Scope Wars

[gf]

Actually DSP can do the job. You can combine a modest analog filter followed by a digital filter to zero out the region in which the aliasing is present.

The keyword here is "modest", though, meaning there is still a minimum requirement. If we want to place the digital cut-off at (say) fs/4, then the analog filter must still provide a sufficient a priori attenuation at 3*fs/4 and beyond.

Anoter question is of course, how many dB to consider/define "suffcient"?
(Individiual requirements may vary)

[nctnico]

Hi,

I am going to speculate that more fundamental issue is the ratio between bandwidth and the sampling rate.

The Nyquist Sampling theorem tell us the bandwidth is half the sampling frequency. Then you get folding and aliasing.

There is some modulation caused by the sampling frequency.
If you look at V(sample) waveform, You have to guess at the input waveform from the sampled data.
There isn't enough data to create the waveform with any degree of confidence.

A 100Mhz scope with 500Msps isn't enough data points.

Sorry, but that is the wrong conclusion. Nyquist says that all the information from a signal is there up the fs/2. It is just that your brain can't make a signal from the dots. But that is a problem in your brain and not in the number of samples. In order to help your brain DSOs have sin x/x reconstruction which connects the dots in a mathematically correct way to show you the signal. Because there needs to be some headroom for an anti aliasing filter most DSOs have a maximum bandwidth of fs/2.5 .

Best we stick with unhacked rated BW and retail prices to avoid all confusion.

Even I find that total nonsense. If you can get more bandwidth from a scope using a simple key generator then it has that extra bandwidth before you bought it.

[2N3055]

Again off the rails...

Back to the topic. Reg KNOWS his DSP. I argue with him when he confuses seismography equipment with a 1 GHz (input bandwidth) scope.

Reg properly said that by choosing proper input filter you will get perfect impulse response.
He properly said that that filtering can be done partially in analog and partially in digital domain.
He properly said that using DSP you can "idealize" (shaping it to achieve proper calibrated response) scope pulse response.
He properly said that by using proper reconstruction filter you can fully reconstruct sinewave with more than 2,5 samples per period. Make note that reconstruction filter is not simply low pass filter.
He properly said that even if scope doesn't have Gaussian frequency response, by choosing reconstruction filter appropriate for it's response, you get the scope to reconstruct signal properly.
He properly said that FFT on most scopes is far from being optimal, not to mention easy to use, or giving directly usable results similar to SA.
He properly said that decimation by simple throwing away samples is wasteful. Data can be downsampled by filter and gain additional dynamic range and lower noise in usable bandwidth. That also filters out all higher frequency components (low pass filter) so it takes care of aliasing . If that is what you want to accomplish.  (edit:that is important if want to further process data, if you simply show it on the screen it's not important)

Things he gets wrong are connected with not knowing how other people use scope, and generally having limited insight into oscilloscope use cases by industry.

Also we are getting back to people again saying " we need at least 10x oversampling".  No you don't. If band limiting and reconstruction is done correctly, you don't. Because, if you do that correctly, there is nothing in the signal reaching the A/D converter in between those samples that doesn't fit perfectly on top of interpolated sine segment. On a 1 GHz scope, looking at a 1 GHz square wave, you have to see perfect sinewave on the screen.

Reason why Keysight itself used to recommend 5x oversampling (which I find good compromise) is that input filtering is not ideal...
On most 1GHz scopes and up nowadays they use brickwall (sharp rolloff) filtering and sampling factor of 2.5 with all channels on.

