EEVblog #878 - Oscilloscope High Res vs Average Modes

[EEVblog]

What's the difference between High Resolution (Boxcar Averaging) and traditional Average acquisition modes on a modern digital oscilloscope?
Dave explains this often misunderstood difference, traps for young players, and gives several demos to highlight the differences.

[Ivan7enych]

Artifical "noise" generated by DDS signal generator is not a good example.

To show benefits of HiRes mode, I would better connect dynamic microphone directly to oscilloscope and try to see the sound shape buried with much RF noise. HiRes can very efficiently clean in out.

[EEVblog]

Artifical "noise" generated by DDS signal generator is not a good example.

It worked just fine.

[wraper]

Artifical "noise" generated by DDS signal generator is not a good example.

To show benefits of HiRes mode, I would better connect dynamic microphone directly to oscilloscope and try to see the sound shape buried with much RF noise. HiRes can very efficiently clean in out.

For this case proper digital low pass filter is better.

[Blaffetuur]

I've used Average and High res mode on my scope but never new exactly what they did other than cleaning up my waveform

[jolshefsky]

So ... umm ... why not leave hi-res mode on all the time? When would "normal" mode be preferable, and is it really most of the time?

[integritetus]

So ... umm ... why not leave hi-res mode on all the time? When would "normal" mode be preferable, and is it really most of the time?

High-resolution acquisition trades off scope bandwidth for an effective increase in vertical resolution.

Here is a Keysight application note, somewhat technical, that explains the tradeoff.
http://literature.cdn.keysight.com/litweb/pdf/5991-1617EN.pdf?id=2318232

Here is a useful summary from the application note:

Use acquisition averaging when:
– The maximum oscilloscope bandwidth is required
– The signal is repetitive
– Large memory depth is not required
– Control of the number of averages is desired

Use high-resolution acquisition when:
– The maximum oscilloscope bandwidth is not required, or the oscilloscope has
excess sample rate relative to its bandwidth
– The signal must be captured from a single trigger
– Deep memory to capture long time ranges is required

[GlowingGhoul]

So ... umm ... why not leave hi-res mode on all the time? When would "normal" mode be preferable, and is it really most of the time?

High-resolution acquisition trades off scope bandwidth for an effective increase in vertical resolution.

Here is a Keysight application note, somewhat technical, that explains the tradeoff.
http://literature.cdn.keysight.com/litweb/pdf/5991-1617EN.pdf?id=2318232

Here is a useful summary from the application note:

Use acquisition averaging when:
– The maximum oscilloscope bandwidth is required
– The signal is repetitive
– Large memory depth is not required
– Control of the number of averages is desired

Use high-resolution acquisition when:
– The maximum oscilloscope bandwidth is not required, or the oscilloscope has
excess sample rate relative to its bandwidth
– The signal must be captured from a single trigger
– Deep memory to capture long time ranges is required

So a certain WünderSchnitzel that's been telling us the "excess" sampling rate of Keysight scopes is useless, USELESS I TELL YOU, failed to recognize the ability to use high res mode at higher useful bandwidth levels, lol!

[Smokey]

Here is a neat trick I discovered recently.  I had a ULTRA noisy signal from a high current, high voltage PWM source (40A at 300V).  Probing anything on that circuit, you lose the signal in the noise no matter what intensity the display is set to.  With normal acquisition settings, even HighRes mode didn't help.

What I ended up doing was set the scope to HighRes mode.... and.. severely restricting the sample memory to like 16k points.  With so few points, the high res mode was able to pull the signal out of the noise to an amazing clarity.  Too high a sample memory and the HighRes averaging wasn't as effective.

[Circlotron]

In HiRes mode, what is the relationship between memory depth, scan speed, and number of boxcars that are averaged per pixel. More to the point, what is the range of number of samples/boxcars per pixel?

[Galaxyrise]

If you do the same high res tests with your Rigol DS2000, you'll get very different results. 

[rs20]

So ... umm ... why not leave hi-res mode on all the time? When would "normal" mode be preferable, and is it really most of the time?

