ADC sample and hold

[JosiahCochran]

It is to my understanding that most ADC's have some sort of sample and hold circuit on the input. I know the basic theory of how these work, but how do they get them to work on the order of GSa/s? As far as discrete sample and hold circuits, most of the things I see online are on the order of microseconds. I tried making a circuit in LTspice and even the GHz op amps have too high of an input capacitance to work at these speeds. Is it possible to make a sample and hold circuit that works on the order of picoseconds with discrete components? how do they do it in high bandwidth ADC's?

[David Hess]

Is it possible to make a sample and hold circuit that works on the order of picoseconds with discrete components? how do they do it in high bandwidth ADC's?

That was possible to do decades ago using various high bandwidth sampling configurations.  Tektronix wrote a book on the subject and details are available in their comprehensive service manuals up until about 1990.

As a practical matter where Tsampling > Tsettling, a 25 ohm source from a terminated 50 ohm transmission line into 2.5 picofarads (double the shunt capacitance of a 4 diode sampling bridge) gets to 2.5 GHz.  Beyond that, sampling at lower than 100% efficiency is used as described by Tektronix above.

[JosiahCochran]

So what I gather is that you can't really do it with op amps, you have to build a high band width transistor amplifier to accomplish these time scales?

[David Hess]

So what I gather is that you can't really do it with op amps, you have to build a high band width transistor amplifier to accomplish these time scales?

Do you mean use operational amplifiers as part of a high bandwidth sample and hold?  No, and in general feedback amplifiers are not suitable.  Transistor choppers are also too slow and suffer from excessive charge injection.

Fast discrete samplers actually operate at the transmission line level.  I am not sure how the integrated ones built into modern ADCs work.

[JosiahCochran]

could you elaborate on what you mean by transmission line level? I am not very familiar with RF circuits. How can a transmission line be gated to sample at a certain frequency? By transistor choppers you mean using an FET as a switch?

[David Hess]

By transistor choppers you mean using an FET as a switch?

Yes, exactly.  Bipolar transistors can also be used as switches.  The switches used for discrete samplers are fast diodes like schottky or microwave diodes in various configurations.  I want to try making one using transistor base-emitter junctions which are also very fast.

Like I said, I do not know how high bandwidth samplers in integrated circuits work.  I suspect most are simply an extension of more common sample and hold circuits but yield better performance because they can take advantage of lower parasitics with such small construction so they have a more traditional RC type of frequency response.

could you elaborate on what you mean by transmission line level? I am not very familiar with RF circuits. How can a transmission line be gated to sample at a certain frequency?

Fast samplers work by moving charge around.  The input looks like a 25 ohm transmission line which gets drained into a capacitor or another transmission line while the sampling switch is closed.  Another design, the traveling wave gate, works by "disconnecting" a section of the transmission line and measuring the total charge inside.

The reason everything looks like a transmission line is that the sampling duration is so short that there is no time for the output side of the sampling gate to "see" any further into the input side.  10s to 100s of picoseconds is typical so this is fractions of an inch to inches.

[T3sl4co1l]

Yes, it is possible.

As far as I know:

A transmission gate is used.  Also known as an analog switch.

These are available as board-level components with quite reasonable dynamic specs (~ohms on-resistance, 10s pF off-capacitance, bandwidth ~1GHz).  Though I don't know that you can pull off a S&H with as much bandwidth, because of the particulars of propagation (I suspect they have some multi-stage transmission-line-like magic which is able to deliver higher on-state bandwidth than Zo*Cp alone implies), let-through, and most importantly of all, commutation (how much time is required for the switch to transition from closed to open: the sample aperture).

Note that such performance levels require very small transistors, and so the supply voltages are low, and the signal voltages must be within those supplies.  (I forget what the best performance is; there might be a near-1GHz switch with +/-15V range, which is very impressive indeed.)

This is all tightly integrated on chip, to charge the sampling capacitor, which is just as it sounds, a small capacitor (some pF) paired with the relevant switches and sense amps which make the ADC an ADC.

From there, it depends on the architecture, but faster ADCs (starting in the low MS/s) are mostly pipelined SAR type, which pair the sample cap with another cap of exactly equal value, that's been precharged to VREF or 0; then its voltage is amplified by a factor of 2, and so on until the desired bits have been obtained.  The secret sauce that constitutes the precious IP of an ADC, includes timing tweaks, gain compensation and so on, so the result is obtained with minimal nonlinearities, no missing codes, and etc.

With everything being neatly on chip, parasitics can be controlled tightly; a capacitor might not be much more than a few transistors and their drain capacitance, or even the capacitance under a transistor's channel*.  You can do some of this with board-level components, but it's impossible to achieve nearly as good performance (between all of: size, accuracy/ENOB, sample rate, power consumption).

*IIRC, this is the mechanism behind the famous analog "bucket brigade" device: as a transistor turns on and off, the channel capacitance itself is modulated, essentially squeezing out the charge under one transistor, into the next and so on.  So the analogy to a real bucket brigade is amazingly apt.  Because charge is conserved between stages, minimal amplification is required, even after many stages.  Doing this with analog switches and fixed capacitors, you would need a buffer after every capacitor, or a gain of 2 after every other capacitor, etc., plus you'd incur the total distortion of all those stages.

(No, this doesn't add much to the present conversation, more just rephrasing or hinting at things that have already been mentioned.)

Tim