That brickwall filter takes care of aliasing. But pulse response of that filter is not the same as Gaussian response filter. It has pros and cons. Pulse response is worse, it overshoots. But, same 1GHz scope will have 15% more accurate rise/falltime measurements. Additionally, it's frequency response will be essentially flat, gaining excellent amplitude accuracy, all the way up to the cuttof. In practice, that means that on 900 MHz you get less than 5 % amplitude error, compared to almost 30% on Gaussian response scope.
Also that distorted pulse response? Well, it isn't there if you don't feed scope stuff over 1.2 GHz, in amplitude that is high enough to show on screen. That overshoot shows only when you feed it signal with components in gigahertz range. You need to buy (or make) special pulser to even be able to create pulse that fast, and connect pulser directly on scope input. Even a foot of coax cable will make things very different.
In practice that is pretty much not going to happen. There are no 40 ps pulses in Raspberry Pi, Arduino, switching and analog PSUs, and pretty much 90% of all electronics. Unless you are working on multigigabit data links, picosecond lasers, or designing MRI machines or such, no worry. And if you do, your employer bought you the fancy stuff. That is no small business or hobby territory.
People who work on 40ps pulses need 20GHz+ scopes and equipment. DUH.

Stuff is, of course, more serious when same things happens on a 100 MHz scope. That bandwidth is pretty much easily pushed into aliasing with anything on your desk.
So question here is how to make it to not alias. Again it is simple. Cheaper scopes are being made with 1x1GS/s A/D for 4 Ch (those that will get in trouble), and those that have 2x1GS/s (those that will be doing better).  Also if aliasing and pulse response is more important than maximum bandwidth, you can always opt NOT to buy highest bandwith machine.
It is quite obvious really. You don't get Keysight MSOX3104T, but deliberately get MSOX3054T or even better MSOX3034T. Those will still sample at 5 GS/s but with maximum 350/500 MHz frequency are guaranteed not to alias. And step response will be perfect. Really, just buy Rigol DS1054Z and DON'T hack it to 100MHz. Take a look at that pulse response...It will be slow, but perfectly behaved.

[gf]

There is some modulation caused by the sampling frequency.
If you look at V(sample) waveform, You have to guess at the input waveform from the sampled data.
There isn't enough data to create the waveform with any degree of confidence.

A 100Mhz scope with 500Msps isn't enough data points.

The example fails to reconstruct the sampled signals exactly due to the following violations of the samling theorem:

1)

The reconstruction filter is not sufficient. The sampling theorem requires a reconstruction filter with a boxcar frequency response, which is equivalent to sin(x)/x interpolation. In the analog domain this is impossible to realize. Digitally it is well possible to up-sample to higher sampling rates with sinc interpolation (of course not to an infinite sampling rate, but in a DSO a finite resolution is still sufficient for the reconstructed waveform).

2)

The signals are not sine waves, but truncated (-> true sine waves had an infinite extent in the time domain). A sine wave modulated with a single boxcar impulse is no longer a band-limited signal, though, thus violating the sampling theorem.

Even if we consider the captured buffer periodic, assuming that it repeats over and over again, then the repetition of the buffer will indeed lead to an infintie (sampled) sine wave for the 100 kHz signal, but not for the 107 and 203 kHz signals, because there are discontinuities at the wrap around from the end of the buffer to the start of the buffer. An exact reconstruction of the infinite sine wave from a finite buffer is only possible if the buffer would contain an exact integral number of signal periods. If the captured buffer is large enough this can be addressed by truncating the sinc interpolation kernel with a (not too short) window function, which still leads to a reasonable approximate reconstruction then (at least at the center of the buffer).

Btw, the audio world also manages to recontruct a 0..20 kHz analog signal from 44 kSa/s CD data (with an accuracy inside the SNR limit which is dictated by the 16 bits anyway).

[gf]

Best we stick with unhacked rated BW and retail prices to avoid all confusion.

Even I find that total nonsense. If you can get more bandwidth from a scope using a simple key generator then it has that extra bandwidth before you bought it.

The consideration is only valid, though, if the hacked frequency response is really the same as for the higher-priced regular models. This would either need to be verified in the first place, or proved otherwise (e.g. by reverse engineering) with sufficient confidence.

Otherwise the hacked one and the higher-priced regular one would rather need to be considered different "models".