Because more often than not, the noise is the thing you're worried about. If High-Res mode destroys the little peaks of noise that, for example, causes your comparator to produce glitch pulses, that's a bad thing! An oscilloscope is a device to display truth, not a device for displaying beautifully thin lines with no respect for the truth. Leaving your scope in Hi-Res mode by default leaves you truly blind to a big part of the picture.

High-resolution acquisition trades off scope bandwidth for an effective increase in vertical resolution.

True, although keep in mind that the bandwidth that is lost does not manifest itself as the visible shape of the curve on the screen, but in the fuzziness of the curve itself.

What I ended up doing was set the scope to HighRes mode.... and.. severely restricting the sample memory to like 16k points.  With so few points, the high res mode was able to pull the signal out of the noise to an amazing clarity.  Too high a sample memory and the HighRes averaging wasn't as effective.

What kind of scope do you have? My Rigol has completely different behaviour: the ADC runs at the sampling rate as determined by the sample memory (so, exactly like normal), but at display time all the points that belong to the same column on the screen are averaged together. So if I do a single-shot capture, it looks noise-free, but as I horizontally zoom in (without doing any further captures, just zooming into the original waveform), the averaging is changed to the new size of the pixels, and more noise is revealed. I think this is rather nice, easy-to-use behaviour, because it's very explicitly de-fuzzing the signal in the way that is most appropriate to how you're currently looking at it.

In HiRes mode, what is the relationship between memory depth, scan speed, and number of boxcars that are averaged per pixel. More to the point, what is the range of number of samples/boxcars per pixel?

This is easy to figure out by playing with your scope, and you'd best do that because different scopes behave very differently (see the different between Smokey's scope and my Rigol above). Connect to the 1kHz square wave test signal on your scope, set your timebase to about 1 second per division, and mess with different memory depths, hi-res vs normal mode, and changing the horizontal timebase (both with live captures, and on stored single shot captures). You should be able to make a 5V 1kHz square wave turn in 2.5V DC, but exactly how you achieve that will depend on the scope (on Smokey's scope, changing the memory depth; on my Rigol, simply changing the timebase).

( Also, pedantically, the boxcar is the shape of the convoluting impulse. So you mean the number of ADC samples per boxcar, not the number of boxcars per pixel. )

[Dave Turner]

Hi all, I've been offline for a while and am trying to catch up.

Regarding Dave's video #878 I tried to replicate his observations on my DS1074Z-S with limited success. I really wanted to see the 'false' results as I had an idea that the 'wrong' display might be useful as a diagnostic tool if one properly understood the resulting waveform.

I used a 1MHz sinusoid with a 10Hz FM modulation, which I think is what Dave used, but couldn't get the same result that Dave did. Either my set-up is wrong or my scope does not use the same algorithms that the scope Dave used did.

Dave, you probably have more scopes immediately available to you than than most of your subscribers. Assuming that I've not screwed up my understanding somewhere perhaps it would be worthwhile seeing how your other scopes respond to the same input.

Note that I'm not contesting your warning but I am wondering how other scopes may show different 'erroneous' readings.

ciao

Dave

[Kleinstein]

The false waveform depends on the way the signal is triggered. Triggering such a signal is difficult and different scopes can fail in a different way to trigger it good.

It can also make a difference if the signal is generated from the scope itself as there the internal clocks are likely locked to each other. A separate generator is more likely to have some real jitter on top and also the time to the next trigger can be different. So I would not wonder to see a different picture.

[Cerebus]

Artifical "noise" generated by DDS signal generator is not a good example.

To show benefits of HiRes mode, I would better connect dynamic microphone directly to oscilloscope and try to see the sound shape buried with much RF noise. HiRes can very efficiently clean in out.

For this case proper digital low pass filter is better.

Erm, most 'high res' modes are exactly that, an FIR digital low pass filter.

[Smokey]

What I ended up doing was set the scope to HighRes mode.... and.. severely restricting the sample memory to like 16k points.  With so few points, the high res mode was able to pull the signal out of the noise to an amazing clarity.  Too high a sample memory and the HighRes averaging wasn't as effective.

What kind of scope do you have?

