/or the number of periods is not an integer fraction of the screen width.

[Wintel]

Hi,

Can SDS2000X HD be hacked to 500MHz using keygen?

[tautech]

Hi,

Can SDS2000X HD be hacked to 500MHz using keygen?

The public one, no.

[Martin72]

Finally with 6Mhz as the first marker, if you can trust the delta display, the -3db point is 580Mhz.


I find this peak hold function very exciting, every scope should have it.

[gf]

The magnitude of the frequency response reminds me (very roughly) of a 3rd order Butterworth.
-15dB at Nyquist is a bit meager as far as alias rejection is concerned.

[Martin72]

It´s a relative smooth roll-off, yes.
In this context, you could see how the filter works in 350Mhz operation - the scope limits this as soon as more than 2 channels are active.
I'll test it later.

[Martin72]

That was nothing...
With a little thought I could have figured it out beforehand instead of wondering a few minutes ago why the FFT graph abruptly stops at 500Mhz.
The only thing I would wish for morons like me would be a limitation of the adjustable horizontal spectrum as soon as the 350Mhz limit is active.
And then I wanted to save the curve and overlay it with the next recorded one.
I wanted to use the new memory function (M1...M4) for this, but that didn't work or I don't think I understood this function properly, because there are many parameters to set and I actually just wanted to save the curve and call it up again.
I then used the normal REF function for this.

[Performa01]

That was nothing...
With a little thought I could have figured it out beforehand instead of wondering a few minutes ago why the FFT graph abruptly stops at 500Mhz.
The only thing I would wish for morons like me would be a limitation of the adjustable horizontal spectrum as soon as the 350Mhz limit is active.
And then I wanted to save the curve and overlay it with the next recorded one.
I wanted to use the new memory function (M1...M4) for this, but that didn't work or I don't think I understood this function properly, because there are many parameters to set and I actually just wanted to save the curve and call it up again.
I then used the normal REF function for this.

Well yes, it lets you display a 1 GHz wide FFT even in full channel mode, when the Sample rate has droped to 1 GSa/s, but certainly not any further and then again this tiny glitch can be even welcome at times.

Memory traces consist of originally acquired (hence undecimated) data. They can be used as a source for FFT, but cannot store the graphical result of an FFT. For the latter, Ref. traces are the way to go.

[Martin72]

Hopefully siglent will also update the manual, a lot has happened/been added in the meantime.
Here are two more pictures, FFT with 350Mhz limit, FFT without limit.
I think I've played with it enough now, I could look at the same thing again with the SDS2104Xplus, but I don't think you can see anything exciting.
After all, playing around with it brought the little bug with the inverted display to light, so that alone was worth something.

[Performa01]

Another vertical zoom demo

Inspired by a R&S promotional video for the new MXO5

https://www.eevblog.com/forum/testgear/rs-to-launch-mxo5/msg5145714/#msg5145714

I thought I’ll do something similar with the humble little SDS2000X HD…

The SDS2504X HD is only 500 MHz (actually 570) and 2 GSa/s, but for a demonstration like this it doesn’t matter, all the more so as the MXO5 didn’t take advantage on its higher bandwidth at all and only low frequency signals have been used – of course, as otherwise the bandwidth limits couldn’t have worked.

The question remains: what if we need to look at higher frequencies? If we limit the experiment to frequencies below 5 MHz, a PicoScope 6242 would be a much better option, delivering 16 bits from the outset.

I’ll show you something different: a 500 MHz 12-bit DSO should be able to demonstrate a resolution advantage with 500 MHz signals just as well…

Here is the signal mix: a 1 MHz 600 mVpp sine with a 500 MHz 10 mVpp sine riding on it:


SDS2504X_HD_Mix_1M_600mVpp_500M_10mVpp_run

While the MXO 5 uses an odd 14.5 mV/div vertical scale in the zoom window, I’ve chosen straight 10 mV/div. The superimposed waveform is noisy, yet clearly visible.

