hckrnws

Wow, great timing!

I recently got a TDS684A and also made the same discovery, and wrote half a blog post about it which remains unfinished/unpublished. I don't have much of an EE background (at least, not on this level) so my article was certainly worse. It's also my only decent scope, so I don't have a good way to take measurements of the scope itself.

Relatedly, I dumped the firmware of mine (referencing Tom's notes on a similar model) and started writing an emulator for it: https://github.com/DavidBuchanan314/TDS684A

It boots as far as the vxworks system prompt, and can start booting the Smalltalk "userspace" but crashes at some point during hardware initialization (since my emulation is very incomplete!) - some screenshots/logs: https://bsky.app/profile/retr0.id/post/3ljzzkwiy622d

Edit: heh, I just realised I'm cited in the article, regarding the ADG286D

> 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.

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.

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

I've added an additional section about the noise issue after doing some additional measurements: https://tomverbeure.github.io/2025/05/04/TDS684B-CCD-Memory.....

It's a repeating pattern of ~195 samples so it's easy to figure out this pattern during calibration and subtract it from the values measured by the ADC.

In addition to that, there is also some interleaving going on.

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.

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.

Could it be some kind of fixed per-bucket offset?

I wouldn't be surprised if the CCD has all sorts of funky analog stuff going on internally which has different impacts on different samples, which would be incredibly hard to deal with on its own.

However, if this behaviour is merely a fixed offset, it would be fairly easy to compensate for this on the digital side: just do a calibration with a known signal, and the measured offset can be used to reverse its effect in future sampling.

Could be. I've already measured with sawtooth and square waveforms to get a better idea about the interleaving and noise, but not yet with a pure DC input.

Another possibility is that there's some charge decay which you could calibrate for.

I mentioned in another comment, but I started writing an emulator for my TDS 684A's firmware. If I can get it a bit further along (most of the business logic seems to be implemented in smalltalk) then maybe we can answer some of these unknowns by looking at what the calibration routines do.

Having an emulator for that is really interesting.

Is the smalltalk program stored? Some kind of byte code or the original source code?

It's something bytecode-y, I haven't been able to make sense of it yet.

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/gettingte...

> Finally, an off-the-shelf device! But why is it only rated for 25MHz?

Because that's more than enough for scanning a screen-width's-worth of samples from the analog CCD snapshot.

In a digital camara, the CCD columns basically capture the image instantaneously. It has an infinite sample rate!

Then the data is shifted out the CCD some rate that basically doesn't matter, as long as it isn't so slow that it takes seconds.

Unrelated but related:

I have an entry level standalone oscilloscope that i got but never used. I once looked for tutorials and unpacked it, ready to test, but:

It's covered in that kind of plastic that goes all gooey if left unattended for a long time.

Any hints on how I can clean it up so i can touch it again?

Its called soft touch or "rubber oil". It's basically silicone suspended in plastic which gives that soft touch feeling before it biodegrades to the sticky goo.. We can easily clean that with isopropyl alcohol. You will be left with the clean hard plastic below that layer.

There are two variants: One with the soft touch sprayed on after molding, the other mixed in with the pellets before molding. You can't do much for the latter.

lighter fluid - https://www.youtube.com/watch?v=mvTSAIV7FlI

Nice work. Just a comment: ADC clock is very sensitive to capacitive loading of the probe as it cannot drive that large load, it most likely fails to generate enough signal swing. It's not advisable to probe clock signals like that.

But probing the ADC clock didn't impact the clock signal itself, only the ADC input signal. How does that happen? FWIW: the clock signal starts at one side of that large PCB and travels to all 4 ADCs, so it must have a serious driver on it.

> If you ever remove the interconnection PCB, make sure to put it back with the same orientation. It will fit just fine when rotated 180 degrees but the scope won’t work anymore!

I remember pulling a 486 out of its socket in the 1990s and putting it back with the wrong orientation. There was a poc and a bit of smoke. Something on the mainboard had burnt and it wasn't working anymore.

I used smell to locate the fault, a big trace on the PCB, which I soldered back and magic, it all worked again...

I have a story like that! Back in my lab rat days, I dropped a scope probe and its grounded sleeve skittered down the back side of a powered-up, prototype memory board (with 240 individual RAMs on it), leaving a trail of sparks. After that it was dead.