This actually was with a Rigol DS1054Z.  Next time I have more of those PWM units going through I'll take some screen shots of before and after memory restriction in highRes mode. 

[vk6zgo]

Hi all, I've been offline for a while and am trying to catch up.

Regarding Dave's video #878 I tried to replicate his observations on my DS1074Z-S with limited success. I really wanted to see the 'false' results as I had an idea that the 'wrong' display might be useful as a diagnostic tool if one properly understood the resulting waveform.

I used a 1MHz sinusoid with a 10Hz FM modulation, which I think is what Dave used, but couldn't get the same result that Dave did. Either my set-up is wrong or my scope does not use the same algorithms that the scope Dave used did.

Dave, you probably have more scopes immediately available to you than than most of your subscribers. Assuming that I've not screwed up my understanding somewhere perhaps it would be worthwhile seeing how your other scopes respond to the same input.

Note that I'm not contesting your warning but I am wondering how other scopes may show different 'erroneous' readings.

ciao

Dave

I was thinking along those lines,too,Dave----if the crossovers,say, were uniquely spaced for a particular frequency deviation,it would be very useful.
I don't have a DSO,& am unlikely to get one,so couldn't check it myself.

[EEVblog]

Regarding Dave's video #878 I tried to replicate his observations on my DS1074Z-S with limited success. I really wanted to see the 'false' results as I had an idea that the 'wrong' display might be useful as a diagnostic tool if one properly understood the resulting waveform.
I used a 1MHz sinusoid with a 10Hz FM modulation, which I think is what Dave used, but couldn't get the same result that Dave did. Either my set-up is wrong or my scope does not use the same algorithms that the scope Dave used did.

Here you go.
Result depends upon sample rate etc.

[AF6LJ]

This was a good video, one more complication I am missing with my Tek analogue scopes.
Good Stuff Dave   

[rsjsouza]

Good video, Dave. Coincidentally this weekend I was playing with modulation on my func gens and today I captured this short clip on my Rigol DS4014. 

[alank2]

Would it be fair to say that:

Average mode = averaging multiple waveform displays together (requires a steady properly triggered display or it goes badly)

High resolution mode = averaging multiple samples from the DAC into a single sample that will be used to make one waveform.

I like both modes, they are both very useful...

[Wuerstchenhund]

What's the difference between High Resolution (Boxcar Averaging) and traditional Average acquisition modes on a modern digital oscilloscope?
Dave explains this often misunderstood difference, traps for young players, and gives several demos to highlight the differences.

Nicely done, as always!    And I do agree, it's annoying when the current sample mode is not shown on the screen but hidden in the settings.

The only thing that I felt was missing was ERES (or maybe I just missed the part), especially now when Siglent uses it (or at least claims that they do). ERES works slightly different, and there are various advantages and disadvantages of each mode.

So a certain WünderSchnitzel that's been telling us the "excess" sampling rate of Keysight scopes is useless,

The only occasions where I said something like that was when comparing scopes, and what I was said was that Keysight can't turn it's slightly higher sample rate into an advantage.

Don't worry, I'm not expecting you to understand the context.

[NiHaoMike]

Why couldn't they check the difference between captures and display a warning if it's unusually large?

[rs20]

Why couldn't they check the difference between captures and display a warning if it's unusually large?

Because depending on where they set the threshold on that heuristic, you'll either end up with a warning that appears all the time and therefore gets ignored, or you'll end up with a crutch that people would assume to be infallible, while giving false negatives. Or worse, the heuristic would end up doing both of the above things.

It'd be like the "Auto" button all over again.

[nctnico]

Why couldn't they check the difference between captures and display a warning if it's unusually large?

That would be like a hammer with a finger avoidance system 

[Kleinstein]

The warning does not need to be a binary one. With graded display, they could superimpose a few normal and the averaged waveform, so the user could see how much variation is there.

[Someone]

The warning does not need to be a binary one. With graded display, they could superimpose a few normal and the averaged waveform, so the user could see how much variation is there.

like the color graduated display options higher end scopes use....

[wraper]

Artifical "noise" generated by DDS signal generator is not a good example.

