Input capacitance

So now that nobody publishes simplified schematics of their ICs any more, perhaps somebody here can enlighten me.

Back in the old days, there was something called an 'input capacitance'. You changed the voltage on an input, it cost you a bit of charge. No worries.

Nowadays there are 'differential' and 'common mode' input capacitances. Still no huge problem--an unbalanced input signal is half differential and half common-mode, so effectively Cin = 0.5[Cin(diff) + Cin(CM)].

But some parts have hugely different CM and differential capacitance, e.g. the OPA657, whose Cin is 0.7 pF differential but 4.5 pF CM!

Anybody know the origin of the big difference?

Cheers,

Phil Hobbs

Reply to
Phil Hobbs
Loading thread data ...

Probably the ESD diodes add the cm capacitance, and they're measuring the differential capacitance 3-terminal style.

2.25

-------+--------||----------- | | _|_ | ___ 0.7 | | |

-------+--------||----------- 2.25 | | | rails

John

Reply to
John Larkin

Specsmanship ;-)

Ignoring the well-to-substrate capacitance allows one to "spec" the tiny differential part.

It's always been there ;-)

...Jim Thompson

--
| James E.Thompson, P.E.                           |    mens     |
| Analog Innovations, Inc.                         |     et      |
 Click to see the full signature
Reply to
Jim Thompson

No

l

I don't really "know", but my speculation is: the CM capacitance is mostly stray C. After all, neither is very large. The DM capacitance is small because the actual JFETs are tiny, and probably cascoded.

Yes, it's a real shame that even the simplified schematics no longer are part of most datasheets :(

-f

Reply to
cassiope

No

l

It could be a function of input bias cancellation circuitry.

Note that input capacitances are usually GBD (guaranteed by design) parameters. That is, assuming wafer parametric test is satisfactory, the capacitance limit should be valid. Expect the value to be guard- banded heavily unless it is very important to the application.

Reply to
miso

Hmm, that's probably part of it, true. I think there's more circuity stuff down underneath there though...the OPA656, which is the unity-gain stable version, specs 2.8 pF CM.

I wish I knew how they were defined and measured, too. For instance, your drawing certainly has 4.5 pF CM capacitance, but if you really drove it differentially, e.g. with a transformer, it'd have 1.825 pF differential capacitance. With a pi network there's no way for the differential Cin to be less than a quarter of the CM Cin.

I thought about the following possibilities:

  1. C_gs of the two input devices appearing in series for differential, but not for CM. Whoops, that should be worse for differential--it's bootstrapped for CM.

  1. Capacitance of current source driving the sources. Right sign this time, but if (1) is small, this has to be small too because it's in series with C_gs.

  2. Miller capacitance caused by non-ideal behaviour of cascodes. Also wrong sign.

But the possibility that strikes me is

  1. They've cross-coupled the gate of each input device to the opposite device's drain to get some negative Miller capacitance, sort of like unilateralizing a push-pull RF power amp. From a circuits point of view that would look like a negative differential capacitance, which might explain how C_diff can be less than C_cm/4. It doesn't help me any in building low-light TIAs, because it's the single-ended capacitance that matters for loop stability with high R_F and so on. I don't see how it helps anything else either, since even when building noninverting amps feedback forces the input swings to be nearly all CM. Mysterious.

The reason this is rotting my socks just now is that National discontinued the LF357 in 2004 (*requiescat in pace*), and there's nothing available with its combination of FET input, +-15V supplies, 2 pF C_in, 12 nV noise, and 15 MHz GBW. The OPA656 is a drop-in replacement but only runs on +-5, so you lose 10 dB dynamic range (it also costs $5 vs $0.75). The OPA627 has low noise and enough speed, but it also has 7.5 pF input capacitance (8 differential, 7 CM). That forces me to drop the feedback resistor by a factor of almost 3, and loses me a decibel of SNR, while costing $18.

This is _after_ the photodiode capacitance has been fixed by the bootstrapped cascode, mind---all I want is for the stupid amplifier to get out of its own way. Grrr.

As Mehitabel would put it, "Toujours gai, toujours gai."

