3 ms·

> The input to the ADC is clearly already chopped into discrete samples, with a new sample every 120 ns. We can discern a sine wave in the samples, but there’s

> The input to the ADC is clearly already chopped into discrete samples, with a new sample every 120 ns. We can discern a sine wave in the samples, but there’s a lot of noise on the signal too. Meanwhile the TDS684B CRT shows a nice and clean 1 MHz signal. I haven’t been able to figure out how that’s possible. Is this maybe using some form of correlated double sampling? > It looks like the signal doesn’t scan out of the CCD memory in the order it was received, hence the signal discontinuity in the middle. Or maybe the samples are also interleaved in the low-order bits in some way. This could be because the organization of the CCD isn't symmetric for the input and output paths, perhaps to reduce area or power, since only one path has to be fast. This would make sense because if you implement the CCD using n parallel CCD bucket brigades you only have to put a fast S&H and multiplexer in front of it, then you can drive the CCD brigades at a fraction of the actual sample rate, and the capacitive load of each of those clock phases is much lower as well.

dp-hackernews 1y ago
I've not read the article, but from your response, is this similar to "oversampling" from the world of CD audio? Oversampling Versus Upsampling: Differences Explained https://www.soundstagenetwork.com/gettingtechnical/gettingtechnical200311.htm https://www.soundstagenetwork.com/gettingtechnical/gettingte...
tverbeure 1y ago
I've been doing more measurements after some discussions on discord after publishing the blog post. There's definitely some kind of interleaving going on, but even when applying a square wave at the input (that gets displayed perfectly fine on the scope CRT) there is a lot of noise on the ADC input that can't be explained away. I will update the blog post tonight after doing more measurements. Some people have also suggested deliberate addition of a pseudo-random signal that gets removed after sampling to counteract some CCD issues. But I don't know how that would work either.
- monster_truck 1y ago
  The shapes in the noise vaguely look like they are repeating. Perhaps something like time%sampleRate (or some other context dependent value)? Easy enough to filter out while still providing enough to know it's a coherent signal
  - formerly_proven 1y ago
    You mean like a signal added to combat / detect aliasing? I'm not sure how that would work. I know HP did something like this, but they did it by introducing a known jitter into the time base. I don't think it could work by merely adding a signal, you need to interact with the incoming signal in the time domain to do something about aliasing. I have no experience with these older CCD-based ones, but the later TDS'es from the 90s could do aliasing. The time base jitter thing was probably patented by HP.
    - photon_rancher 1y ago
      Yeah, this is a well known technique for boosting resolution. For example: You add a dithering signal which can be processed out. If the signal has the right properties (for example, random but evenly distributed noise bounded to one LSB), you can then average out multiple samples to get more effective resolution than the ADC has. The additional number of bits scales something like 2^n samples, although if you don’t take sufficient samples this mainly just reduces your SNR. It also requires a periodic input. However you can also pull similar tricks in the time domain or using simultaneous sampling with multiple ADCs. You can also interleave slower ADCs with a phase shift. This produces stitching artifacts unless you average them out though because ADCs generally are not well matched at the limits. You can bin or calibrate this out somewhat if you can characterize the error. You can do a similar thing in the frequency domain if the ADC sample window is narrow enough but it has arguably the worst artifacts. Lo-pass the first ADC around N/2. The second ADC use a bandpass from N/2 upto N. The third is N upto 3N/2 etc… but the fourier transform will have a bunch of junk at the stitching points. Or you can take the sampling scope approach using a fast but low sample rate ADC and many triggers. I’ve seen most of these done on commercial instruments if you dig into the settings. Some of them you can see in normal operation (like the stitching in the frequency domain). But I think the other poster was suggesting the first case applies - if you think about it there are certain periodic signals you can add instead of a random signal. That has the advantage of limiting SNR degradation and can also be filtered out easier/ detected i n the data.

dp-hackernews 1y ago

I've not read the article, but from your response, is this similar to "oversampling" from the world of CD audio? Oversampling Versus Upsampling: Differences Explained https://www.soundstagenetwork.com/gettingtechnical/gettingtechnical200311.htm https://www.soundstagenetwork.com/gettingtechnical/gettingte...

tverbeure 1y ago

I've been doing more measurements after some discussions on discord after publishing the blog post. There's definitely some kind of interleaving going on, but even when applying a square wave at the input (that gets displayed perfectly fine on the scope CRT) there is a lot of noise on the ADC input that can't be explained away. I will update the blog post tonight after doing more measurements. Some people have also suggested deliberate addition of a pseudo-random signal that gets removed after sampling to counteract some CCD issues. But I don't know how that would work either.

monster_truck 1y ago
The shapes in the noise vaguely look like they are repeating. Perhaps something like time%sampleRate (or some other context dependent value)? Easy enough to filter out while still providing enough to know it's a coherent signal
- formerly_proven 1y ago
  You mean like a signal added to combat / detect aliasing? I'm not sure how that would work. I know HP did something like this, but they did it by introducing a known jitter into the time base. I don't think it could work by merely adding a signal, you need to interact with the incoming signal in the time domain to do something about aliasing. I have no experience with these older CCD-based ones, but the later TDS'es from the 90s could do aliasing. The time base jitter thing was probably patented by HP.
  - photon_rancher 1y ago
    Yeah, this is a well known technique for boosting resolution. For example: You add a dithering signal which can be processed out. If the signal has the right properties (for example, random but evenly distributed noise bounded to one LSB), you can then average out multiple samples to get more effective resolution than the ADC has. The additional number of bits scales something like 2^n samples, although if you don’t take sufficient samples this mainly just reduces your SNR. It also requires a periodic input. However you can also pull similar tricks in the time domain or using simultaneous sampling with multiple ADCs. You can also interleave slower ADCs with a phase shift. This produces stitching artifacts unless you average them out though because ADCs generally are not well matched at the limits. You can bin or calibrate this out somewhat if you can characterize the error. You can do a similar thing in the frequency domain if the ADC sample window is narrow enough but it has arguably the worst artifacts. Lo-pass the first ADC around N/2. The second ADC use a bandpass from N/2 upto N. The third is N upto 3N/2 etc… but the fourier transform will have a bunch of junk at the stitching points. Or you can take the sampling scope approach using a fast but low sample rate ADC and many triggers. I’ve seen most of these done on commercial instruments if you dig into the settings. Some of them you can see in normal operation (like the stitching in the frequency domain). But I think the other poster was suggesting the first case applies - if you think about it there are certain periodic signals you can add instead of a random signal. That has the advantage of limiting SNR degradation and can also be filtered out easier/ detected i n the data.