To show benefits of HiRes mode, I would better connect dynamic microphone directly to oscilloscope and try to see the sound shape buried with much RF noise. HiRes can very efficiently clean in out.

For this case proper digital low pass filter is better.

Erm, most 'high res' modes are exactly that, an FIR digital low pass filter.

You have some filter with unknown properties which does something, but you are not sure what exactly. And there is no way to adjust it. Take for example GW Instek GDS-2000E series scope, and you can set any cutoff frequency as you want, low pass or high pass. And result is absolutely predictable.

[Cerebus]

Artifical "noise" generated by DDS signal generator is not a good example.

To show benefits of HiRes mode, I would better connect dynamic microphone directly to oscilloscope and try to see the sound shape buried with much RF noise. HiRes can very efficiently clean in out.

For this case proper digital low pass filter is better.

Erm, most 'high res' modes are exactly that, an FIR digital low pass filter.

You have some filter with unknown properties which does something, but you are not sure what exactly. And there is no way to adjust it. Take for example GW Instek GDS-2000E series scope, and you can set any cutoff frequency as you want, low pass or high pass. And result is absolutely predictable.

So you didn't mean, as you said, 'proper' you meant 'configurable or 'fully-fledged'. Assuming, from what looks to be an Austrian flag, you're not a native English speaker - the way you used 'proper' implies replacing something that is not proper, something that is pretending to be or is a substitute for something proper, perhaps even something that is 'ersatz' like wartime coffee.

[wraper]

Assuming, from what looks to be an Austrian flag, you're not a native English speaker - the way you used 'proper' implies replacing something that is not proper, something that is pretending to be or is a substitute for something proper, perhaps even something that is 'ersatz' like wartime coffee.

it's not Austria. Well, I consider high res exactly as "'ersatz' like wartime coffee". And it's not the same as digital low pass filter, it just averages nearby samples. (EDIT: I mean, it is kind of low pass filter but pretty inferior and not very predictable thing with frequency response far from perfect).

[rs20]

Take for example GW Instek GDS-2000E series scope, and you can set any cutoff frequency as you want, low pass or high pass. And result is absolutely predictable.

As a side note, this is the one feature I miss on my scope, and I wish it were more standard. Neither the Agilent Hi-res approach* ( cutoff frequency is a hard-to-figure function of record length and timebase ) nor the Rigol approach ( cutoff frequency is just a function of currently displayed timebase ) are anywhere near the utility of the fully configurable FIR filter.

* Not claiming that Agilents don't have the configuring cutoff as well...

[Cerebus]

Assuming, from what looks to be an Austrian flag, you're not a native English speaker - the way you used 'proper' implies replacing something that is not proper, something that is pretending to be or is a substitute for something proper, perhaps even something that is 'ersatz' like wartime coffee.

it's not Austria.

What flag is that then? It's really useful having some idea where people are from but I find the tiny flags on here problematic. They are too small to immediately recognise any but the most iconic and if they aren't in the top twenty or so well known flags they're damn near useless. I've heard of, say, Turkmenistan but I couldn't tell their flag from a novelty tea-towel.

Well, I consider high res exactly as "'ersatz' like wartime coffee". And it's not the same as digital low pass filter, it just averages nearby samples. (EDIT: I mean, it is kind of low pass filter but pretty inferior and not very predictable thing with frequency response far from perfect).

What is an FIR digital filter but a weighted average of nearby samples? That is exactly how (practical, non-infinite) FIR filters are implemented.

[vk6zgo]

Assuming, from what looks to be an Austrian flag, you're not a native English speaker - the way you used 'proper' implies replacing something that is not proper, something that is pretending to be or is a substitute for something proper, perhaps even something that is 'ersatz' like wartime coffee.

it's not Austria.

What flag is that then? It's really useful having some idea where people are from but I find the tiny flags on here problematic. They are too small to immediately recognise any but the most iconic and if they aren't in the top twenty or so well known flags they're damn near useless. I've heard of, say, Turkmenistan but I couldn't tell their flag from a novelty tea-towel.

Well, I consider high res exactly as "'ersatz' like wartime coffee". And it's not the same as digital low pass filter, it just averages nearby samples. (EDIT: I mean, it is kind of low pass filter but pretty inferior and not very predictable thing with frequency response far from perfect).