Cheers,

Phil Hobbs

Reply to
Phil Hobbs

I saw a simple circuit from Bob Pease for "rolling your own" JFET front end for TIA's. It's printed out and tucked into a folder somewhere.

George

Reply to
ggherold

Thanks. A couple of BF862s will do a good job at AC, true. The problem is the amount of extra crapola you have to hang on them to get decent DC behaviour. (Also one begins to wonder why one bothers with an op amp at that point.)

The basic issue is that 300K resistors are so very noisy compared to almost anything else, so you have to use really big ones to overcome their Johnson noise, which in turn leads to all sorts of bandwidth problems like this.

I have a new design idea based on using all current feedback that more or less avoids this problem by making the feedback currents very quiet. (This is done by dropping lots of voltage across the resistors in the current sources.) Still a partly-baked idea at this point, but it should be possible to do a better job with a simpler circuit that way.

Cheers,

Phil Hobbs

Reply to
Phil Hobbs

It is if you do a 3-terminal C measurement. My old green Boonton c-meter would measure that 0.7 pF accurately and ignore much bigger caps to ground. So it depends on the capacitance model they used but didn't define.

John

Reply to
John Larkin

Interesting. I've never played with one, but given the very strong tendency to believe test equipment, that might well be it. I suppose the only thing to do is get out the old soldering iron and see--and here I was trying to mail the MS tomorrow.

So I suppose one good way to measure C_CM would be to build a zero-gain amplifier (a diff amp with the two inputs tied together), but with much higher impedance in the + arm. Feedback would pretty well bootstrap out C_diff, so I should just get sin(2 pi f R C_CM).

Using the loop behaviour to estimate C_diff would require quite a bit of curve fitting. Any suggestions on how to measure that?

Cheers,

Phil

Reply to
Phil Hobbs

[Oh, yeah, Rob looked at the thing - he's smarter than I am - and it looks fine. I'd like you to discuss *fast* (GHz maybe) photodiode amps at some point... 3e maybe. I'm certain we're doing it wrong.]

Open-loop, it's easy. Poke an AC voltage into one input, and mesaure the current that flows to ground from the other. That's what the Boonton does. It outputs (I recall) 20 millivolts at 1 MHz on the driven side, and has a low-impedance, high gain tuned amp and a synchronous detector on the sense side.

formatting link

formatting link

The low range is 1 pF full scale, and it has a bias input on the back, so it's cool for characterizing diodes and such.

John

Reply to
John Larkin

Thanks. (Thanks to you too, Rob.) I'll try fleshing the partly-baked idea out a bit and post something if it's interesting looking. Might be worth a short article if so--it looks a dB or two better than the bootstrapped cascode even, and it might help get round the op amp drought.

That would be a bit of a learning experience for me too--I've been a bit spooked by the guys I know who design TIA chips and so on...very fast, all 50 ohms and 20 pA 1-Hz current noise, with PDs integrated on the chip for the most part. However, there's probably an interesting niche in the 1-2 GHz range that they don't care about anymore, so I can build little prototypes without running into any more scorn and derision than usual. There's more high frequency stuff in this edition, but it's mostly reactive matching networks, Bode's bandwidth vs return loss theorem, and so on.

I'd love to know how Miteq makes those amps with the 25 K noise temperature. They work at 77 K, so they have to be FETs of some sort---they aren't parametric amps, for sure. If I can find an excuse to buy one, I might buy two and can-open the spare.

Yeah, I was afraid of that. I sort of doubt the capacitances will be the same with the amp unpowered or railed, don't you?

Cute. Looks useful, too.

Cheers,

Phil Hobbs

Reply to
Phil Hobbs

My problem with the commercial TIA chips is that their gain is so high that they saturate at sub-mw levels, and worse that they all have AGC. AGC is hell on calibrated, dc-coupled receivers. All telecom stuff, basically. We just did a pretty clean 1 GHz o/e converter, made from regular parts on boards, 2 mw full range, but the noise level is going to be horrible compared to the telecom pin-tia gadgets.

Fet distributed amps? Nf improved by the square root of some silly value of N?

