fast ramp follies

hat

his

It's a CCS, so the collector load's the ramp cap. I guess that dominates. It's oscillating, so there's feedback somewhere. From another guy I'd ask about magnetic coupling in the layout, but I doubt your layouts would do that. That's why I figure it's in the transistor itself, and why it changes with Q1.

Usually, yes. Here I was thinking the upper (d.c.) transistor could be slow, the cascode transistor fast.

Wirebond stuff is usually well above the 70Mhz oscillation of the 1st bjt you tried.

Hopefully minimal, with SMD layout.

-- Cheers, James Arthur

That was me. They don't generally oscillate, meaning that's not a general, usual result. They certainly can--I made a cascode oscillator-frequency multiplier once. It worked amazingly well.

Hopefully base inductance is minimal. If L(base bond wire) is a typical SOT-23-ish 1.5nH, that resonates with 10pF at 1.3GHz. I assumed that's not the feedback mechanism, since John's first transistor oscillated at 70MHz.

If the op-amp output is some complex impedance, which it well could be, that falls apart.

If that's the mode, a base resistor would kill it, agreed?

-- Cheers, James Arthur

The -6 rolloff is usually done somewhere deep inside the chip. At 100 MHz, the opamp gain as such is very low and the rolloff is a lot steeper than -6. I think the transistor oscillation is more dependant on wirebonds and esd diode capacitance and stuff like that, nothing that an opamp Spice model is likely to sim usefully.

Fast transistors used as classic emitter followers (stiff bypassed base and collector) tend to oscillate. They don't need additional inductance. A base resistor will kill the oscillation, but an opamp output impedance usually won't. Somebody with a VNA could measure the impedance of some opamp outputs in the hundreds of MHz... that would be interesting. I suspect the low frequency loop (follower, inverter, whatever) wouldn't change things much at high frequencies.

Maybe I'll TDR some opamp outputs. I don't have a VNA but I can do that.

Yeah, sounds like a cascode has about the same hazards as my current source transistor, except that the cascode transistor isn't connected to the opamp output, with its own unknown complexities.

If you terminate them as you're supposed to, you only get 0.4 volts single-ended swing, not enough. The NE3508/9 can be driven nicely from a full-swing ECL (eclips lite, eclips plus) gate, with margin to enhance a bit (+0.2 to -0.6 maybe), or full-swing ECL can drive the neat Avago enhancement parts.

MC10EL/EP89 swings about 2 volts, which can drive meatier depletion fets that want more swing. Beyond that, it's hard to get fast edges with volts of swing.

I took a bunch of data on the NE3508/9 if you're interested.

Bummer that the RF boys usually furnish pitiful DC/time domain specs.

ta

Is this the one you're thinking of?

-- Cheers, James Arthur

ost uses.

What might work is a cascode with a capacitor to ground on the middle node; that would look weird. Transistors are less likely to oscillate if their emitters are at RF ground. The base could be bypassed, too... no soggy Miller effect.

That one is worth thinking about. All it needs now is a cute way to fold the cascode base current error back into the loop.

The beta feedback trick is cute, but it relies on the op amp to compensate, and so only works at low frequency. It also doesn't fix Early effect at high frequency, whereas a properly designed cascode does.

Cascodes do double the base current error, and at low supply voltages, you have to reduce the voltage across the emitter resistor of the current source by a V_BE drop, which hurts temperature stability as well as initial accuracy. (RF transistor data sheets rarely mention V_BE, and IME never specify it, so you have to assume that it might vary by 50 mV or so.)

PNP RF transistors have fairly horrible betas--the BFT92 is 20 minimum,

With matched transistors, e.g. the Intersil HFA series, you can do very well using a Wilson current mirror. It compensates for beta exactly to leading order, and that works well up to frequencies of order f_T/5 or something like that, where you have to worry about phase shifts and second-order effects. With a beta of 20, its error is only about 0.2%. Unfortunately those arrays cost $5 or so.

The NPN/PNP series-shunt pair (the one that folks with nothing better to do are are slagging each other off about at the moment) runs the driver transistor at constant I_C, and lets you use a much faster NPN to stiffen the PNP at high frequency. Try simulating it--it's really good medicine.

(I spent most of July up to my ears in high performance laser noise cancellers, which rely on a lot of those sorts of tricks, so I'm pretty well up on all this at the moment.)

Cheers

Phil Hobbs

he

beta

is

most uses.

We aren't slagging one another off about the complementary Darlington as such (or whatever you've chosen to call it) but about the right name to use for it. If you give Google "complementary Darlington", "Sziklai Pair" appears at the top of the list. "NPN/PNP series-shunt pair" doesn't do as well.

"Sziklai pair" is probably the most direct reference, but it doesn't mean much if you haven't run into the patent.

Just as voltage regulators rely on bypass capacitors at high frequency, I'd think of a series inductor to keep the current source happy at high frequencies. I wonder, though, how far one could get with B-E capacitance (with a HF transistor that has relatively small B-C capacitance)? Either way, oscillation is possible and layout will matter.

Early effect should only add repeatable curvature, I'd be concerned only because it isn't usually a well-tested parameter.

A b-e cap on a fast RF PNP, inside my slow opamp loop, is interesting. I could get rid of some other parts, maybe. The cap could be pretty big, because it doesn't do any harm to the opamp loop dynamics.