Uh oh. We needed that board. What to do? Well, it can't hurt to try. We had "freeze spray" for debugging purposes, so got a bottle of white-out handy (what's that?), frosted up the board really well on the component side, powered up, and quickly marked the devices that defrosted notably quicker than the rest.

Got the solder station lady to replace all those parts and it worked again.

Old days...

Using your comment as a crutch to post a link to my Eye of Sauron function generator story: https://tomverbeure.github.io/2023/01/02/HP33120A-Repair-Shu.... That was an incredibly satisfying repair if you ignore days of dealing with after effects of burnt PCB material in your lungs.

These days they do that using a vape, cover the PCB with a thick layer of vapor and see where it raises the fast after power-up. Youtube has videos.

What a relief! That you dared to try it. Much harder today...

I didn't even know such things exist. When I wanted to know the delays of 7400-series logic gates, I built a logic analyzer using Arduino and stroboscopic effects which provided resolution on order of 1 ns. But with CCD it seems that it could be done much easier, and allowed to measure exact voltage and not just logic levels.

In the 90s Conrad (think Germany's RadioShack) sold a little audio recorder module that I believe was CCD based. It could only store 16 seconds and quality was terrible but audio stored completely in solid state without moving parts was revolutionary back then.

Thinking about it, I might still have the device somewhere in the attic.

Kit with three of those round digitast buttons and a wide DIP-24 or something like that chip?

Exactly! That's it.

Maybe this CCD/delay-line/bucket-brigade trick can finally inspire someone to make a cheap DIY device that can sample USB3 signals, useful e.g. to check signal integrity. A missing tool in my toolbox.

I mean, I think this would be a very nice project for someone with hardware skills and some time on their hands, and it would be useful too.

This whole exercise started because I wanted to see if it's possible to add a sniffer to the memory chips of the TDS 684B so that I can use it as the analog frontend of a USB scope. It's not possible because of this CCD trick.

I've also looked at the existence of CCD memory, but it doesn't seem to be a thing anymore. I didn't find any such modern chips.

Have you checked this category:

https://www.digikey.com/en/products/filter/delay-lines/688

I didn't know that kind of product existed! I don't think it's useful to slow-sample a fast signal (e.g. they delay digital signals, not analog ones), but some of those chips have a BW of 1.5GHz.

> e.g. they delay digital signals, not analog ones

I think they can be used with analog signals too. After all, they are (I suppose) just a chain of transistors holding charge, like in a CCD. It is just the most efficient implementation, since doing it with logic gates will bring more overhead. EDIT: maybe not, it seems at least some of these chips have digital input/output stages. Maybe it could work if you put a very fast 8-bit ADC in front of the delay line, and used 8 delay lines, one for each bit? :)

Anyway, I have totally zero experience with these chips.

But I can imagine you clock-in the signal using a fast clock (maybe the internal clock), and then you clock-out the signal using a slow clock (slow enough for a subsequent ADC chip).

Also, perhaps you can put a bunch of these combinations in parallel to increase bandwidth, or to increase sample depth.

https://www.psi.ch/en/drs

-not quite the same but similarly novel.

My motivations were similar, although for my needs the 15K buffer + GPIB is just about good enough. I am curious about the unpopulated SRAM footprints on the acquisition board though, I wonder if there's any possibility it could be upgraded to 30K.

The Thunderscope story had a comment about USB3 sampling as well. Here, you cite "check signal integrity," as a reason. I'm curious. Is USB3 signal integrity really an active area of research or troubleshooting? I would have figured this has all been well characterized and mostly solved by device manufacturers selling working USB3 components by the boat load for years now.

High-speed signals are incredibly fragile. The chips themselves are a solved problem, but you still need quite a bit of skill and experience to design a PCB for them. It's quite easy to mess up in subtle ways which look like they should work, but actually ruin your signal to the point of being completely unusable.

If you're designing a board, not being able to look at its signals is a major limitation. Is something wrong with the transmitter, receiver, cable, connector, pcb, firmware, driver? Who knows! It doesn't work, and that's all you're going to get. Have fun randomly tweaking stuff in the hope that it is magically going to work.

> I'm curious. Is USB3 signal integrity really an active area of research or troubleshooting?

Yes, but: You don’t need a general purpose realtime high bandwidth scope for verifying signal integrity.