Build a regular Ri/Rf inverting amp. Crank the input level and frequency up until you have significant swing at the - input. Measure the ac current that sloshes out of the grounded + input, with another opamp maybe, or maybe just a spectrum analyzer. Just make sure the rails are very stiff.

John

Reply to
John Larkin

but

lem

t

"Thanks. A couple of BF862s will do a good job at AC, true. The problem is the amount of extra crapola you have to hang on them to get decent DC behaviour. (Also one begins to wonder why one bothers with an op amp at that point.)"

Yup, it would be getting very busy.

"The basic issue is that 300K resistors are so very noisy compared to almost anything else, so you have to use really big ones to overcome their Johnson noise, which in turn leads to all sorts of bandwidth problems like this."

Hmm, could you use some sort of coorelation scheme to reduce the noise? If you had two TIA's running off the same PD current you could generate two signals Vs+Vn1 and Vs+Vn2, where Vs is the common signal and the Vn's are the uncorrelated noises from each amp. Now if I had a "common mode" amplifier, (as opposed to a differential amp), I could amplify the "common" signal and not the noise... Unfortunately the only thing I can think of is summing the two signals. Which "I think" gives me a 1.414 increase in the SNR. Is there any way to do better?

" I have a new design idea based on using all current feedback that more or less avoids this problem by making the feedback currents very quiet. (This is done by dropping lots of voltage across the resistors in the current sources.) Still a partly-baked idea at this point, but it should be possible to do a better job with a simpler circuit that way."

Ahh do tell, or post a message/link if things pan out. Are you talking about using CFA's or some other configuration? I've always wanted an excuse to learn about CFA's.

George Herold

Reply to
ggherold

This works for voltage noise, as in the two-point correlation method, which does something similar--you measure the voltage twice and cross-correlate the two measurements. It relies on time-averaging, so you need stationary statistics. Thus you can do a great job measuring noise, but not so great measuring signal.

An ensemble average, where you use 10 voltmeters, form the 45 unique pairwise products, and average those, relaxes the requirement for stationary statistics, but still measures the squared magnitude. You win linearly with the number of voltmeters, because the improvement is sqrt(#pairs) and the number of pairs goes as N**2. Note that it only works for voltage noise--current noise is a real current that comes out the input terminals of each amplifier, so they all measure the sum of their noise currents, and this will survive the averaging since it's indistinguishable from signal.

Making a linear measurement isn't possible that way, unfortunately. You can get 3 dB reduction in the circuit noise by connecting a TIA to each end of the PD, as you suggest (which is an interesting idea). You'd actually want to subtract the results, because the two currents are equal and opposite. It requires some fancy footwork on the power supplies, of course, because you lose much more than 3 dB by not reverse-biasing the diode.

quiet from a voltage gain perspective (T_noise ~ Tj/beta) but look like

150 kelvin resistors from a current gain perspective, like all ideal diodes. So if you pick a superbeta transistor and connect the PD to its base (like a phototransistor), you're stuck with 150 kelvin T_N and significant nonlinearity. On the other hand, if you enclose it in a current-feedback loop that pulls all but epsilon times the photocurrent out of the base circuit, and optimize epsilon, you can win overall noise temperature like sqrt(beta). All the feedback currents have to be way sub-Poissonian, so you have to drop a lot of voltage across the emitter resistors of the current sources, but in principle with a beta of 1000, you can make a TIA whose noise temperature is of the order of 10 kelvin, with a reasonably favourable tradeoff of noise vs bandwidth.

Vanilla CFB amps were initially puzzling to me back in the day, because their bandwidth is not a strong function of the feedback gain. It became a lot clearer when I imagined making a summing amp from a regular VFB op amp--if you put a small resistor between the summing junction and ground, the VFB amp's bandwidth becomes gain-independent as well. (CFBs do this without the horrible noise degradation that the stupid resistor would cause.)

