De-capping & circuit analysis of hybrid modules

[D Straney]

HP/Agilent mystery module
To follow up on that one, whose workings were very clear, here's another module from Agilent back in 2001, which I know nothing about.

This one again came to me already decapped, so I don't have a part number, and was unable to find a reference to anything that looks like it.

It's clearly centered around one IC, plus some resistors and capacitors.  Two side-by-side possibly-differential connections enter on the top edge in the photos.  The bottom and bottom-right edges also contain 14 almost-identical traces that run in parallel from the IC to the castellated connections - it's possible this is a 14-bit ADC or DAC with a parallel interface?

Let's take a look at the IC:

The function is still not obvious here.  There's a dense maybe-digital section in the middle, surrounded by more widely-spaced analog sections with visible resistors and capacitors.  A large bank of capacitors is at center-right.

I'm curious about the white squares scattered throughout, which look like exposed metal on the top layer (same as the bond pads).  The way these are arranged, these look like test points - but there's a lot of them, and they're very small compared to the bond pads.  Were these maybe internal testpoints probed during development, rather than for manufacturing tests?

Here's some closer views of the analog sections:




...and the dense middle section:


Anyways, let me know if you have any idea what this might be.

[D Straney]

Teltone M-958-02 "PCM-compatible DTMF receiver"
Here's a module from some kind of telecom setting:

The actual function is slightly strange.  According to the datasheet, it receives a digital data stream, converts it to an analog signal, and then feeds it to an analog DTMF decoder.  DTMF was used not only for dialing on landline phones, but I think also for digital signaling in between switching offices.  It's kind of funny to me that the DTMF tones, in this case, encoded digital signaling inside an analog signal inside a digital signal (the PCM input), but I guess that's the kind of situations you see when there's the backwards compatibility, slow equipment migrations, and long-lived standards needed for very large installations like the telephone network.  A modern approach would be to directly filter the incoming data stream digitally, and produce the decoded DTMF numbers without ever converting to analog, but the detour through the analog domain makes the chips much more interesting to look at here.



The chip on the left is the DTMF decoder.

This is made by Silicon Systems Inc., and is marked with SSI250, which sure looks like a part number.  It doesn't seem to be an off-the-shelf part, as I couldn't find any matching part numbers in their catalogs at bitsavers, so this is probably a custom modification of one of their standard DTMF decoder ICs.  We can get an idea of how it works by looking at these off-the-shelf DTMF decoders - the SSi201, for example, uses an array of switched-capacitor bandpass filters matched to each of the DTMF tones.  You can see a lot of metal squares in the die photo below, which are probably those capacitors (switched-capacitor circuits are going to be a theme here, by the way).  If you're interested, there's an entire application note here describing how the SSi201 works.

The chip on the right is the PCM decoder / DAC, which is most likely an AMI S3502 (large PDF warning; go to p.189).

This is not just a serial-input DAC, but also incorporates "µ-255 law" compression:

This kind of non-linear companding (compression/expanding) is common in radio/telephone/etc. as it lets you use fewer bits in your ADC/DAC for the same perceived sound quality.  With audio, a single LSB makes much more difference when outputting a small value (LSB is a larger percentage of the output), vs. when outputting a large value (LSB is small percentage of the output).  Using a logarithmic/exponential(-ish) transfer function gives the bits more equal weight, which allows eliminating some unused bits and stuffing more parallel voice channels on the same transmission line.

Here's the block diagram of the S3502 so you can see all the other functional blocks:

The way that the DAC is implemented is also pretty interesting.  The most common technique for digital-to-analog conversion is to use an R-2R ladder, however this uses a more complex scheme.  There's a small resistor ladder here that converts a few bits, but the bulk of the digital-to-analog conversion is done by charge balancing in binary-weighted capacitors.  I'm guessing that the benefit has to do with on-chip resistors having very wide tolerances: previous R-2R DAC ICs I've looked at have almost always needed laser trimming on the resistors.  With on-chip capacitors, however, I think the relative values should have a much tighter tolerance, as their capacitance only depends on the (fairly precise) metal layer geometry.  You can see all the metal squares from those capacitors in the die photo above.  There's also a switched-capacitor reconstruction filter that follows the DAC, so that's another source of the many, many on-chip capacitors.  There's an IEEE paper that describes how this chip works in fascinating detail: https://ieeexplore.ieee.org/document/1051304
If you don't have access through work or school, you can find it through a convenient website whose name rhymes with ShyCub. Oops, no I was wrong about that sadly.

Now that you know what's on this chip, here's my best guess of how the functionality maps to the different physical sections, based on the pin numbering & connections:

Let me know if you see any mistakes there.