This is run mode. In stop mode, we can see that most of the noise is just modulation by spikes:


SDS2504X_HD_Mix_1M_600mVpp_500M_10mVpp_stop

In stop mode we get a basically clean waveform with some distortion. Yet this is just 12 bits without any additional tricks.

We cannot limit the bandwidth – for obvious reasons. We can use the Average acquisition mode though:


SDS2504X_HD_Mix_1M_600mVpp_500M_10mVpp_Avg16_run

16 times average is enough to get the waveform pretty clean even in run mode – and that without any speed penalty. The implicit resolution enhancement of this measure is 4 bits, so that the DSO is effectively working with 16 bit data now.

[Andreas]

16 times average is enough to get the waveform pretty clean even in run mode – and that without any speed penalty. The implicit resolution enhancement of this measure is 4 bits, so that the DSO is effectively working with 16 bit data now.

Hello,

Mathematically perhaps. But resolution has to do with noise (= power).
So you need to have a factor 4 to average to gain one bit resolution.
4 Bits resolution enhancement (signal to noise ratio) equal to a average of 256 waveforms.

with best regards

Andreas

[gf]

Mathematically perhaps. But resolution has to do with noise (= power).
So you need to have a factor 4 to average to gain one bit resolution.
4 Bits resolution enhancement (signal to noise ratio) equal to a average of 256 waveforms.

I'm afraid you lost me at the "noise = power" bit. As we are measuring voltages, why would resolution not depend on voltage noise? Could you explain a bit further please?

N-fold averaging has the following effects:

- It reduces the power of uncorrelated noise components by a factor of N.
- It reduces the voltage of uncorrelated noise components by a factor of sqrt(N).
- It enhances the resolution by a factor of N.

But the term "resolution" has nothing to do with noise. It is basically just the the quantization step size, or expressed differently, the number of quantization steps that make up the full scale.

[gf]

the sqrt(N) behaviour of the voltage noise is more intuitive to me, since that's how the standard deviation of a set of N samples behaves.

And for variance, which is stdev^2, it is the factor N. Consider variance as equivalent to noise power.

[ Honestly, for calculating the sample variance, we actually divide by N-1, in order that the sample variance becomes an unbiased estimator for the variance of the distribution from which the samples were drawn. But let's neglect the -1 for a large number of samples. ]

[Performa01]

Yes, I was referring to resolution, not ENOB.

I've mentioned it numerous times that the ERES acquisition mode and math function is related to ENOB, hence 1.0 bit ERES means a theoretical increase of 1 bit ENOB, but 2 bits resolution. This is also how the 10 bit mode in the SDS2000X Plus works: it is just ERES 1.0.

Of course we get even better results with a higher number of averages. The sole reason why I made do with just 16x averaging was because I wanted to demonstrate how even a very moderate averaging can be very effective already.

Since Average qcquisition mode in SDS2000X HD is pretty fast, there is no problem to use up to 128x averaging even when virtually instant real-time updating of the screen picture is required.


SDS2504X_HD_Mix_1M_600mVpp_500M_10mVpp_Avg128_run

[crossan]

I am now the proud owner of an SDS2104X HD and would like to join the club.

I've been a bit busy with work to have a good play yet, but I'm already noticing unexpected benefits. A really nice offset range for one.

I tried the LPF filter Math function on a single shot capture and it seems to me it needs some sort of decimate option. If you have a single capture and need to try an out of range filter you can't change your time base like you would in continuous mode.

Better still would be an ERES decimate parameter.

Anyway it's a fantastic scope, I'm enjoying using it.



 

[tautech]

I am now the proud owner of an SDS2104X HD and would like to join the club.

I tried the LPF filter Math function on a single shot capture and it seems to me it needs some sort of decimate option. If you have a single capture and need to try an out of range filter you can't change your time base like you would in continuous mode.

Have you tried using Fixed Memory mode in the Acquisition menu ?