[nctnico]

Best we stick with unhacked rated BW and retail prices to avoid all confusion.

Even I find that total nonsense. If you can get more bandwidth from a scope using a simple key generator then it has that extra bandwidth before you bought it.

The consideration is only valid, though, if the hacked frequency response is really the same as for the higher-priced regular models. This would either need to be verified in the first place, or proved otherwise (e.g. by reverse engineering) with sufficient confidence.

Otherwise the hacked one and the higher-priced regular one would rather need to be considered different "models".

Nowadays the difference is in software only. The dead giveaway is that you can buy options in the form of license keys to increase the bandwidth.

[Fungus]

Best we stick with unhacked rated BW and retail prices to avoid all confusion.

Alright, let's do that then.

How much would a 4-channel, 50MHz DSO have cost back then?

(It would have been sooooo easy for you to just say "100MHz is cheating but a 50MHz DSO would have cost XXXXX back then")

Nowadays the difference is in software only. The dead giveaway is that you can buy options in the form of license keys to increase the bandwidth.

Yep, and Rigol KNOWS we hack them and you can bet they're not losing money.

If you want to go to 2-channel DSOs you can a lot cheaper. Owon, Hantek sell 2-channel scopes for a lot less than $350. How much would they have cost back then?

Heck, you can even got one of these for under $150: https://www.eevblog.com/forum/testgear/fnirsi-1013d-100mhz-tablet-oscilloscope/

That's probably a match for most 20Mhz analog CROs but it has batteries, touch screen and lots of measurement functions including FFT.

The claim that things aren't advancing in the o'scope world is ridiculous.

[tv84]

Let's keep the war at the technical level.

20 years ago the chinese weren't in the game so you can easily imagine how much the prices were inflated...

[Jay_Diddy_B]

Hi,

I am going to speculate that more fundamental issue is the ratio between bandwidth and the sampling rate.

The Nyquist Sampling theorem tell us the bandwidth is half the sampling frequency. Then you get folding and aliasing.

There is some modulation caused by the sampling frequency.
If you look at V(sample) waveform, You have to guess at the input waveform from the sampled data.
There isn't enough data to create the waveform with any degree of confidence.

A 100Mhz scope with 500Msps isn't enough data points.

Sorry, but that is the wrong conclusion. Nyquist says that all the information from a signal is there up the fs/2. It is just that your brain can't make a signal from the dots. But that is a problem in your brain and not in the number of samples. In order to help your brain DSOs have sin x/x reconstruction which connects the dots in a mathematically correct way to show you the signal. Because there needs to be some headroom for an anti aliasing filter most DSOs have a maximum bandwidth of fs/2.5 .

Sure, there are ways to reconstruct a sinewave from 2.5 samples per period.
But I believe the signal has to be continuous and have no frequency content beyond fs/2

If you reconstruct a sinewave from a small number of data points, you are making the assumption that the signal was sinewave in the first place.

Regards,
Jay_Diddy_B

[nctnico]

Hi,

I am going to speculate that more fundamental issue is the ratio between bandwidth and the sampling rate.

The Nyquist Sampling theorem tell us the bandwidth is half the sampling frequency. Then you get folding and aliasing.

There is some modulation caused by the sampling frequency.
If you look at V(sample) waveform, You have to guess at the input waveform from the sampled data.
There isn't enough data to create the waveform with any degree of confidence.

A 100Mhz scope with 500Msps isn't enough data points.

Sorry, but that is the wrong conclusion. Nyquist says that all the information from a signal is there up the fs/2. It is just that your brain can't make a signal from the dots. But that is a problem in your brain and not in the number of samples. In order to help your brain DSOs have sin x/x reconstruction which connects the dots in a mathematically correct way to show you the signal. Because there needs to be some headroom for an anti aliasing filter most DSOs have a maximum bandwidth of fs/2.5 .