[D Straney]

Burr-Brown 3650 isolation amplifier (optical)

Here's an isolation amplifier, that transmits an analog signal across an electrical isolation barrier.  This is useful for industrial equipment (where you may need to sense current through a high-voltage conductor and transmit it to an earth-grounded control system, for example), medical equipment (where EEG & EKG leads need to be heavily isolated to avoid deadly shocks through low-resistance connections), specialized measurements, etc.  This particular one uses light, and has a bandwidth of 15 kHz.  Let's open it up:



You can see a few bare dies here - the 3 largest ones are op-amps.  All identical designs, but I wasn't able to figure out the manufacturer and part number:


There's also two silicon capacitors (no images, as they're boring to look at), and a resistor:


The middle section, with the white rectangle covered by the 2 black lids, is where the magic happens.  I removed these, and found the optical parts underneath - the lid structure is flipped over, on the right-hand side, so that you can see the two channels inside it:


Here's a closer look at the main substrate:

All 4 of these dies are photodiodes.  There are 2 separate channels here, each in one slot of the lid structure to block light from each other and from outside.
What transmits the light?  There's 2 blank spaces you can see in the photo above if you look carefully: it turns out the LEDs were pulled off the board when I removed the lid.  They're still stuck to the bottom of the lid structure:

It looks like these channels are filled with a transparent potting material, which the LEDs stuck to.  The bondwires also stayed in the potting material.  This potting allows light to pass from the LEDs (on the transmitter side) to both photodiodes (one of which is on the receiver side), while still providing some high-voltage electrical insulation (rated at 2 kV).
You can also see the negative shapes left by the photodiodes in the potting material, which is kind of fun.

So, we have 2 optical channels, each holding an LED and 2 photodiodes.  Let's talk about how this all works:

Each of the LEDs is coupled to 2 photodiodes: one on the transmitting side, and one on the receiving side.  This is for linearization: LED brightness vs. current is not linear at all, and varies a lot with temperature.  Photodiodes are better with the linearity of their their light-to-current characteristics (as long as you stay far above the dark current), but still not perfect.  So if you use 2 matched photodiodes, you can use a feedback scheme where the transmitting side adjusts the LED current until it sees the correct output from its photodiode; the receiving side will then be seeing the same output from its photodiode too.

This is the popular approach behind the HCNR201 analog optoisolator IC, and a few similar ones: the theory of operation and example circuit sections of that datasheet are very much worth reading if you're interested in details.  I've looked into this approach pretty deeply, back when I worked on power electronics and was frustrated by the lack of good high-side gate drive probing (but didn't get much of anywhere with these off-the-shelf parts because the bandwidth is low and the CMRR is not great).

The reason for the 2 separate channels, is because the input signal is bipolar (can be positive or negative), while each LED/photodiode set can only work with one polarity.  D1, D3, and D5 handle the positive signals, while D2, D4, and D6 handle the negative signals.

Even though the output is a voltage, the input is in the form of currents: if you have a voltage input, just connect a resistor in series.  The currents are effectively summed at the inputs of U1, with some kind of RC snubber to tune the frequency response I guess.  Putting the LEDs in-line with the op-amp power supply pins is a clever approach: this eliminates the need for separate LED-driver circuitry that eliminates the crossover distortion between positive & negative.  R1 provides a dummy load on the op-amp's output so that its supply currents will be roughly proportional to its output voltage.  Because the positive and negative photo-currents are subtracted on both transmitting and receiving sides, the LED currents also don't have to be precise: the feedback loop corrects for everything.

For a vague idea of how it works on the receiver (output) side:
D6 conducts a photosensitive current through R9 that creates a positive voltage, proportional to the D6 current.  U3 buffers this voltage.  (Connecting D6's cathode to this buffered voltage, instead of Vcc, is done to maintain 0V across it and therefore mostly remove any bandwidth-limiting effects of D6's capacitance, if I remember correctly).  U2 then looks at the difference between D6's "negative output" voltage, and D5's photo-current through R10, and generates the output voltage based on that.

[magic]

You can see a few bare dies here - the 3 largest ones are op-amps.  All identical designs, but I wasn't able to figure out the manufacturer and part number

Some sort of 741.
https://www.eevblog.com/forum/projects/opamps-die-pictures/msg6234899/#msg6234899

[D Straney]

Tektronix 155-0236-00 channel-switch module
This comes from the Tektronix 2400 series of oscilloscopes:



There's only one die, so I guess most of the the hybrid area is to fit the heatsink?
All the information I have on this hybrid comes from the wonderful Tek Wiki, which has a detailed description including where it appears.  It also includes this schematic excerpt which shows roughly what it does - there's two separate switches, which feed two amplifiers.  These amplifiers then have their outputs directly connected together.  Everything is differential, presumably because it's all pretty high-bandwidth and it's easier to keep unwanted crosstalk (and grounding imperfections) from causing problems that way.

