Scope Wars

[David Hess]

Even ignoring design flaws, the Rigol DS1000Z series would come up last in my list only because Rigol is deliberately misleading with their documentation and design.  Being inexpensive has nothing to do with the manufacturer deliberately deceiving its customers.

Would you please document those assertions?  I'd like to prove or disprove them and compare the performance metric to other OEM products.

I have posted about them multiple times in the past.  Off the top of my head:

1. Confusing peak detection with envelope detection - the former requires hardware support while the later can be done in software.
2. Confusing dual timebase operation with magnification - if these were the same thing, then how could other DSOs with actual dual timebase support have both?  There was also something here about confusing delayed trigger as applied to dual delayed timebase operation with a qualified delayed trigger which is completely unrelated.
3. Ground coupling which is implemented late in the amplifier chain - this makes the instrument seem to have less input noise than it really has.  This would not be such a big deal if they did not take advantage of it in their marketing materials touting low noise which was anything but; it was low noise because they were not amplifying the noise.  This might be responsible for the inability to null small offsets which users have noticed; the instrument can't see them because ground coupling attenuates them.
4. Lack of sufficient full power bandwidth to support the transition time or specified bandwidth - this is why users report wildly different bandwidths for the same model of instrument.  It also shows up as a non-linearity from 2 to 10 nanoseconds because of overload or cutoff after a fast edge at certain time/div settings.  I suspect this comes about because the preamplifier stage operates over 10 times the voltage range with the same transistors that other oscilloscopes manage and for good reason.  Tektronix for instance used transistors which were 3 times faster with an input range which was 1/20th as large.

[StillTrying]

As you're such a fan of Rigol, I hope you didn't miss the Capactor Charging In Steps thread.
https://www.eevblog.com/forum/chat/does-a-capacitor-charges-smooth-or-in-stairs/?action=dlattach;attach=1006164;image
https://www.eevblog.com/forum/chat/does-a-capacitor-charges-smooth-or-in-stairs/msg3103617/#msg3103617

[SilverSolder]

I don't know the details of this instrument, but HPAK introduced sample dithering which phase shifted the clock from sweep to sweep.  If you divide the actual sample rate by 20 and shift the phase of the clock by that amount on each sweep, then you get an effective sample rate on a repetitive waveform which is 20 times greater.  For spectrum analysis which has traditionally been done by sweeping the LO, that is a very viable approach and will produce the same answer.
[...]

That is very clever.  I found an HP Journal article that explains how their random sample dithering works to dramatically reduce aliasing artifacts (attached for reference), when doing equivalent time sampling.

Presumably this scope is still limited by the analog front end, which is why we don't see anything past 200MHz even if the sampling circuitry is capable of it (FFT mode).

It looks to me like the FFT function in this scope is exactly the same as in its bigger brother models that have 1Gsample/second sample rates and more analog bandwidth.  So, in the lower range models, the FFT is better than the input circuitry is actually able to supply...

[SilverSolder]

Two tests using the FFT in an HP 54542A oscilloscope. 1 MHz square wave in each case, using an HP 3310A Function Generator and an HP 8011A Pulse Generator.

This looks like superb performance.

What is the difference between the first and the second test?

[CharlotteSwiss]

I missed the developments, will my sds1202 also be tested, or has the 4 channels been preferred?
 

[rhb]

I still guess that it might be the same reason as on mine: The FFT is calulated from the samples in the display buffer which are interpolated and re-samped at a higher rate in order to fill the gaps between the original sampling points.

You may be right -  the FFT number of points is 2048, and the display is 1024...   so by doubling up the number of points and interpolating, the "new Nyquist" would now become 200MHz, which is what we can see in the screen grab.   

But...   how can there be information in the signal beyond the 100MHz bandwidth of the analog front end, even if you interpolate in the screen buffer you are not really adding signal information to what was there in the first place - right?

Perhaps they are actually doing wizardry along the lines that @Reg suggests above.

If the ADC is triggered on the rising edge of the clock, just flipping the phase of the clock would collect the samples between the samples.  That could easily be done with a NAND gate on the clock line.   It can't be done by interpolation  because you don't have the information.

Reg

[Tomorokoshi]

Two tests using the FFT in an HP 54542A oscilloscope. 1 MHz square wave in each case, using an HP 3310A Function Generator and an HP 8011A Pulse Generator.

This looks like superb performance.

What is the difference between the first and the second test?

The HP 3310A function generator happens to have a square wave setting, while the 8011A is purposefully built as a pulse generator.