What is an FIR digital filter but a weighted average of nearby samples? That is exactly how (practical, non-infinite) FIR filters are implemented.

If you"hover" the mouse over the flag,a small box comes up to show you the country.
Also,re "A Native English Speaker"---not all such hail from the UK.
In Australian English,wraper's usage of "proper" is quite proper.

[Cerebus]

Assuming, from what looks to be an Austrian flag, you're not a native English speaker - the way you used 'proper' implies replacing something that is not proper, something that is pretending to be or is a substitute for something proper, perhaps even something that is 'ersatz' like wartime coffee.

it's not Austria.

What flag is that then? It's really useful having some idea where people are from but I find the tiny flags on here problematic. They are too small to immediately recognise any but the most iconic and if they aren't in the top twenty or so well known flags they're damn near useless. I've heard of, say, Turkmenistan but I couldn't tell their flag from a novelty tea-towel.

Well, I consider high res exactly as "'ersatz' like wartime coffee". And it's not the same as digital low pass filter, it just averages nearby samples. (EDIT: I mean, it is kind of low pass filter but pretty inferior and not very predictable thing with frequency response far from perfect).

What is an FIR digital filter but a weighted average of nearby samples? That is exactly how (practical, non-infinite) FIR filters are implemented.

If you"hover" the mouse over the flag,a small box comes up to show you the country.
Also,re "A Native English Speaker"---not all such hail from the UK.
In Australian English,wraper's usage of "proper" is quite proper.

That's not English, that's Strine. 

I wasn't trying to be prescriptive over English but, given how many non-native speakers we have on here, just trying to make sure we were on the same page. i.e. I didn't want to berate him just because he might have mis-spoken.

[wraper]

What is an FIR digital filter but a weighted average of nearby samples? That is exactly how (practical, non-infinite) FIR filters are implemented.

Only pretty basic and crappy ones.
http://ptolemy.eecs.berkeley.edu/eecs20/week12/freqResponseRA.html

However, if the intent was to design a lowpass filter, then we have not done very well. Some of the higher frequencies are attenuated only by a factor of about 1/10 (for the 16 point moving average) or 1/3 (for the four point moving average). We can do much better than that.

[Cerebus]

What is an FIR digital filter but a weighted average of nearby samples? That is exactly how (practical, non-infinite) FIR filters are implemented.

Only pretty basic and crappy ones.
http://ptolemy.eecs.berkeley.edu/eecs20/week12/freqResponseRA.html

However, if the intent was to design a lowpass filter, then we have not done very well. Some of the higher frequencies are attenuated only by a factor of about 1/10 (for the 16 point moving average) or 1/3 (for the four point moving average). We can do much better than that.

I think you don't understand how an FIR filter works. They are nothing but a series of delays (or samples in time) multiplied by filter coefficients for each delay and the results all summed together. So all FIR filters are, pretty much by definition, a weighted average give or take some overall gain or attenuation. Not just the good ones, not just the 'pretty basic and crappy one' but all of them. All the link shows that is it possible to create mediocre FIR filters as well as good ones, that doesn't change how they are implemented.

Look, if you're trying to say 'I don't like boxcar filters' - fine, say that and be done, your opinion will be respected. But whatever you say won't alter the facts that the boxcar filter we're discussing is a digital low pass filter, that it's normally implemented as an FIR filter, and an FIR filter is  a weighted sum of samples taken over time. Alternatively, if you've got a design for a digital low pass filter that doesn't use an FIR filter, ls better performing and can be economically implemented at 1 Gsps then please share.

[ConKbot]

Ugh, the rigol 1074z needs to work on its high-res mode. I forget the exact details, but I was probing a 1 MHz SPI bus with more than enough samples/second to avoid aliasing, but had the time base zoomed way out because of a 18 ms pause during the spi transaction. More than enough sample memory, should be just fine.  NOPE. Some clock/data bits were only a single point long, no where near the 2.5V they should be at, etc. Decode wouldnt work until I had ~3 bytes on the screen.  Other stuff, where a fast clock waveform should have just been a fat fuzzy trace as the timebase I was at, but wasnt.  Finally figured out it was on hi-res from a previous time using the scope, put it back on normal and I can zoom in on the SPI transaction from way out like I should, and it can decode the transaction once it was wide enough to display the letters. The clock trace turned into a fuzzy line like it should.

