Sub: Rigol's DHO800 Oscilloscope (Gibbs Effect & Aliasing Misunderstanding)

[nctnico]

https://cdn.teledynelecroy.com/files/whitepapers/wp_interpolation_102203.pdf

The Simple Test described in this paper was also my reconstruction validity criterion when I wrote

But you have to keep in mind that sin x / x is not always implemented correctly. I have come across a few oscilloscopes which had broken sin x / x implementations. An easy check is to feed a sine wave into an oscilloscope with a frequency of 2.5 times less than the samplerate (fsample / 2.5). You should be able to get the sine wave on screen properly. If you get the wrong amplitude or a distorted sine wave, then the sin x / x implementation is incorrect.

[Mechatrommer]

Well, let's assume it comes from the interpolation filter. Why would you design an interpolation filter that cuts off 50% of the recoverable bandwidth? Sure, in practice you have to sacrifice a little, but why so much? Maybe I am missing some information to understand the intention As a result, the interpolation filter does not only reconstruct, but additionally it acts as a bandwidth limiter. And the latter, of course, can still introduce ringing, even if the reconstruction alone at (almost) full bandwidth would possibly not do that.

i'm not sure how internal Lanczos/Sinc filter is designed/coded, possibly cut off more bw to cut cpu processing speed, there's been try and error by the people who know it, for example change Lanczos kernel width or what not, we will have more or lesser gibbs effect. and secondly, we have to take into account the imperfection of sampling machine in real world.. non consistent sampling interval... a little bit of ADC noises etc. if we read theory, they seems to be interesting, given the limited BW and enough parameters, "you should be able to reconstruct waves PERFECTLY!", but i guess in practice there will be catches here and there to save resources. ymmv.

[mr ed]

Anything inbetween the samples is made up data. It's - art. Pick your interpolation. Or rather, it's done for you by the manufacturers. Any digital scope when the sampling gets thin will show some artifacts of some sort. Then the BW limiter adds confusion. Even with it off, the pcb wiring will still have some minor parasitics.

I hate SW options. Hack away!

[Fungus]

Anything inbetween the samples is made up data. It's - art. Pick your interpolation.

Nope. It's correct data, providing the bandwidth conditions are met.

[nctnico]

Anything inbetween the samples is made up data. It's - art. Pick your interpolation.

Nope. It's correct data, providing the bandwidth conditions are met.

Indeed. It is signal reconstruction, not interpolation. Signal reconstruction is only there to make a signal visible for humans (as a visual aid) in case the sample points are too far apart on the screen to see the shape. However, when dealing with steep edges, the sampled data represents a step function (infinite risetime) which can not exist in the real world. So there sin x /x will produce pre and post ringing (Gibbs ears).

[Martin72]

I had read through the whitepaper I linked to and connected the bodnar pulser to the scope again, in this case the DHO804(100Mhz).
The signal from the bodnar is 10Mhz, so it should be interpolated cleanly, regardless of whether linear interpolation or sinx/x is used.
But it doesn't, even at 1.25GSa/s you can already "discover" something and in the worst case, the 312.5MSa/s, it already looks pretty blurred.
This can only have something to do with the very fast rise time of the signal(~40ps).
So I fed in a square wave signal from the SDG2122X with the same frequency and amplitude.
The only difference is that the signal is very "slow" in terms of rise time - 9ns, which shouldn't be a problem for the interpolation, and lo and behold, it is.
(Why the "roof" looks so "wobbly", I'll have another look tomorrow).
Now I also understand the behavior of the DHO4204, thanks for the food for thought@rf-loop.

[Mechatrommer]

3ns sample interval... 0.04ns risetime, how many points? 3ns sample interval... 9ns risetime, how many points? Just for thinking.. too bad dho cannot show dots... harder to think... 2GSa/s also Gibbs

[nctnico]

The only thing that matters where it comes to getting Gibbs ears (as in reconstruction artefacts) is when the signal your are measuring becomes a step response in the sampled aka digital domain. You really need to think about what a sampled signal looks like in the digital domain and how that depends on the samplerate.

[bdunham7]

too bad dho cannot show dots...

Is that true--it does not have a dot mode?

[Martin72]

Yepp, no dot mode and no interpolation choice.
Neither the smallest (DHO800) nor the largest (DHO4000) model.

[wasedadoc]

Anything inbetween the samples is made up data. It's - art. Pick your interpolation.

Nope. It's correct data, providing the bandwidth conditions are met.

It is only correct data if the bandwidth conditions are met and the interpolation is "sufficiently perfect".  In theory, perfect interpolation requires an infinite length filter but in practice it is pointless going beyond a length and arithmetic precision where the ADC quantising error dominates the combined quantising plus interpolation error.

[Fungus]

I didn't say it was perfect, I said it was correct. 

To be "perfect" you need an infinite amount of samples.

[Mechatrommer]

too bad dho cannot show dots...

Is that true--it does not have a dot mode?

yep. maybe someone need to burst tantrum all over rigol sites on this. or else, we just need to get along with this limitation. make sure only work with risetime within nyquist range, or at least need 3 or 4 points sampling for risetime alone. need mental calculation everytime. btw: a "step" in sampling reality got dt is not zero, that means, not perfect vertical, more like diagonal slope, so its real, not only exist in theory, but when doing Sinc interpolation, and ADC noise included af f = sampling rate / 2 and aliasing and all, things can become funny as already presented.

btw, recently probing 150MHz digital lines with DS1054Z, gibbs and ringing everywhere, probably energy content in GHz range? change to 500MHz tektronix 1:10X probe reduced those artifacts and digital signal can be seen clearly, enough for my need, probe with one channel (2 channels 500MSa/s possible) and then save long memory to PC one by one channel, done all 14 channels lane and do post-processing analysis later on, job done. just dont expect doing signal integrity check on 1-2 points per risetetime, that request is unrealistic with $400 scope.. ymmv.