How's it gonna oscillate with the base shorted to the emitter? [1]

So, while 70 MHz seems lowish, it's not impossible.

Yes. But, even at 100 Ohms, the margin would not be really large.

And it adds to Rbb, which is the primary noise generator of the transistor. (internal Rbb' = 5R typ.) Ok, in a cascode that would have to be divided by beta.

It's really tough to get the sort of SRF you'd need to do it with a real inductor. A good RF transistor has an output capacitance of well below

One interesting trick for this is to use a simulated inductor, i.e. a one-pole capacitance multiplier floating--like so:

--> I_in

With another resistor across the capacitor, you could use a depletion pHEMT and get another order of magnitude in bandwidth and probably two orders in Q.

An RC would be a repeatable curvature too, and would have the additional advantage of being known *a priori*, as opposed to Early effect, which is actually not linear and at least as variable as beta. You're right, though, the Early curvature would be ~10x smaller.

Fun stuff.

Cheers

Phil Hobbs

I'm doing a bunch of that at the moment too, but only at low frequency. It would be really great if analogue muxes were as good as a handful of

Cheers

Phil Hobbs

Am 15.08.2012 22:15, schrieb John Larkin:

Usually in the VAS stage, but the output of the OpAmp is where the feedback is taken from, so loading that node will have a first-order effect.

A few octaves at -12 would guarantee > 180 deg at unity gain. Then we do not have to search for other culprits.

Well, the output impedance of an OpAmp is meant as max. Ohms worst case, so a nice op amp will feature less...

Somebody with a VNA could measure the

Aaaaargh... exploding daytime schedule, and you drop sth. interesting b4 my feet... Preferred type, conditions? May take a week!

:-)

Yes, I got the promised double ECL swing from SY100EP89, but the ADCMP580 was so much faster :-)

I have played with the ADN2525 laser driver, I would prefer the 2531, but digikey does not stock them. Should yield

YES YES YES!!!

Gerhard

I like the pseudo-Darlington, I'm just leery because I had those scream wickedly once using fast transistors. I don't think I ever fixed it.

It has a bunch of advantages here, including reducing the base current i(b) error.

Here's a skeleton idea for a Bloggs-Hobbs hybrid:

Vcc Vcc -+- -+- | | | .-. R1 | U1 | | 120 | LM385 '-' | -adj | .---' | A

R2's a pathetic provision for isolating U1 from the nS stuff. The whole thing needs stabilizing against UHF oscillation, these are just the bones.

-- Cheers, James Arthur

No luck on that link so far.

I know of a part that will swing 0.8 volts in 40 ps, and it's almost affordable, roughly dinner for two with beer.

Hittite and Inphi make really fast flops and gates. Some of the Hittite parts have 12 ps edges... for $500.

We are using the ADN2871 as a driver for, umm, other stuff, not a laser. It was hard to convince them that we were worthy of using it. We are DC coupling into its data inputs from 2.5 volts CML. It's a "single loop" part so is somewhat less weird than some of their other drivers. Maxim has stuff, too.

See if this works...

It wouldn't need much more than that, I don't think. Series-shunt pairs have poor stability when the transistors are about the same speed, due to having two lags in the loop, but the BFP640 is ~10x faster than the BFT92, which helps a lot. (Reducing the BFT92's collector current makes the difference wider and helps with stability.)

I used the BFP640 as a cascode for an ATF38143 pHEMT in a very quiet charge-sensitive preamp (noise floor ~10 electrons per inverse bandwidth), and all it needed was an 0402 bead in the base--no worries at all.

Cheers

Phil Hobbs

Back in The Day, there was a thing called a Reactance Modulator. This was typically a pentode hung off the local oscillator tank circuit, with a relatively large resistance coupled to the grid. The remaining grid capacitance (which is almost exactly G1-G2 and G-K capacitance; pentodes have G-P capacitance enviable even by PHEMT standards!) acts against the series resistance as an almost exact 90 degree phase shift, thus the tube looks like a reactance (exercise for the student: prove this relation and what phase shift it produces). By varying the gain (by varying bias), the reactance is changed, and therefore the effective inductance (or capacitance) of the tank circuit.

It's like that, but with more phase shift ;-)

Tim

P.S. Had some nice local craft beers this evening. I'm happy ;-)

Happy or hoppy? Most "craft" beers are way too bitter for my taste, and most gourmet red wines are way too tannic. Just got some nice wheat ale that my engineers brewed at home... gotta try it.

Hey, how about this?

Ludicrously simple. No Miller nonsense. Might not oscillate. MMBTH81 would have about 2 pF out, or BFT92 would have about 0.75 but lower beta.

If this works, I should have done it years ago.

I like darker things, there's a lot of good stouts, porters and ales around here. Hops can be good and bad. Something like, say, Heineken, is very hoppy but also light and drinkable. A lot of lagers are excessively hopped and unappealing.

I don't see a cap helping you there. Ccb puts signal directly on the op-amp, coupling the active node to its nonlinearity, and Cbe in turn couples this to R_E, which puts a (second degree) resistor back in the circuit. Might as well use a ferrite bead, but you already know how sloppy that works out.

Better to use a cascode, where the "outer" transistor is as close to the active node as possible, with as little base resistance as possible (just enough to keep it stable), so you get Ccb incorporated as closely as possible into the node. At that point if you still need temp or bias compensation, you won't be any worse off, and you'll most likely be much better off than any other method.

Tim