Siglent SDS6000A DSO's 500MHz-2GHz

[bdunham7]

ESR in LeCroy oscilloscopes is well explained in attached LeCroy white paper.  But this is LeCroy.

Thanks for the link to the paper.  Meh, I'm sort of underwhelmed by the concept and whatever benefits the system may have for measurement and analysis, marketing it as an equivalent to a higher sample rate is dubious, IMO.  And here they are comparing 2.5GSa/s rather than 5GSa/s to their '10G estimated points/s.  Is that really better?

[rf-loop]

The LeCroy white paper reference by 2N3055 is a must read for anyone using any of these mentioned instruments.

http://cdn.teledynelecroy.com/files/whitepapers/enhanced-sample-rate-whitepaper.pdf

Edit: A quick search shows TI has candidate 12 bit ADCs, as does AD, and they aren't cheap at ~$1K per chip.

Best,

Just same what was attached pdf in my previous message. 

[rf-loop]

Why you name it as Siglent ESR
@bdunham7 corrected it. Now OK 

[nctnico]

ESR in LeCroy oscilloscopes is well explained in attached LeCroy white paper.  But this is LeCroy.

Thanks for the link to the paper.  Meh, I'm sort of underwhelmed by the concept and whatever benefits the system may have for measurement and analysis, marketing it as an equivalent to a higher sample rate is dubious, IMO.  And here they are comparing 2.5GSa/s rather than 5GSa/s to their '10G estimated points/s.  Is that really better?

Well, it helps to improve the visualisation of a signal. That is what interpolation on a DSO is all about; showing an image which is as close to the actual waveform shape as possible. Then again, how much of that pre-edge ringing is the actual signal and how much is due to the Gibbs effect?

[bdunham7]

Well, it helps to improve the visualisation of a signal.

Yes, but like sharpening a photo, there's a limit to the amount of improving you can do before you are distorting.  In this case, I'm not seeing the 'improvement', at least visually.

[bdunham7]

Why you name it as Siglent ESR (your original msg attached image name Siglent ESR.png) . Where you have seen Siglent do this compare or presents the matter with that image. These are from Teledyne LeCroy.
Siglent sell outside China only 8 bit SDS6000A.

I'm underwhelmed by the concept and I dislike the marketing of it, no matter who presents it.  However, in the interests of accuracy and just for you, I've changed the title of the photo.  Now is Siglent's 10GSa/s 'ESR' on the 8-bit SDS6000A scopes they sell outside China significantly different in either operation or the overall fact that the extra samples are interpolated?

[Performa01]

While ESR doesn’t increase the effective sample rate, hence has no effect on the Nyquist frequency, it generates more data for screen rendering and automatic measurements. As always, there are pros and cons to this approach:

+ improves the accuracy of automatic measurements in many cases
+ can improve the visual representation of a waveform in some (rare) cases
+ works silently in the background and we get the benefits without firmware change

– uses up twice as much memory for data that could be generated at any time later in the process
– may result in ambiguous waveform rendering even in dots mode when Nyquist has been violated

The best feature of ESR is that it can be turned off.

The attached screenshots demonstrate the improved accuracy of transition time measurements in a long record.

The main window contains a 5 µs long capture of a ~160 MHz square wave with approx.. 210 ps rise time. The zoom window shows the details at 2 ns/div. We measure the transition times in both the main and the zoom window – and these are different for a reason:

In the main window, a 25 kpts long record provides more data than what is required for displaying the waveform, hence no interpolation required. Consequently, the resolution of time measurements is limited by the sample rate: 5 GSa/s -> 200 ps sample interval. This means we have to live with an average error of 100 ps.

In the 20 ns wide zoom window, there are only 100 samples, hence interpolation is required for the measurements and we get results much closer to the truth for comparison purposes.

First a standard measurement without ESR. The zoom window measurements hint on a rise time (signal source & scope combined) of some 310 ps. The measurement on the long record reads some 100 ps higher, just as predicted (always look at the mean values in the measurement statistics!).