I dont know if hi-res decimates the capture down to however many pixels wide the screen is or what, but it definitely wasnt being very useful. DIdnt see any options for how much decimation was performed but wasnt looking too hard.  I'll take my tools at home over my toys at work any day   

[rs20]

I dont know if hi-res decimates the capture down to however many pixels wide the screen is or what...

Yes, this is what Rigols do, as already discussed. Being able to explicitly dictate cutoff frequency is a high-end/rare feature that is unreasonable to expect in a bottom-end Rigol. Even on Dave's agilent, you only kinda get to indirectly manipulate the "cutoff frequency" by changing the record length.

I mean, to deliberately oversimplify things, the whole point of Hi-Res mode is to get rid of the fuzz. If you're complaining that Hi-Res mode is giving you a non-fuzzy line, maybe that's a PEBKAC and you shouldn't be using High-res mode!

[ConKbot]

I dont know if hi-res decimates the capture down to however many pixels wide the screen is or what...

Yes, this is what Rigols do, as already discussed. Being able to explicitly dictate cutoff frequency is a high-end/rare feature that is unreasonable to expect in a bottom-end Rigol. Even on Dave's agilent, you only kinda get to indirectly manipulate the "cutoff frequency" by changing the record length.

I mean, to deliberately oversimplify things, the whole point of Hi-Res mode is to get rid of the fuzz. If you're complaining that Hi-Res mode is giving you a non-fuzzy line, maybe that's a PEBKAC and you shouldn't be using High-res mode!

Eh, I've done data acquisition for years, when I know the signal coming in, and see a sample rate on the screen, I have expectations. If the scope is only showing 1024 or however many samples over 20ms, it should show 51.2ks/s. I don't care what the ADC is running at, I care what is being acquired. I shouldn't have to track down what hidden/non-indicated setting is on reducing the sample rate to some unknown quantity.  Cutoff freq isn't needed but setting the decimation to 2^n or displaying the decimation  really shouldn't be difficult. 

On a spec-an I use regularly, there is 16 letters in a box, and with a glimpse I can tell if the detector is normal, max hold, min hold, power rms, and which of 4 averaging modes, and a few other settings for each of the 4 traces.  An  'N' 'P' 'A' 'H' next to the sample rate wouldn't hurt anyone

[rf-loop]

If scope have what ever "High-Res" mode and if manufacturers claim that in this mode resolution is example up to 12bit or even more. Lot of talking about Hi-Res, averaging and so on. But, I have only one question.

How we get out this 12bit data from scope?

For what is 12bit resolution in 8bit ADC oscilloscope if there is not 12bit resolution acquisition data available out from scope.  Just for "nice picture" or what.  Even old Tektronix TDS can do it, up to 16bit and this data is also available out from acquisition memory.

[rs20]

rf-loop,

How to retrieve the 12-bit data out of a scope is an extremely scope-specific question. But the question of how a scope with an 8-bit ADC can output 12-bit data is very easy to answer: take 16 8 bit samples, add them up, and you now have a 12 bit number. If you average 256 samples, the noise is reduced by 4 bits. So depending on how exactly you defined "12 bits", averaging 16 or 256 8-bit samples will get you there. Further reading: http://cp.literature.agilent.com/litweb/pdf/5991-1617EN.pdf

Note that a delta-sigma ADC is the ultimate manifestation of this principle - at its core, it's basically a 1-bit ADC, but processing vast numbers of samples gives resolutions in excess of 24 bits. Disclaimer: delta-sigma ADCs are actually more advanced that just a 1-bit ADC followed by an FIR, but the general principle stands.

[rf-loop]

rf-loop,

How to retrieve the 12-bit data out of a scope is an extremely scope-specific question.

And this I wonder why there is nearly nothing about this important thing in this thread.