You can use an equivalent time sampling scope (~10ksps to 100ksps) to measure the eye diagram and ensure that the received eye matches the mask specified by the spec.

The PCB is easy to screw up, so are the cables. And scopes that go that high are expensive but customers expect USB3 even on low volume low cost products so there’s definitely advantages.

There can even just be manufacturing defects the FR4 weave that mess up SI that you might want to check for as a QA step on the assembly line. For hi volume that gets slow or expensive

The ICs are a solved problem, but hobbyists like me designing their own USB3 capable PCB is a different story. Whether or not having a high BW scope would be helpful to detect signal integrity issue is not clear to me: measuring these kind of signal is an art in itself.

"Whether or not having a high BW scope would be helpful to detect signal integrity issue is not clear to me"

Indeed. Even if you have an instrument, and the skill to operate it and produced a meaningful result that reveals some problem, that doesn't spell out a solution. It may help, but high frequency is subtle.

Yes, I experienced a lot of trouble with USB3 cables, cameras, and storage devices in scientific/industrial environments where lots of data is transferred.

There seems to be no way to debug this stuff unless you are a big company that can pay large sums of money.

This isn't nearly as technologically radical, but a Tek scope of that era that we had in the lab then had a sequential colour display. A monochrome CRT with a fast-changing electronic colour filter in front of it. Totally sharp (for the era) colour graphics with no subpixels. Until you moved your head and you briefly saw the individual colour frames. The HP stuff was stodgy and boring and predictable by comparison.

The HP logic analyzers back then had a really neat touchscreen interface based on criss-crossing infrared beams in front of the CRT face. The only thing that I've ever used that felt even better than a capacitive touchscreen, though obviously lower resolution.

I'm pretty sure that this TDS 684B has the same kind of CRT. When I take a photo of the display, you can often clearly see it switching from one color field to the next in the middle of the screen.

As for the HP touch screen, I tore down a bunch of HP 16500A logic analyzers and reverse engineered the touch screen PCB. It uses a pretty simply LED/photo sensor matrix. You can see the PCB in one of the pictures here: https://tomverbeure.github.io/2022/06/17/HP16500a-teardown.h....

That looks fairly similar to what the Kobo Touch still uses today! Pretty cool.

Modern video projectors also display color channels sequentially and also exhibit that effect. Also, this particular CRT technology was not only used in oscilloscopes: https://www.youtube.com/watch?v=z-q8ehzHeQQ

Not all video projectors, only color wheel DLP projectors (which were once the majority of the home projector market but not sure if that’s still the case). Always drove me nuts but apparently most people don’t notice it.

Somewhat related but I just remember when Numega had a color image digitizer with a black and white camera and color filters for the Amiga. I guess the idea is very old because IIRC the very first color photos were taken that way. And speaking of projectors - I think there were projectors / projection TVs that used three black and white CRTS with filters that were merged to produce the final image. Could be higher resolution that way.

Skip forward to 9 minutes and the fantastic vid talks quite a bit about how Tektronix invented the LCD colour shutter tech calling it NuColor.

HPs portable benchtop LAs of the that era with the same X11R5 GUI were better than the modular touchscreen lab units because on the latter you only had a jog wheel and were forced to navigate menu hierarchy with touch for everything else. The portable models had hard buttons for every major mode you'd want to switch between, plus an alpha keypad for convenient label editing.

Not any kind of EE, just a nerd: could this kind of CCD circuitry be used to model a neural network, but in analog-land?

There are similar methods to model neural networks in analog. Switched capacitor circuits like CCD are semi digital as they are time discrete. This makes them very interesting for analog computing. It kinda never got popular though, as capacitors are much larger than transistors, and you need an ADC and a DAC to interact with a digital computer. In fact you need a lot of them, and they consume even more power than the digital alternatives.

The biggest use case for this is sensor interfaces where the signal is still analog (not passed through an ADC yet). Voice recognition is a typical example where analog neural networks are used to certain level of success, and now people are pushing for insge recognition but the architecture of a digital camera isn't compatible with that, so I don't see much happening there.

Funny fact is these kind of circuits are used in analog portion of the chips to implement rather complex calibration/training loops (correlation, LMS optimization, pattern recognition etc) heavily since early 90s. There's a lot of analog computing happening in every SoC.