Second is the same measurement, but with ESR enabled. The error of the long record measurement is now only some 50 ps. ESR does of course not affect the zoom measurement.

[nctnico]

It is an interesting technique and I think it is valid for improving measurements. With sin x/x reconstruction (IMHO interpolation is the wrong term) the original sample points are kept but a more true shape of the waveform is created. My guess is that measurements normally use linear interpolation which is highly affected by the position of the sample points and thus resulting in extra jitter in the measurement result.

[tv84]

With sin x/x reconstruction (IMHO interpolation is the wrong term) the original sample points are kept but a more true shape of the waveform is created.

It's a reconstruction based on interpolation. So using "interpolation" is perfectly correct (and shortens the text).

[nctnico]

With sin x/x reconstruction (IMHO interpolation is the wrong term) the original sample points are kept but a more true shape of the waveform is created.

It's a reconstruction based on interpolation. So using "interpolation" is perfectly correct (and shortens the text).

There is a little bit more to it. In measurement land interpolation is a very dirty word because interpolation means estimating values that lie between points. So it is a word to avoid that if you can when talking about measurements & equipment because it will turn certain people off. Where it comes to DSOs, sin x/x does not interpolate but reconstructs so NOT calling it interpolation is valid (note how the ESR paper from Lecroy puts so much emphasis on meeting the Nyquist criteria to make sin x/x valid).

[tv84]

sin x/x does not interpolate

  I must have skipped some lessons...

[nctnico]

sin x/x does not interpolate

  I must have skipped some lessons...

Indeed you have  . When the Nyquist criteria is met (with some margin) you can prove mathematically that you can draw only 1 waveform through the sampled points which is what sin x/x does. This means you are not interpolating (estimating) but reconstructing.

[2N3055]

Yes we are reconstructing by using sin(x)/x to interpolate parts of the curve (or points, in the end there will always be physical plot points on the screen at least) between..

Estimation is not a synonym for interpolation.

Interpolation is a specific type of value estimation, that is closely related  to approximation.
Hence we approximate missing values by using sin(x)/x function, that we curve fit to sampled data points.
Therefore reconstructing the data between sample points.  We say that data was interpolated.

Or I don't understand it right...?

[nctnico]

I don't want to drag out this semantic discussion out any further so this is the last thing I'm going to write about it: As I explained before you have to be really careful with the word 'interpolation' because it has a negative tone in combination with using test equipment. And there is reason for it since interpolation means estimation and thus implies inaccuracies / fantasy data. This has become very clear to me in previous discussions so I avoid using the word unless it is actually appropriate.

[bdunham7]

The best feature of ESR is that it can be turned off.

Hardly a ringing endorsement...

The attached screenshots demonstrate the improved accuracy of transition time measurements in a long record.

I certainly don't want to dismiss the efforts of some very talented people at LeCroy, as I'm sure this 'feature' is of some use and I certainly don't know enough technically to pass judgement on it.  It's the marketing angle where I'm sure the sales department was giddy with excitement being able to write "10Gsa/s blah blah".  That's an area I know enough to pass judgement on, b/t/w.

As for what it does, according to your explanation, it appears that without ESR the scope only performs the sinx/x interpolation on the screen display (and to adjust the trigger point on the fly) and not the whole capture?  So then ESR is simply using sinx/x to generate points in between the actual captured points so that the measurements can be done in the way they normally are--not using sinx/x interpolation--on a greater number of points.  OK.

However, I can't quite reconcile that with the photo excerpt that I posted from LeCroy's paper.  More samples means more Gibbs ears?  As I commented there, I fail to see the improvement in that case, and since the screen display is already being processed by sinx/x, what are they actually doing?

[Performa01]

The best feature of ESR is that it can be turned off.

Hardly a ringing endorsement...

It has to be the best feature, if only because I have requested it.
And the reason was … see the cons in my previous posting.

I certainly don't want to dismiss the efforts of some very talented people at LeCroy, as I'm sure this 'feature' is of some use and I certainly don't know enough technically to pass judgement on it.  It's the marketing angle where I'm sure the sales department was giddy with excitement being able to write "10Gsa/s blah blah".  That's an area I know enough to pass judgement on, b/t/w.

