Why is 1M waveforms/sec so impressive?

[EEVblog]

The likelyhood to catch a rare glitch is independent from the screen refresh rate, it depends on the waveform capture rate. Once catched and stored in screen buffer and intensity set to infinite, it will appear on screen with 60Hz.

Yes, the screen refresh rate is a red herring, it's got nothing really to do with the O/P's question. To answer the O/P question, 1M waveform/sec is impressive because it allows you to potentially capture a random glitch, on average 1000 times faster than a scope with only 1K waveform/sec.
http://cp.literature.agilent.com/litweb/pdf/5989-7885EN.pdf
I have demoed this in a video.

[Wuerstchenhund]

Almost by definition, that is a GPU inside the Agilent ASIC, it's essentially a Graphic Processing Unit. What's your point?

The point is that you can either have a GPU or do all the calculations in software in a general purpose CPU which is slower. GPUs can help with many of the calculations in a modern scope, but come at a price.

I think it's very unlikely they have used an off-the-shelf GPU architecture, it would be purpose written verilog/VHDL to do the specific task and nothing more.

Not necessarily, as using a common GPU architecture (even if it's a relative simple one) would bring certain advantage in terms of functional expandability. It can also bring cost benefits in not having to re-invent the wheel.

But at the end of the day we will never know as I'm sure Agilent won't tell us.

[KedasProbe]

I do wonder if it's not more useful to provide the max sample speed without skipped time of a scope.
Then if you set your scope to sample at twice the speed you know you are missing at least half of it.

edit: Showing % dead time would be nice on a scope screen. (although I doubt they will do that)

[EEVblog]

Not necessarily, as using a common GPU architecture (even if it's a relative simple one) would bring certain advantage in terms of functional expandability. It can also bring cost benefits in not having to re-invent the wheel.
But at the end of the day we will never know as I'm sure Agilent won't tell us.

This is the 4th generation MegaZoom ASIC, so would almost certainly build upon technology in previous generations.
http://www.hit.bme.hu/~papay/edu/Lab/MegaZoom.pdf
Previous versions went through the CPU to the display, but the latest one doesn't it uses what Agilent call the "plotter"


It's almost certainly custom instead of some off-the-shelf GPU core. Its is also the reason why the Agilent can't resize the display window etc, it's a fixed size and location in the custom display logic. That crippled the 4000 series too that used the bigger screen but the same ASIC chip.

[Gunb]

I do wonder if it's not more useful to provide the max sample speed without skipped time of a scope.
Then if you set your scope to sample at twice the speed you know you are missing at least half of it.

One thing is independent from the other. Using an ADC with twice of the speed only takes twice of the samples than before - nothing else. Assumed the timebase is the same you've got only higher horizontal (at least graphical) resolution of the same waveform, but the dead-time between subsequent waveforms remains the same - nothing gained.

wfm/s cannot be replaced with higher sample rate of the ADC.

[KedasProbe]

You are right that nothing else changes including the processing power.
Hence the CPU can't follow and you start to get dead time.

[Marco]

AFAICS there simply is too little demand for high waveform/s capture and some type of histogram display on the low end ... the purchasers of low end oscilloscopes are not discerning enough.

It's a shame too, the front ends are getting pretty decent ... but the DSP is atrocious.

[marmad]

I have to say, while I can appreciate the technological achievement of 1M wfrm/s, the recent posts by kg4arn showing the X3000's zoomed display of undersampled waveforms makes me wonder what price is being paid for that speed.

[free_electron]

these machines don't use GPU's. not needed. the approach is completely different. you don;t need shaders and polygon stuff . it's not a gaming machine. what you do need is a 2d image composer that has high bandwidth. you can't cram that through pci , you'll be boggin the system down.

i have had a technical explanation on how they do these things. what i am explaining now is based on the infiniium architecture ( and the older 546xx series. ) in essence the megazoom system.

1) let's assume we have infinite acquiition memory. we have a convertor that digitizes the incoming signal at a fixed speed ( as hard as it is designed to go). lets say 1 gigasample per second.
there is a trigger comparator that looks at the data coming from the convertor. when 'trigger' this logic block resets a counter.
the counter controls the address lines of this infinite memory.

so now you have a basic acquisistion system. we have a valid trigger event : reset the counter , at every subsequent clocktick a sample is written to memory and the address counter is incremented so the next sample will fall in the next memory slot.

so far so good ( forget about pretrigeer and other stuff , i'll come back to that later...
this memory has a second address bus and second data bus . this port allows random access to samples in the memory. this bus runs at amuch lower speed. This kind of memory is called a FISO : Fast-in , slow out. it's not a fifo ! data will be overwritten when a trigger comes. the output is not sequential as in a fifo but random as in normal memory. the agilent scopes today still ave this FISO memory, it is on board the a/d convertor and runs at full throttle. pure static ram. the backend pipe goes to the slower ddr ram. ( ddr can't follow the breakneck speed the frontend runs at)

2) we have a display system. the display consists of an lcd deisplay that is refreshing itself at a reasonaby low speed. 60 times a second it can refresh its pixels.

the display gets its data from an image buffer. this buffer is dual ported. this means that the block filling the buffer does not have to wait for the display.

the image buffer itself gets its data from an image composer. and this is where the magic happens.