But the question of how a scope with an 8-bit ADC can output 12-bit data is very easy to answer: take 16 8 bit samples, add them up, and you now have a 12 bit number. If you average 256 samples, the noise is reduced by 4 bits. So depending on how exactly you defined "12 bits", averaging 16 or 256 8-bit samples will get you there. Further reading: http://cp.literature.agilent.com/litweb/pdf/5991-1617EN.pdf

Of course but ...
16 give two extra bit for resolution. After 16 we get, yes 12 bit number but only 10bit have valuable data, there need do two times right shift because these two lowest bits do not carry valuable data. After then we have 10 bit resolution.

Perhaps bit better than Keysight papers what are maany times bit colored "truth" for marketing.
http://www.atmel.com/images/doc8003.pdf
and there 3.3


Also in attached image from Texas Instruments.
This table is also equal with Siglent ERES mode. (disadvantage (or bug) there is that final result is then truncated to 8bit in memory so it only reduce noise but do not give out for use this enchanged resolution)

[rs20]

How to retrieve the 12-bit data out of a scope is an extremely scope-specific question.

And this I wonder why there is nearly nothing about this important thing in this thread.

Because it's an extremely scope-specific question.

But the question of how a scope with an 8-bit ADC can output 12-bit data is very easy to answer: take 16 8 bit samples, add them up, and you now have a 12 bit number. If you average 256 samples, the noise is reduced by 4 bits. So depending on how exactly you defined "12 bits", averaging 16 or 256 8-bit samples will get you there. Further reading: http://cp.literature.agilent.com/litweb/pdf/5991-1617EN.pdf

Of course but excuse me, not.  This is (partially) wrong.
16 give two extra bit. After 16 we get, yes 12 bit but only 10bit have valuable data.

Only if AWGN noise dominates over quantization noise. You seem to be confusing ENOB with resolution (think accuracy vs resolution). If you take 8-bit samples of a practically noiseless signal with appropriate offsets, you can get a genuine 12-bit output with only 16 samples. Ultimate extension of this principle is an  8-bit flash ADC is just 256 1-bit ADCs after all. Fair enough, this isn't relevant to an oscilloscope ADC, but it's not "wrong"!

In any case, I clearly stated 256 samples are required for matching noise reduction. You just clearly prefer my second definition of "12 bits".

[rf-loop]

Yes now it is ok. Also as you see I modify my msg before I see this your answer. (because also I find this problem in my original answer)

[Cerebus]

If you take 8-bit samples of a practically noiseless signal with appropriate offsets, you can get a genuine 12-bit output with only 16 samples. Ultimate extension of this principle is an  8-bit flash ADC is just 256 1-bit ADCs after all. Fair enough, this isn't relevant to an oscilloscope ADC, but it's not "wrong"!

In any case, I clearly stated 256 samples are required for matching noise reduction. You just clearly prefer my second definition of "12 bits".

It's probably worth pointing out, 'for those following along at home',  that in a real-world signal with noise the noise floor is improved by multiple sampling and improves roughly as the square root of the number of samples that are averaged whereas the resolution improves with the number of samples (e.g. Your example, where 16 samples gives you 4 more bits resolution but only 2 more bits accuracy). That obviously only improves things in the presence of uncorrelated Gaussian noise (i.e. 'white' noise) not systematic or correlated noise. In the ultimate case it's possible to drag a signal out from below the noise floor if you can take enough samples and the signal is sufficiently strongly correlated from sample to sample.

[thm_w]

If scope have what ever "High-Res" mode and if manufacturers claim that in this mode resolution is example up to 12bit or even more. Lot of talking about Hi-Res, averaging and so on. But, I have only one question.

How we get out this 12bit data from scope?

For the rigol at least, I had to download the full sample memory in 8-bit and perform external FIR filtering via matlab.
To add this export option in firmware, depending on the output format, is probably not much work. But I rarely see complaints about it.

[generic_username]

Hi, I was a long time absent in the forum and while I was just searching for the high res mode on the internet I found out that dave made a vid about that 
It is a very nice vid explaining the differences between those modes but I'm still wondering why does the ampplitude drop at about 8:40 when increasing the timebase in high res mode?
cheers.