[mawyatt]

We've used 2 samples per rise/fall edge for sample rate minimum, or Sample Rate => (1/tr)*2.

So for something like 100ps rise/fall time this equates to having a sample rate => 20GSPS.

Best,

[gf]

I had read through the whitepaper I linked to and connected the bodnar pulser to the scope again, in this case the DHO804(100Mhz).

Is it really the 100MHz model? Not hacked? There are no measurements enabled, but the risetime seems to be in the order of 1.6ns. This risetime rather suggests that the actual bandwidth is twice as much.

The signal from the bodnar is 10Mhz, so it should be interpolated cleanly, regardless of whether linear interpolation or sinx/x is used.

No, that's the wrong conclusion. It is not the fundamental frequency of the square wave which matters here, but the "smoothness" of the edges and corners.

But it doesn't, even at 1.25GSa/s you can already "discover" something

The superimposed persistent traces from subsequent acquisitions line up quite well and are not "blurred" (ilke in 312MSa.png), which suggests that there is no interpolation/reconstruction issue here. Most likely, the small overshoot is either already present in the original signal or introduced by the analog frontent.

and in the worst case, the 312.5MSa/s, it already looks pretty blurred.

Which is indeed an indication for a reconstruction problem at this bandwidth / sample rate combination.

So I fed in a square wave signal from the SDG2122X with the same frequency and amplitude. The only difference is that the signal is very "slow" in terms of rise time - 9ns, which shouldn't be a problem for the interpolation, and lo and behold, it is.

IMO It is not. Again, the superimposed persistent traces are not "blurred" (ilke in 312MSa.png), which suggests that there is no interpolation/reconstruction issue.

(Why the "roof" looks so "wobbly", I'll have another look tomorrow).

Check the waveform also on a different scope. I could well imagine that the top is not perfectly flat. A square wave is not an easy task for an AWG either, which is based on Nyquist-Shannon reconstruction from samples.

Now I also understand the behavior of the DHO4204, thanks for the food for thought@rf-loop.

I still do not understand it. Your doubt was well justified when you asked why the rise time of the 2Gsa/s trace was shorter than the rise time in the 4Gsa/s trace. This suggests that the front-end was running at a higher bandwidth when capturing the 2GSa/s trace than when capturing the 4GSa/s trace. But this would not make sense.

[Martin72]

Is it really the 100MHz model? Not hacked? There are no measurements enabled, but the risetime seems to be in the order of 1.6ns. This risetime rather suggests that the actual bandwidth is twice as much.

Not bad, not bad...
It´s the 70Mhz model hacked to 100Mhz - But for whatever reason, I had measured 200Mhz (-3dB).
As for the rest, thanks for that, I'll have to look into it further.
I think it's important to know this, or to know it more precisely.

[Fungus]

It´s the 70Mhz model hacked to 100Mhz - But for whatever reason, I had measured 200Mhz (-3dB).

The 100Mhz isn't an officially available  option so who knows what it's doing internally. They're all capable of much more than 200Mhz.

Let's hope Rigol doesn't (or can't) address this issue.

[Martin72]

Let's hope Rigol doesn't (or can't) address this issue.

It´s not a bug, it´s a feature..
Good for us.

[Martin72]

@gf:

IMO It is not.

I had forgotten the "not"...

[gf]

I had read through the whitepaper I linked to and connected the bodnar pulser to the scope again, in this case the DHO804(100Mhz).
The signal from the bodnar is 10Mhz, so it should be interpolated cleanly, regardless of whether linear interpolation or sinx/x is used.
But it doesn't, even at 1.25GSa/s you can already "discover" something and in the worst case, the 312.5MSa/s, it already looks pretty blurred.

I looked through the whole thread and could not find a 625MSa/s screenshot for the DHO800.
All screenshots are either 1.25GSa/s or 312.5MSa/s.
Could you please also post a screenshot of a Bodnar signal taken at 625MSa/s?
Does it also show overshoot, and blur/jitter due to aliasing?

[Martin72]

Hi,

Just made a few pictures...

[Martin72]

Here again a comparison between the original 70Mhz version and now, the "unleashed" version, which has 200Mhz bandwidth (although you can only hack on 100Mhz and only 100Mhz are displayed, maybe the reason why after the last firmware update the bandwidth option is no longer displayed).

[Fungus]

Here again a comparison between the original 70Mhz version and now, the "unleashed" version, which has 200Mhz bandwidth (although you can only hack on 100Mhz and only 100Mhz are displayed, maybe the reason why after the last firmware update the bandwidth option is no longer displayed).

As noted earlier: The "70MHz" version of the DHO800 has 125MHz measured bandwidth.

I think Rigol's marketing department is under-labelling the DHO800 to fool people into buying the DHO900.

(Which makes perfect sense - they probably make 3x profit in return for a $1 edge connector!)

[Martin72]

Just for fun, bodnarpulse on a lecroy WR9054, 500Mhz, 20GSa/s.

[iMo]

Hi,
Just made a few pictures...

You may see with 3ns/point (312.5Msa/s) the sinc function tries to interpolate the points such the "sinc waves" periods fit that 3ns between the points..