Sure, there are ways to reconstruct a sinewave from 2.5 samples per period.
But I believe the signal has to be continuous and have no frequency content beyond fs/2

If you reconstruct a sinewave from a small number of data points, you are making the assumption that the signal was sinewave in the first place.

No. Just read more about sampling theory and Nyquist. The signal doesn't need to be continuous but there is a requirement that the original signal (before sampling) doesn't have frequency content above fs/2 which is what the anti-aliasing filter is for. See gf's example of digital audio where they push the bandwidth to fs/2.2 .

[David Hess]

Well, a Pentium Pro CPU of 1995 managed a 150Mhz clock frequency.  I would guess (but haven't done the math) that a modern processor is roughly 100x to 1000x better, given increased clock frequencies, higher IPC (instructions per clock), and high amounts of parallelism / multiple cores.

Processor speed is not what limits the performance of an oscilloscope.

Practically all DSOs are limited by processor speed, which is what leads to dead time between acquisitions even with double buffering.  Producing a usable display from a multiple Gbyte per second stream of data takes massive bandwidth.  The highest performance hardware manages it with heroic amounts of parallelism but nobody here is going to pay for that.

Display refresh rate has nothing to do with it because processing should accumulate every acquisition during each frame; slower frame rates just mean more acquisitions for each one.

I mentioned earlier designing a DSO taking processing limitations into account.  The primary limitation is memory bandwidth for decimation or processing the acquisition record suggesting that on a modern high performance processor, record length will be limited by cache size for a given performance with higher level caches being significantly slower but allowing larger record lengths.  Offhand I do not know of any DSOs which take advantage of this but I suspect some do without mentioning it.

[Fungus]

Sure, there are ways to reconstruct a sinewave from 2.5 samples per period.

Or even less than that.

The problem is that the width of your reconstruction window (and the number of samples that need to be processed) approaches infinity as you head in that direction. 2.5 Is a practical limit in real life.

[Fungus]

If you reconstruct a sinewave from a small number of data points, you are making the assumption that the signal was sinewave in the first place.

No, you're making the assumption that the signal is bandwidth limited to the Nyquist frequency.

[2N3055]

Hi,

I am going to speculate that more fundamental issue is the ratio between bandwidth and the sampling rate.

The Nyquist Sampling theorem tell us the bandwidth is half the sampling frequency. Then you get folding and aliasing.

There is some modulation caused by the sampling frequency.
If you look at V(sample) waveform, You have to guess at the input waveform from the sampled data.
There isn't enough data to create the waveform with any degree of confidence.

A 100Mhz scope with 500Msps isn't enough data points.

Sorry, but that is the wrong conclusion. Nyquist says that all the information from a signal is there up the fs/2. It is just that your brain can't make a signal from the dots. But that is a problem in your brain and not in the number of samples. In order to help your brain DSOs have sin x/x reconstruction which connects the dots in a mathematically correct way to show you the signal. Because there needs to be some headroom for an anti aliasing filter most DSOs have a maximum bandwidth of fs/2.5 .

Sure, there are ways to reconstruct a sinewave from 2.5 samples per period.
But I believe the signal has to be continuous and have no frequency content beyond fs/2

If you reconstruct a sinewave from a small number of data points, you are making the assumption that the signal was sinewave in the first place.

Regards,
Jay_Diddy_B

You are absolutely correct. If you are looking at 100 MHz squarewave signal on 100 MHz scope you should get 100 MHz sinewave on screen.
Because your scope must not show 300 MHz and 500 MHZ and 700 MHz at any amplitude to be visible on screen. Built in filtering must take care of that that neither of those frequencies and sharp changes in signal ever reach A/D converter.

If you need to look at 100 MHz squarewave, you need scope with 1GHz bandwith, to see at least first 9 harmonics..

[gf]

The signal doesn't need to be continuous but there is a requirement that the original signal (before sampling) doesn't have frequency content above fs/2

The problem in practice is that even a band-limited non-periodic signal still has an infinite extent, i.e. you'd need to capture an infinte number of samples in order to enable sinc reconstruction, which is not feasible.