I don’t know exactly how LeCroy advertise their gear, but I do know what Siglent’s strategy outside of China is.

Have a look at the international SDS6000A instruments – do you see 10 GSa/s printed on them anywhere?
Have a look at the company webpage. There it says “5 GSa/s (10 GSa/s ESR) per channel” over and over again, so clearly hinting on a genuine samplerate of 5 GSa/s. It is also clearly stated what ESR is – some form of 2x interpolation. So nobody gets fooled.
Have a look at the datasheet. It’s the same as stated above.

Sorry, I cannot see what’s wrong with the marketing. Should the ESR feature be concealed?

As for what it does, according to your explanation, it appears that without ESR the scope only performs the sinx/x interpolation on the screen display (and to adjust the trigger point on the fly) and not the whole capture?  So then ESR is simply using sinx/x to generate points in between the actual captured points so that the measurements can be done in the way they normally are--not using sinx/x interpolation--on a greater number of points.  OK.

ESR has nothing to do with the general post processing strategies, except that it doubles the number of data points regardless of anything else.

It is only logical that there will neither be linear interpolation “x” or a sin(x)/x reconstruction if there are already enough samples to fill the screen.

So for the display, there will be a 1:1 sample mapping at a certain timebase, e.g. 20 ns/div.
Interpolation or reconstruction is used at faster timebases – where the number of samples gets less then the horizontal screen pixels.
Decimation or agglomeration is used at slower timebases.

You seem to suggest that this is somehow awkward, suboptimal or at least unusual? Just think about it:

In an extreme situation, you use the fastest timebase, which is 100 ps/div. At a sample rate of 5 GSa/s, you only get one sample per two divisions, that is five samples total. That means the interpolation or reconstruction has to provide some 1200 additional values, i.e. multiplying the initial amount of samples by 240. So there are situation, where you need to create that many additional data. Now what does that mean for the memory consumption?

Even ESR, which only doubles the number of samples already requires twice the memory for that. You probably say, for enhancing measurements on long records we do not need to multiply the data by 240 – what do we use then? 1 ps resolution requires a 200 fold expansion of the sample data – our scope then has to have either 200 times the memory (= 200 Gpts) or we leave it as it is and can only advertise it as “2.5 Mpts memory length.”

Even if we make do with just 10 ps, then it is still a factor of 20.

So it’s easy to criticize, but you should also consider the consequences.

However, I can't quite reconcile that with the photo excerpt that I posted from LeCroy's paper.  More samples means more Gibbs ears?  As I commented there, I fail to see the improvement in that case, and since the screen display is already being processed by sinx/x, what are they actually doing?

Yes, I’m with you here.

As my screenshots demonstrate, there is no difference in the Y-t view – at least not for a rather benign signal with ~200 ps rise time. I do not have a 30-40 ps signal source, maybe I should get one and see what happens then (most likely nothing unexpected).

Anyway, you can be sure that you’ll hardly ever see a whitepaper with questionable screenshots, that create more questions than answers, from Siglent.

[bdunham7]

Sorry, I cannot see what’s wrong with the marketing. Should the ESR feature be concealed?
So it’s easy to criticize, but you should also consider the consequences.

My only 'criticism' re Siglent is of the marketing, and IMO the way it is stated makes it too easy to confuse with ETS.  I would not have listed it as a headline feature like that. 

As for the technical issues, the I understand the tradeoffs in making measurements.  You could generate a huge number of interpolated points, which would use too much memory, or (in some cases--like pulse width or rise time) just do the interpolation on-the-fly locally where it is needed for each measurement, just like is done with the trigger interpolation.  That would probably use too much processor power and I have no good ideas on how to actually implement that, so there are no technical criticisms from me regarding this feature. 

And I'm just curious--why stop at 10GSa/s?  If you have a short capture, why not interpolate away until the memory is fully used, then do measurements?  You'd be able to measure pulsewidth as nicely as the trigger interpolation is able to lay successive traces right on top of one another.