Cheers,

Phil Hobbs

Reply to
Phil Hobbs

e
t

amp

o
d
d
y
u
n
s
r
s

"Making a linear measurement isn't possible that way, unfortunately. You can get 3 dB reduction in the circuit noise by connecting a TIA to each end of the PD, as you suggest (which is an interesting idea). You'd actually want to subtract the results, because the two currents are equal and opposite. It requires some fancy footwork on the power supplies, of course, because you lose much more than 3 dB by not reverse-biasing the diode. "

Yes I was thinking about this last night. As you say you need to hook a TIA to each end of the PD, and then you will have to subtract (and not add) the two signals. And it looks like you lose the ability to bias the PD. But I "think" I've seen the circuit trick to beat that. (Though I haven't tried it.) So the non-invering input of one TIA is grounded, and then you make the non-inverting input of the other TIA the bias voltage. This does limit your bias voltage to be less than the op-amp supply voltage, but I don't think that is much of a constraint. It does look like it might limit the dynamic range a bit....I'd have to draw all this out. It seems like one could "play games" with where the ground was defined and still retain most of dynamic range.

And Thanks for the CFB lesson... I'll think about it.

George Herold

Reply to
ggherold

It's potentially a nice solution to the +-5V restriction on many of the good FET op amps e.g. the OPA656.(*) You can run the op amps off +15/+5 and -5/-15 (suitably protected against power supply sequencing funnies, loss of ground, and so on). You can do the level shifting like this:

----------*--RFRFRF-------*---R1R1R1--- +15 | | | | | +15 | | | |\\ | | | | | \\| | | ---|- \\ | / | | \\ |< | | >-----| | | / |\\ | +10---|+ / | \\ | | /| | | -(1+R_F/R_1)*I_photo +I_bias | |/ +5 | V | | | *------ | | | | R | | R | ----- R ------ / \\ R | - | /___\\ | | | | GND | |----0 Out (plus bias) | | | | | R | + | | R ------ | R | | R | | | | | -5 *------ | |\\ | | | | \\| | | (1+R_F/R_1)*I_photo - I_bias | -10---|+ \\ | / V | | \\ |/ | | >-----| | | / |>

| ---|- / | \\ | | | /| | | | |/ -15 | | | | ----------*--RFRFRF-------*---R1R1R1--- -15

This gets you back some of the dynamic range you lose by using the +-5V op amps. Also, the current multiplication relaxes the noise requirements on the subsequent stages.

Cheers,

Phil Hobbs

Put BJT follower in the feedback path (base to op amp output, emitter driving R_F), which produces (for level shifting by driving the base with the op ampwith a BJT in series with each feedback resistor

(*) TI wants everybody to use the 657, correctly pointing out that almost all TIAs run at some gigantic noise gain on account of the PD capacitance, so that you don't have to worry about using a decompensated op amp even though it looks like a unity gain application. The thing they don't tell you is that the 657 has twice the input capacitance of the 656, so it's worse at low photocurrents.

Reply to
Phil Hobbs

On our list of things to do is a really fast DC-coupled o/e converter. One end of the photodiode goes into a blinding-fast AC-coupled amp, and the other end to a slow DC amp. Combine them later.

The real bear is to find a good way to calibrate the fast gain path. Trimpots have been demonstrated to not work!

John

Reply to
John Larkin
[snip]

I presume there's a DC path at the input to the "blinding-fast AC-coupled amp" ??

...Jim Thompson

--
| James E.Thompson, P.E.                           |    mens     |
| Analog Innovations, Inc.                         |     et      |
 Click to see the full signature
Reply to
Jim Thompson

Do you mean some way to keep the electrons from piling up on one side of the pd? Sure.

Or do you mean that the initial coupling, into the first fast gain stage, should be DC coupled? That would be fine... the photodiode current will be low, so a mmic or whatever wouldn't mind the dc component.

MMICs have horrendous input and output dc offsets, more horrendous over temperature, so it makes sense to ac couple the fast path.

We actually used opamps on the one we just finished, all DC coupled at

1 GHz net o/e bandwidth. But that's about the end of the line for opamps.

John

Reply to
John Larkin

ElectronDepot website is not affiliated with any of the manufacturers or service providers discussed here. All logos and trade names are the property of their respective owners.