Setup was done using the HP 3310A. Lots of tweaking of the FFT setup. Aliasing was very easy to encounter even when careful about the setup.  I calculated the FFT coefficients to check against the ideal square wave spectrum. I normalized the calculations at 1 MHz. Compared to the 8011A the 3310A drops off more quickly, so by 25 MHz it's down by 8 dB. The 8011A is off by only 2.5 dB at 25 MHz. However, the 8810A has higher power in the even harmonics.

[SilverSolder]

Two tests using the FFT in an HP 54542A oscilloscope. 1 MHz square wave in each case, using an HP 3310A Function Generator and an HP 8011A Pulse Generator.

This looks like superb performance.

What is the difference between the first and the second test?

The HP 3310A function generator happens to have a square wave setting, while the 8011A is purposefully built as a pulse generator.

Setup was done using the HP 3310A. Lots of tweaking of the FFT setup. Aliasing was very easy to encounter even when careful about the setup.  I calculated the FFT coefficients to check against the ideal square wave spectrum. I normalized the calculations at 1 MHz. Compared to the 8011A the 3310A drops off more quickly, so by 25 MHz it's down by 8 dB. The 8011A is off by only 2.5 dB at 25 MHz. However, the 8810A has higher power in the even harmonics.

Ah, I misunderstood, I thought you were using the pulse generator on both.  - Very cool!   

[rhb]

To keep things simple, one thing at a time:

1. Confusing peak detection with envelope detection - the former requires hardware support while the later can be done in software.

I'm not sure I understand what you mean by this.  One does not have to sample the peak to determine the magnitude or when it occurs.

The input here is the 100 ps 10 MHz repetition pulse.  The scope samples are 1000 ps apart. It's quite clear that the band limited pulse is well captured in dot mode and shows a ~ 1 ns rise time.

So why are these supposed to be different and how?

I should like to note that the small dip in most of the displays just prior to the rise of the pulse is real.  The pulse actually starts there, but there is phase dispersion and delay in the analog circuitry.  I'll demonstrate that later using a 20 GHz SD-26 sampling head which can more accurately display a 100 ps pulse.

 The ringing before the step in average mode is the result of using a zero phase sinc interpolator.  The correct interpolator is the minimum phase impulse response of the AFE which is precisely what is shown using persistence and dot mode.

Watching the live display, the distribution of the dots varies with time.  I suspect it is simply the result of clock jitter rather than a design feature.  However, dot mode and persistence let you get a very good representation of the pulse.

Have Fun!
Reg

[rhb]

Here's the same setup on the GW Instek MSO-2204EA.  Interestingly we get a clear picture of the waveform at 5 ns/div, but not at 2 ns/div

Reg

Edit:  Sigh.  the server has borked itself and the 5 ns/div Instek image has been replaced by the Owon image in the next post by me.

[rf-loop]

Here's the same setup on the GW Instek MSO-2204EA.  Interestingly we get a clear picture of the waveform at 5 ns/div, but not at 2 ns/div

Reg

Look this Rigol trigged edge time position. Quite ok.

Look then this Good Will trigged edge position. Totally out of order.  Is this GoodWill model at all with digital trigger engine..  In 5ns/din display scale position is 2.5ns wrong when sampling interval is 1ns. 2ns/div time scale just "game over"...

[rhb]

I don't recall the Instek behaving this way before, but it gets a little difficult keeping track of what does what with so many scopes around.  I plan to reexamine the Instek result and inquire with Good Will about it. 

The shift in the trigger point is because I was triggering off the sampling scope trigger signal produced by the pulser to see if that would help the 2 ns/div display and forgot to switch back to triggering on channel 1 when I made the screen dumps.

Here are the Owon XDS-2102A and Rigol DS1102E with the same setup.


BTW My previous FFT result for the DS1202Z-E was dominated by user error.  I'll be redoing that soon. 

Have Fun!
Reg

[nctnico]

Look then this Good Will trigged edge position. Totally out of order.  Is this GoodWill model at all with digital trigger engine..  In 5ns/din display scale position is 2.5ns wrong when sampling interval is 1ns. 2ns/div time scale just "game over"...

If you look more closely you'll see the actual trigger point is off-screen. What you see is the next pulse. I don't know what rhb did in the second image. I can't reproduce it on my unit.

[rhb]

Look then this Good Will trigged edge position. Totally out of order.  Is this GoodWill model at all with digital trigger engine..  In 5ns/din display scale position is 2.5ns wrong when sampling interval is 1ns. 2ns/div time scale just "game over"...

If you look more closely you'll see the actual trigger point is off-screen. What you see is the next pulse. I don't know what rhb did in the second image. I can't reproduce it on my unit.