The values of the nodes in a network could be represented by charged capacitors. The functions which combine the voltages from multiple nodes could be operational amplifiers, with a clock driving everything: telling the nodes to sample their input that is calculated from the current values of nodes.

Can someone recommend a good starter scope?

Depends on what you’ll use it for, but the DS212 has been a lovely low spec and cheap scope for me. There may be better options now, mine’s pretty old.

Tektronix instrumentation from this era (as well as HP/Agilent, and many of the military-affiliated labs) used pretty magical engineering tricks.

In order to be able to design equipment, the instrumentation generally needs to outperform the equipment, sometimes by a significant margin. If I'm looking at the eye of a digital signal, I need to capture much faster than the signal.

It'd be fun to have a book of tricks from this era. At some point, it will fade into obscurity. Right now, it's a whole different bag of tricks for the state-of-the-art. They feel less... more textbook and less clever.

On the other hand, what's nice is that in 2025, decent equipment is cheap. There's a breakpoint where below around 100MHz, you can't do basic work, and above you can. That's roughly where FM pickup and a lot of oscillations sit. That used to cost a lot, but as technology progressed, we're at a point where a decent home lab can be had for well under a grand.

   In order to be able to design equipment, the instrumentation generally needs to outperform the equipment, sometimes by a significant margin.

To which he replies: why don’t we use what the guys from the oscilloscope are using?

> It'd be fun to have a book of tricks from this era.

Though it isn't a book, the Hewlett-Packard Journal is a gold mine for this type of content: https://web.archive.org/web/20120526151653/http://www.hpl.hp...

E.g. An 8-Gigasample-per-Second, 8-Bit Data Acquisition System for a Sampling Digital Oscilloscope (October 1993): https://web.archive.org/web/20120526151653/http://www.hpl.hp...

>It'd be fun to have a book of tricks from this era.

I think you'll get a kick out of:

"Analog Circuit Design: Art, Science and Personalities"

https://www.amazon.com/Analog-Circuit-Design-Personalities-E...

Any recommendations for a sub-grand scope that a hobbyist will never be limited by?

Never is too absolute of a statement, but I bought a Siglent SDS2304X 7 years ago. It has 350 MHz of bandwidth and a 2 Gsps sample rate. I don’t think I’ve ever been limited by what it can do.

I paid the full price for it (>$2000), but almost all Rigol scopes have lower and higher BW versions with the same chassis and electronics. You can buy the lower BW versions for a fraction of the higher one and apply a software patch to unlock the higher BW.

This has been possible for years. Despite many SW revisions, Rigol has never made an effort to block this. I think they know that in the grand scheme of things, they make more money this way.

So research 100MHz Rigol scopes and check if they are hackable. Chances are high that they are.

on some of the latest models rigol put in a few hardware features that are only on the higher bandwidth versions, but the lower bandwidth are still hackable to higher bandwidth.

It's all about extracting as much as they can from each customer, but I'm happy they're willing to allow "willing to do hacking stuff" to be a market segmenter for them.

Unfortunately, I think the GP post question of a device a hobbyist can't exceed has a negative answer. I really like my DHO924S and there are a huge number of tasks where it is way more than enough (and it's very portable too). But around me there are kinds of computer and video and radio devices that run much higher than 250MHz, and going to 350 or 500MHz doesn't really change that fact. Scope prices go pretty exponential after a couple hundred MHz. ... so if you want to be snooping on a SFP+s sfi signal, USB3 some hdmi thing or whatever via a scope anything but a lucky surplus find is unlikely to be inside a hobby budget.

Yet I think it's totally reasonable for a hobbyist to want to work on the high speed digital signals that surround them in their own home.

(hobby solution to fast digital busses is to make custom boards with inexpensive FPGAs I guess, rather than using a 5+-figure oscilloscope)

In practice, you can get very far with high speed interfaces while never using a high-speed scope. I've spent years professionally doing HDMI, DisplayPort and LVDS interfaces with FPGAs and never had the need for scope shots. (Admittedly, a colleague did use it, but that was always for certification related measurements which is a non-issue for hobbyists.)

Part of the reason is that the specifications for these fast interfaces are created for robustness. They can sustain a lot of PCB design abuse and still work fine in practice even when out of spec.