The problem is that the width of your reconstruction window (and the number of samples that need to be processed) approaches infinity as you head in that direction. 2.5 Is a practical limit in real life.

It is in fact always infinite, since a sin(x)/x impulse has generally an infinite extent.

The only special case is when the captured signal is periodic and the captured buffer happens to contain an exact integral number of periods for each frequency contained in the signal. Then the (finite) buffer can be considered being repeated periodically, enabling e.g. exact sinc up-sampling via FFT.

In the general case, you always have to truncate the reconstruction window at some point in practice, enabling only an approximate reconstruction then.

[Fungus]

In the general case, you always have to truncate the reconstruction window at some point in practice, enabling only an approximate reconstruction then.

I prefer the term "good enough".

[Jay_Diddy_B]

Hi,

To change the direction of the thread a  little, other considerations when using a scope:

1) What is the bandwidth at the probe tip?

2) Is the 50 input really 50 ?

2b) does it change with the attenuator settings?

2c) can you help a lot by putting a 6dB pad on the input?

I have measured the input impedance of my

DSOX 3034A
MDO4104
TDS3032

and there are some differences.

I don't have any 100MHz class scopes to measure.

Regards,
Jay_Diddy_B

[Sighound36]

Must admit chaps is thread still going  my goodness can we not just settle for say six fundamental requisites that a novice could use to select a scope for their own specific purposes?

I mean I respect those that have the time to deep dive into this subject but do you really, really need to? Are the scopes these days that bad at all?

If I took into account every possible variant before choosing a scope, I would be left with a mental break down FFS.

Agree we do not wish to end up with the device for our own needs and yes some manufacturers are not so truthful with their figures, however figures only tell you so much the real world use is where the real point of decision is made imho.

More T&E companies should have more extensive loan stock for genuine potential purchasers of their products imho, some are really good other are 'Mey' as our esteemed leader would say.

The Keysight MXR demo has gone awfully quiet to.

[tv84]

The Keysight MXR demo has gone awfully quiet to.

LeCroy rulez??   

[rhb]

Here's a 20 MHz, 200 mVpp square wave from my 33622A.

First the gospel according to the 485.



Now the MSO-2204EA with normal sampling



Then in peak detect mode.  Notice that the aliases that fall below 200 MHz become visible in the FFT.



So same thing with the DS1202Z-E in sample mode:



And peak detect:



I forgot to set the anti-alias filter to on, but turning it on/off has no affect.  I'm using a triangle window.

I am forced to conclude that on the Instek "peak detect" turns off the anti-alias filter.  I am unable to interpret the Rigol behavior.  I'm using a 6k memory buffer for the FFT but unable to determine whether the spurious responses are internal noise or numerical artifacts.   An off by one indexing error in an FFT produces artifacts that are very difficult to interpret  unless you can do the FFT on a pure impulse and a constant.

If you have an analog filter which does no achieve -6 dB/bit  at Nyquist, energy above Nyquist is folded back.   If a digital filter that suppresses everything that gets through the anti-alias filter follows you can remove the aliased frequencies above the frequency where the aliased signals are less than -6 dB/bit.  Naturally this leads to a steep skirt and lots of ringing as we see commonly.

I'm pretty well aware of what sorts of things people use a scope for, I've been using one for quite a long time and have repaired both a 60 MHz  Dumont 1062 and a 100 MHz Tek 465 analog scope with major faults due to cracked solder joints.  I'm also enumerating the full range of use cases, but that's  a subject for another time.

What I'm doing here is performing tests which show the actual internal engineering of the DSOs.  So I'm considering the various options for implementing a DSO and then making up tests to see what particular scopes are doing.

That's an iterative process.  They are not all the same and so one has to devise new tests as information is gained.

The input signal to a scope is an analog waveform.  The sina qua non is that the scope accurately display the analog waveform within the limitations of the instrument.