Reg doesn't know either.  Rather reminiscent of my 11801 which from time to time gets into a weird state and starts producing bogus events until I do an "autoset".  Fortunately, I've learned to recognize the artifacts.

So here's the 5 ns/div display that got trashed by Dave's server and a 1 ns/div after doing an "autoset" and readjusting the time base.  I don't think I was able to get faster than 2 ns/div when I made the other displays.

Have Fun!
Reg

[rhb]

Here's  an FFT from the DS1202Z-E with sensible parameters.  Sample length is 12K samples and I'm using a triangle aka Bartlett window in the time domain which is a sinc(f)**2 in frequency giving a sensible looking display of a fast spike train.

Turning anti-aliasing on or off seems to have no effect.  I've gotten a FW update, but have not installed it yet.

I have discovered that if I stop the acquisition I can save a PNG file *much* faster.

Have Fun!
Reg

[SilverSolder]

So, if the DS1202Z-E is a 200MHz scope,  how is it able to display any frequency content at 500MHz with any accuracy in its FFT function?

Doesn't the analog front end limit the signal before the FFT code even sees it?

[David Hess]

To keep things simple, one thing at a time:

1. Confusing peak detection with envelope detection - the former requires hardware support while the later can be done in software.

I'm not sure I understand what you mean by this.  One does not have to sample the peak to determine the magnitude or when it occurs.

That doesn't refer to the DS1000Z series which supports peak detection, which was why the DS1000Z series was the first of their DSOs to draw my serious consideration.  The previous Rigol series, which the DS1000Z replaced, lacked the hardware to support peak detection which was a common feature by then (1) so Rigol confused the issue by conflating envelope detection with peak detection so they could claim to support it.

The input here is the 100 ps 10 MHz repetition pulse.  The scope samples are 1000 ps apart. It's quite clear that the band limited pulse is well captured in dot mode and shows a ~ 1 ns rise time.

So why are these supposed to be different and how?

Peak detection applies when the maximum sample rate is not available, usually because of a limited record length.  It is also not exclusively for peaks; it also prevents aliasing due to a lower than maximum sample rate.

So use a shorter record length or a slower time/div to see its significance.

(1) Peak detection has been around since the late 1980s.

[rhb]

So, if the DS1202Z-E is a 200MHz scope,  how is it able to display any frequency content at 500MHz with any accuracy in its FFT function?

Doesn't the analog front end limit the signal before the FFT code even sees it?

That's a *very* good question.  Just what *is* going on?  I just did an update that is supposed to fix the CSV save bug.  We shall see.  Once I can get a CSV file I can do an FFT and see what the actual AFE response is.  I was able to save a 120 Kpts file, but it's been sitting for about 5 minutes saving a 24 Mpts file.

Have Fun!
Reg

[Andie]

Some DSOs do not allow sin(x)/x reconstruction to be disabled and even if they do, sin(x)/x reconstruction is still present for the digital trigger and *that* is what smears the results; the trigger point is changing depending on the relationship between the sample trigger and input edge.  DSOs with analog triggers do not suffer from this problem and sampled points align correctly.  If they did not, then equivalent time sampling would not work.

Trigger is indeed a good point. When I think about it, any sub-sample resolution triggering requires interpolation, (a) in order to determin the trigger point location between two adjacent sample points (in case of a digital trigger - as you say), and (b) in order to time-shift the captured samples by a fraction of the sampling interval. Even an analog trigger still requires (b), since the pre-trigger samples are already captured (with an arbitrary ADC clock pahse) when the trigger fires, and I guess even for post-trigger samples it were hard to enforce a different ADC clock phase momentarily, if I assume a PLL generated clock. On the other hand, if sub-sample resolution triggering is renounced, then the trigger point gets time-quantized, leading to trigger-jitter of +- 1/2 sampling interval (for single-shots, this does not matter of course, as subsequent frames don't neet to line-up on the time axis).

HP describe how they did that in the HP54100A:

https://www.hpl.hp.com/hpjournal/pdfs/IssuePDFs/1986-04.pdf

Page 7 "Interpolator"

[rhb]

Peak detection applies when the maximum sample rate is not available, usually because of a limited record length.  It is also not exclusively for peaks; it also prevents aliasing due to a lower than maximum sample rate.

So use a shorter record length or a slower time/div to see its significance.

(1) Peak detection has been around since the late 1980s.

So "peak detection" substitutes the peak value during the sample period for the value at the time of the sample clock?  If so, that has some very *interesting* effects.   None of which seems to me desirable.