[Performa01]

My only 'criticism' re Siglent is of the marketing, and IMO the way it is stated makes it too easy to confuse with ETS.  I would not have listed it as a headline feature like that. 

Well, it is the proper description of the acquisition system, listed together with other key features. ESR in parenthesis, nothing that sticks out in any way, nothing like a bigger or bolder font.

As for the possible confusion of ESR with ETS … if we cannot rely on EEs to be able to properly understand what they read, then we get some more problems indeed. ETS went largely out of fashion anyway.

Overall, we could criticize pretty much all manufacturers, because there will always be some room for confusion and/or unjustified expectations from the marketing material. At least I firmly believe that anyone can get honest and accurate information from the Siglent material, as long as they read carefully and are capable to understand what they’ve just read. I’m aware that we cannot generally presume that though…

As for the technical issues, the I understand the tradeoffs in making measurements.  You could generate a huge number of interpolated points, which would use too much memory, or (in some cases--like pulse width or rise time) just do the interpolation on-the-fly locally where it is needed for each measurement, just like is done with the trigger interpolation.  That would probably use too much processor power and I have no good ideas on how to actually implement that, so there are no technical criticisms from me regarding this feature. 

Yes – and I can give you some good reasons why this is not really feasible…

Have you had a look at the screenshots in my earlier posting? Look at the counter in the measurement statistics for rise and fall times. You will notice that the count is vastly different in the main and zoom window. You might conclude that every single transition is analyzed separately, and a single record can provide a very high number of measurements.

So, interpolating all the transition regions in a record is certainly a lot more effort than fine-adjusting the (single) trigger point. Just for fun, let’s have a look at my example:

800 full cycles of a 160 MHz signal fit into a 5 µs long record. Each cycle has two transitions, so there is a total of 1600 transitions in a single record. But that’s not the only problem by far. Now we need to find the 10% and 90% points for our measurements. Where are they located? We do not know until we have a valid measurement result – that’s where the cat bites her own tail. We would have to analyze the entire region around the transition. How much? Well, it depends on the transition time. In theory the signal could have slow transitions, up to the point where it turns into a triangle, where we need to look at almost the entire signal period.

Long story short: We can either do it in advance, with a rather stupid algorithm, i.e. for the entire record and with massive hardware support, so that it doesn’t cost any additional time, or we do it as part of the software based measurements, and slow down the frame rate to an extent that everyone will complain very loudly.

And I'm just curious--why stop at 10GSa/s?  If you have a short capture, why not interpolate away until the memory is fully used, then do measurements?  You'd be able to measure pulsewidth as nicely as the trigger interpolation is able to lay successive traces right on top of one another.

A good question.

Remember that ESR usually is just there, nothing configurable at all. As mentioned earlier, it was not even switchable originally.

The main reason will be the limited bandwidth of the data bus and the memory. You cannot push the sample rate up to the sky, because you need to be able to transfer that amount of data into the acquisition memory in real time. It really doesn’t pay off to fit the bus hardware and sample memory required for a high-end system with 100 GSa/s or more, just to have some interpolation that only produces more or less redundant data from a low end 5 GSa/s ADC in the end.

There is a major difference to ETS, where you take a number of consecutive records to assemble the final representation of the input signal. The bandwidth requirements here remain the same as for standard real time sampling. It just takes a number of acquisitions until a complete ETS record is built. That’s why ETS is only applicable for static signals.

[edigi]

There is too much marketing blah-blah surrounding the ESR/enhanced sample rate term (and most engineers have little time to read it) so I'm trying to guess what's it really about (please correct or confirm).
Unfortunately the core of the very good description from bdunham7 touching the same essence was not really discussed.

The sinc interpolation is nothing new, in fact every modern DSO I've seen has it as an option and very useful when the timebase is so low that there as significantly more display points than samples available (for example 200ps with 5GSa/s).
For a small number of samples the micro (ARM or whatever) has sufficient power to do this (but can be done also in FPGA). This is about the visualisation of the signal and it's not about doubling the sample points but getting a "sample point" for each display pixel.