3) the image composer behaves like a stack of transparencies . remeber the plastic sheets used on a projector. you draw with colored pen. you can put one or more transparencies on top of each other to create a composite image. that is exavlty what the image composer does. it has drawing 'planes'
one plane per channel ( 4 channels is 4 planes ) 1 plane for static images like the graticule , 1 plane for quasi static text , 1 plane for the graphical menuing system and so on.

what you see through the projector is a composite image.

the cpu plots its text and buttons on the allocated planes. it never gets to see the trace data as that is drawn on other planes , by the trace plotter.

4) the trace plotter. for clarity and ease understanding. let's assume our viewport on the lcd display is 16 pixels wide. i know very bad resolution but this is explanation.

there is two thing. timebase and time offset.  timebase dictates how far the points are apart from each other. time offset dictates how far from the trigger point we begin.
if i display at fastst timebae with zero time offset i need to show sample 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15 and im done.
if i scroll in time i show sample 25,26,27 .... up till (25+16)

if i decrease timebase 1 step the distance between samples changes.
i now show sample 0,2,4,6,8 ...
take it down another notch and it becomes 0,5,10,15,20,25 .. and so on.
if i scroll now. it could be sample 27,27+5 , 27+10 , 27 +15 .... to 27 + (16x5)

so this image composers goes to to FISO and grabs only the samples it is interested in. it does not care about the rest ! the sample value gives it the vertical position on the screen the pixel needs to be plotted.

it puts a byte there with value 255 (max intenisty )

5) the grader. this blok does the image decay. a sample is a point taken from memory and plotted with max intensity. on every sweep of the lcd refresh all the pixels on the plane are now decremented in value (intensity ! , NOT position ! )
that is how you create the afterglow.
control the speed of the decay  (how many times a second you decrement and by what value you decrement ) and you control image persistence.

and there you have it. the basics of the megazoom system

there ismore at play. the megazoom actually interpolates between groups of pixels so you don't miss spikes

the memory is not infinite , bt is timed in such a way that the a/d sampling speed does not need to come down for a long time. the deeper the memory the longer you can go flat out . once you hit slower timebase speed you hit the end of memeory before a screen refreh and then you need to throttle down the a/d

the image composer keeps writing as hard as it can to the plane. even if the lcd can only refresh 60 times a second the drawing plane can collect millions of waveforms. so that is what they mean with the waveform rate : how many waveforms can we overlay on each other inbetween an lcd refresh cycle. the more you can write there the bigger the chance is one waveform will be the glitch you were looking for. on the next lcd refresh it gets shown on screen and then the decay kicks in..

[jpb]

AFAICS there simply is too little demand for high waveform/s capture and some type of histogram display on the low end ... the purchasers of low end oscilloscopes are not discerning enough.

It's a shame too, the front ends are getting pretty decent ... but the DSP is atrocious.

I don't think it is a case of not being discerning - it is more a matter of lack of choice and trades off. Agilent have high waveforms/s but are expensive (for purchasers of low-end scopes) so if you have a limited budget you've got to decide between different scope features and waveforms/s is only one of them. The 2000X series were cut down in terms of features too far and Agilent now seems to realise this and are upgrading with more memory and serial decodes as options. The 3000X series are very nice but they are two or three times the price.

Waveforms per sec as a feature is difficult to assess. Agilent push it quite hard as it is one of their selling points, but even at 1M WF/s the scope is blind most of the time at faster time bases so in app notes the rate of occurrence of rare events has to be selected to be infrequent enough for other scopes not to spot them but frequent enough to be seen with 1M WF/s. It is difficult to get objective information as to how often the feature is really needed.

[Marco]

Just to get a feel of the numbers I'm going to quickly calculate the necessary bandwidth to get limit-case waveform capture on a 8 bit 1 GS/s scope for the simplest type of display (ie. completely binary, no ability to do some type of histogram calculations). AFAICS for every sample you need to do a RMW on a 256 bit word, with each bit indicating whether a signal was ever present at that amplitude at that moment in time after the trigger). So total memory bandwidth 64 GB/s and with access being completely linear any type of commodity DRAM can be used ... have I got that right?

Quite a bit of bandwidth, but doable.

[Gunb]

It is difficult to get objective information as to how often the feature is really needed.

Exactly. I would say, higher wfm/s is more wanted than needed.

Many scopes have got only 2000-4000 wfm/s, but they've got plenty of measurement features you need rather than higher wfm/s. That's the reason why the Hameg HMO beats the RIGOL DS4000. I've bought both scopes but wfm/s wasn't an issue at all, rather a nice-to-have add on.