The Instek is clearly a better instrument than the DS1202Z-E.  So the real question is will Rigol fix the bugs I find.  However, at current prices the Instek is 2-3x the Rigol.  You can hack a GDS-2072E to 200 MHz, but it's almost twice the cost of the DS1202Z-E.  And if you buy the GDS-2202E it's over 3x the Rigol.

Most people neither need nor can afford a $10k+ instrument.  For $300, the Rigol is not bad.  I know how it compares to the Instek, so the remaining question is how does it compare to a Siglent in the same price class.

[bdunham7]

No, the BW in these scopes is determined by the band limiting amplifier IC in front of the ADC. All software does is sends commands to switch the band limits. Therefore in order to figure out a scope's front end filter shape  all one needs to do is refer to that IC's datasheet.

I wish that were true, but experimentation shows otherwise.  At 500MSa/S, my Siglent 1104X-E will alias a 300MHz or 400MHz signal quite clearly, in fact the aliased 400MHz signal has greater amplitude than the genuine 200MHz! 

However, if I shut off CH2, the 300MHz signal is completely suppressed.  If the BW filters were all before the ADC, I don't see why that would happen.  To me this is a semi-serious shortcoming and the lesson is that if you want a good 2-channel scope the 1104X-E is it.

The first 4 pictures are 1GSa/S and 100, 200, 300, 400MHz, then 500MSa/S and 200, 300, 400MHz.  I appear to have missed the shot of 500MSa/S @ 100MHz, but it looked similar but with slightly greater amplitude.

[tom66]

@rhb, you have two channels enabled on the Rigol, which will reduce the sample rate to 500MSa/s.  Does it look any different with CH2 off?

[bdunham7]

Sorry, but that is the wrong conclusion. Nyquist says that all the information from a signal is there up the fs/2. It is just that your brain can't make a signal from the dots. But that is a problem in your brain and not in the number of samples. In order to help your brain DSOs have sin x/x reconstruction which connects the dots in a mathematically correct way to show you the signal.

As the audiophools like to say, Nyquist is just a theory.  It isn't just your brain that struggles to make a signal from the dots.  The few dots there are, the harder it is to reconstruct the signal accurately.  Just look at the example he posted.  Small amplitude errors can result in huge reconstruction errors at low sample counts while at fs/10 one slightly misplaced dot won't hurt too much.  Yes, in theory you can reconstruct a sine wave at fs/2.1 with a 48-pole unity-gain filter or its digital counterpart, but in practice that becomes quite difficult. I can show you distortions on an actual A-brand scope that are similar to the fs/3.3 example, even though is well within Nyquist, like fs/6 or so.  This particular scope is a bit older and has a sample rate 10X the BW.  I guess they agreed that 10X was a good number.

And then, consider that Nyquist, AFAIK, doesn't cover triggering, an important part of oscilloscopy.

Sorry about the terrible picture, I had to use my phone to get the picture off the camera since I can't find my cable at the moment...
This is a 300MHz signal (the scope attenuated it quite a bit), it shows averaging in use because I forgot to take a picture without it, but it still had similar issues in the SAMPLE mode. 

[SilverSolder]

Must admit chaps is thread still going  my goodness can we not just settle for say six fundamental requisites that a novice could use to select a scope for their own specific purposes?

I mean I respect those that have the time to deep dive into this subject but do you really, really need to?

[...]

Well...  if you go to a track day with your car, you kinda have to expect people to talk a lot about cars...     

[rhb]

To try to get some data about  the anti-alias filter in the Instek I decided to really go up far past the rated BW and Nyquist.

Here we have the 8648C:



Here's what the Instek shows:



I have to turn up the gain on the scope to get it to trigger, but 700 MHz on a 200 MHz scope is rather interesting, though clearly there is somewhat of a sampling theory issue.

NB As you approach Nyquist the phase and amplitude resolution is reduced.

Have Fun!
Reg