Things however get radically different if we start to talk about measurements that's supposed to operate on large number of samples (so even when the timebase is not so low) and that needs the "sinc interpolation" done in the FPGA. Due to the massively more data this means not getting hugely more virtual sample points, but some kind measurement result based on interpolated (although this may not be the best term here) data.
What is achieved with this is that the measurement accuracy gets to a comparable level with it with a 5GSa/s DSO as without it with a 10GSa/s DSO.

So for visualization it has negligible impact (in fact the 10GSa/s DSO has an edge here due to the more physical samples) and it's mainly about measurement?

Am I way off with my thinking, or kind of OK just missed something (I mean in the essence of the story I know that probably there are way more in the dirty details)?
 

[Performa01]

Unfortunately the core of the very good description from bdunham7 touching the same essence was not really discussed.

Oh? have I missed something?

Things however get radically different if we start to talk about measurements that's supposed to operate on large number of samples (so even when the timebase is not so low) and that needs the "sinc interpolation" done in the FPGA. Due to the massively more data this means not getting hugely more virtual sample points, but some kind measurement result based on interpolated (although this may not be the best term here) data.
What is achieved with this is that the measurement accuracy gets to a comparable level with it with a 5GSa/s DSO as without it with a 10GSa/s DSO.

I'm not quite sure if I really understand your question. ESR is a technique applied very early in the digital signal processing chain. We can probably compare it with the acquisition modes Avg and ERES which process and substitute the data on their way from the ADC to the sample memory.
ESR is special in that it doubles the amount of data, so all the subsequent processing cannot tell the difference from a true ADC with twice the sample rate.
Because it has to be done in real time, all these acquisition modes need to be accomplished by hardware.

ESR gives twice the amount of data and doubles the sample rate, hence the resolution. Measurements gain accuracy, not only time measurements, but also amplitude, because peaks can be located more precisely which in turn leads to a better estimation of the peak amplitude for instance.

Measurements use the sample memory on records with more than some 1200 samples (more than fits on the screen). So the resolution of the data there is proportional to the sample rate. For shorter records, measurements use some secondary buffer with the sin(x)/x reconstruction of the waveform in it.

So for visualization it has negligible impact (in fact the 10GSa/s DSO has an edge here due to the more physical samples) and it's mainly about measurement?

Am I way off with my thinking, or kind of OK just missed something (I mean in the essence of the story I know that probably there are way more in the dirty details)?

For visualization, it should have very little to no impact, because for longer records we don't need any additional samples for good visualization, whereas for short records it should make no difference whether we have a two-stage process in applying sin(x)/x with only doubling the number of samples first and then again in the display buffer to do the rest - or if everything is accomplished in a single run in the display buffer just once.
 

[edigi]

OK, clear explanation, thanks. I think we're on the same page now.

[simone.pignatti]

Hello the, we just got the unit, I like to share a nice picture of it, by the way the front cover comes as standard with the oscilloscope 

[hpw]

Hello the, we just got the unit, I like to share a nice picture of it, by the way the front cover comes as standard with the oscilloscope 

So please report to us, what probes you may use and any nice fast pulse measurements (as Bodnar fast pulse 40..50ps) to show...

Hp

[hpw]

Hello the, we just got the unit, I like to share a nice picture of it, by the way the front cover comes as standard with the oscilloscope 

So please report to us, what probes you may use and any nice fast pulse measurements (as Bodnar fast pulse 40..50ps) to show...

Hp

So no beef seen so far (no tear down or test reports), all waiting for the SDS6000 Pro for any better bits... or are those all out of turtles budged 
   

[maxwelllls]

Some folks here already noticed that the price difference between the various bandwidth models is rather low - so this is a first pointer that
a) the hardware is identical for all models - other than for the SDS5000X
b) this is clearly aimed for the professional market and there is no particularly cheap entry model for all the hackers out there.

c) This is the first Siglent model with "licensing hack" by design.

Are there any threads for SDS6000 hacking? I recently got an SDS6000 PRO (Chinese version)