It looks like it selects between various signals to send to the vertical amplifier: some of these inputs are the actual input channels of the oscilloscope, while some are connected to fixed voltages (maybe unused?) and Ch. B comes from the trigger voltage, so that the trigger position can be viewed on-screen.

Now that we know what it's supposed to do, here's the die:

Nice colors - the green rectangles are resistors, used as input termination in some places and put in series with inputs or outputs in other places.

Here's the most important pin assignments, and a rough idea of how each area of the die matches the functions:

[D Straney]

Meant to mention, the 155-0236-00 wasn't used in all the 2400-series scopes: only a few specific all-analog portable models.  These models go up to 300 Mhz bandwidth, so the bandwidth of the switches and amplifiers on this die must've been somewhat higher than that.

Next, from a similar background we have the...
Tektronix 155-0240-00 sweep hybrid
This was also used in the 2400 'scope series, as well as the 11302 scope and SCD1000 digitizer.



Yet again, Tek Wiki comes to the rescue with lots of information!  It looks like this module is in charge of the horizontal sweep generation.  In the left-hand IC, some control bits are shifted in via a serial interface.  These control bits both control a DAC to generate sweep currents, and also select between different capacitors for sweep-speed ranges.  A 2nd DAC generates different delays for starting the sweep.  The right-hand die handles the logic of when to start and stop each sweep, and switching the currents appropriately.

If this is hard to read, there's also block diagrams of the left-hand IC and right-hand IC from an internal Tektronix catalog of their in-house products, as referenced from Tek Wiki.

Here's the left-hand die (Sweep DAC), with the digital controls, 2 DACs, and the capacitor selection switches:

It's relatively easy to figure out what's happening here, since there's only a few functional blocks, which all have distinctive repeated sections and connect directly to their associated pins:


The right-hand die (Sweep Integrator), is much harder to figure out though.

Even after tracing the main signals on the metal layers, I could only place a few of the functional blocks:


If you want to know more about the details of this sweep generator, you can read patent US4705961A (as referenced by the Tek Wiki), or follow other links from the Tek Wiki page on this hybrid to internal Tektronix documents.

[D Straney]

National Semiconductor AF133 active filter
This little guy is a self-contained 5th-order elliptic filter.  It's a low-pass meant for anti-aliasing (before digitization) in communications systems: the goal is to have a flat frequency response up to 3.2 kHz, and then > 16 dB attenuation at 4 kHz, for an 8 kHz sampling rate.

Anyone who's gotten into higher-order analog filters probably knows that the component tolerances matter a lot - if the various values don't balance out correctly, you can end up with more passband ripple than you expected, or with the zeros not attenuating as much as they should.  A hybrid module, with the ability to have very precise laser-trimmed resistors directly on the ceramic substrate, therefore is a good way to go about this.

This actually has 2 covers, one on the top and one on the bottom:


The top has all the resistors...


...while the bottom has all the silicon:


The two sides are connected together only by the external pins, which clip onto both top & bottom.  Therefore, most of the pins on this module are marked in the datasheet as "no connection", as they're used only for internal connections between the top & bottom sides.

The two "normal" dies on the bottom side are LM148 quad op-amps:


The other flat-gray-looking pieces of silicon are capacitors.  They're similar but not all exactly the same, and each one has a number laser-etched into it.  I haven't been able to figure out what these numbers mean.




The circuit isn't too complicated to trace fully:

Take the resistor values with a massive heaping of salt, as I had to measure them in-circuit: so the values are inaccurate and only there to give you a ballpark sense.

There's two weird-looking biquad filters, I think (U2D/A/C & U1D/A/C) with a single purely real pole added (U2B).  The dual integrators are a dead giveaway, even if they're not fed directly from the input signal.  It doesn't match any of the standard biquad implementations I could find - the KHN was the closest, but still distinctly different.  I decided to take half an hour and figure out the transfer function and then the parameters as functions of component values, for myself...and then quickly lost my notes with the results.  There was nothing that stood out though: most components affected every parameter, so there was no independent control over the corner frequency and Q (by changing separate components) like in a state variable filter.  So, not being a filter expert myself, I have no idea why they used this particular filter topology.  If you have any insights or see any mistakes on my part, let me know.