If that is a correct interpretation I cannot understand why anyone would want to do it except for lack of comprehension of sampling theory.  Aliasing is a consequence of regular sampling, so there is *no* possibility of "peak detection" preventing aliasing.  This is dead obvious if you simply draw the time domain multiplication and the corresponding frequency domain convolution.

Seismic data is often acquired at not quite regular spatial sampling and there have been hundreds of professional papers written on how to mitigate the effects of that on the data as it causes a lot of problems in seismic imaging.

Have Fun!
Reg

[SilverSolder]

Peak detection applies when the maximum sample rate is not available, usually because of a limited record length.  It is also not exclusively for peaks; it also prevents aliasing due to a lower than maximum sample rate.

So use a shorter record length or a slower time/div to see its significance.

(1) Peak detection has been around since the late 1980s.

So "peak detection" substitutes the peak value during the sample period for the value at the time of the sample clock?  If so, that has some very *interesting* effects.   None of which seems to me desirable.

If that is a correct interpretation I cannot understand why anyone would want to do it except for lack of comprehension of sampling theory.  Aliasing is a consequence of regular sampling, so there is *no* possibility of "peak detection" preventing aliasing.  This is dead obvious if you simply draw the time domain multiplication and the corresponding frequency domain convolution.

Seismic data is often acquired at not quite regular spatial sampling and there have been hundreds of professional papers written on how to mitigate the effects of that on the data as it causes a lot of problems in seismic imaging.

Have Fun!
Reg

The reason would be to avoid missing a short pulse on a slow sweep.

[Andie]

Peak detection applies when the maximum sample rate is not available, usually because of a limited record length.  It is also not exclusively for peaks; it also prevents aliasing due to a lower than maximum sample rate.

So use a shorter record length or a slower time/div to see its significance.

(1) Peak detection has been around since the late 1980s.

So "peak detection" substitutes the peak value during the sample period for the value at the time of the sample clock?  If so, that has some very *interesting* effects.   None of which seems to me desirable.

If that is a correct interpretation I cannot understand why anyone would want to do it except for lack of comprehension of sampling theory.  Aliasing is a consequence of regular sampling, so there is *no* possibility of "peak detection" preventing aliasing. [...]

I think a lot of people use an oscilloscope for troubleshooting and finding components of signals which they didn't expect in the first place. In this case an oscilloscope is a tool to gather as much information as possible. A peak detector can capture transients which would otherwise be unnotable.

To me, an oscilloscope, which performs antialiasing or lacks a peak detector (or equivalent features like high samplerate and lots of memory), would be useless.

Andreas

[nctnico]

Peak detection applies when the maximum sample rate is not available, usually because of a limited record length.  It is also not exclusively for peaks; it also prevents aliasing due to a lower than maximum sample rate.

So use a shorter record length or a slower time/div to see its significance.

(1) Peak detection has been around since the late 1980s.

So "peak detection" substitutes the peak value during the sample period for the value at the time of the sample clock?  If so, that has some very *interesting* effects.   None of which seems to me desirable.

If that is a correct interpretation I cannot understand why anyone would want to do it except for lack of comprehension of sampling theory.  Aliasing is a consequence of regular sampling, so there is *no* possibility of "peak detection" preventing aliasing.  This is dead obvious if you simply draw the time domain multiplication and the corresponding frequency domain convolution.

Well, you have to understand that an oscilloscope's purpose is to look at a signal and not to acquire a signal for processing later on. Without peak-detect you can get all kinds of weird looking signals / miss parts of a signal when the samplerate is too low. I regard peak detect as an essential feature on a DSO. Try to check a 1PPS signal from a time reference for example. The pulse might be too narrow compared to the samplerate available at time/div settings of 500ms/div.

[Fungus]

If that is a correct interpretation I cannot understand why anyone would want to do it except for lack of comprehension of sampling theory.

They do it when they don't want to display the curve, they only want to count the number of peaks.

The Rigol DS1000Zs do all their calculations using the "on screen" data. The main CPU doesn't appear to have access to the original sample data, only the 1200 byte decimated version displayed on screen. All the measurements are done from that. This leads to a couple of hacks.

The only exception to this is the "memory" FFT which Rigol added after too many people noticed that the 1200 byte "screen" FFT was useless.

[SilverSolder]

With careful attention to the FFT settings to keep the span within range of the actual bandwidth of the scope, an almost credible looking FFT comes out of the 54622d. 

It is so sensitive to even minor tweaking that it is hard to know when you can trust what you are seeing...

Presumably the magic of HP's dithering methods is what allows greater than Nyquist performance.  Scope sample rate is 200Msamples/sec, so Nyquist is 100MHz, at the center vertical line in the screen - but the FFT clearly performs well beyond that.

Offset -50dB,  20dB/division