Great River OpAmp

GroupDIY Audio Forum

Help Support GroupDIY Audio Forum:

This site may earn a commission from merchant affiliate links, including eBay, Amazon, and others.

Samuel Groner

Well-known member
Joined
Aug 19, 2004
Messages
2,940
Location
Zürich, Switzerland
Hi

I'm currently trying to get some inside in discrete opamps; there's a great one from Dan Kennedy on FFs page (Great River Opamp2.pdf) which I tried to understand. Let's see:

* Q1: diff. input pair (I'm great, no? :green: )
* Q2-Q4: current mirror for Q1
* Q5: current source for Q1
* Q6: another current source
* Q9-Q11: another current mirror

Now I'm stuck with Q7 and Q8; at first they look like another diff. pair; but I would expect Q8 to be connected to Q9 for this. In "Designing Analog Chips" by Camenzind I found a similar topology (page 4-6, fig. 4-12) which would explain Q7. Q8 looks like a miller compensated voltage gain stage, but this is really a hopeless guess.

Any help?

Samuel

PS: you can download the mentioned book here.
 
Q8 IS one-half of a diff pair. Its collector however becomes the source of current for the output emitter followers, and Q7 goes to the wilson current mirror whose output is the other polarity of output drive. In a sense Q8 IS connected to Q9, it's just doing so through the output stage bias network.

This is a high-performance design---very cool!

I was wondering why more people didn't use the Wilson mirror, as it has extraordinary performance---and here it is used twice!

Brad
 
Thanks, Brad, for jumping in.

In a sense Q8 IS connected to Q9, it's just doing so through the output stage bias network.

I see. The current mirror forces the two transistors to act as a diff. pair, altough on of them would love to behave different. So Q7/Q8 is the miller compensated stage, right? I wonder about the advantage over a "normal" VAS? Anyway, why do we need higher voltage transistors in this stage?

I was wondering why more people didn't use the Wilson mirror, as it has extraordinary performance.

I guess it is the required matching. In fact I wonder why Dan did not choose a four-transistor mirror, which should be even better. Maybe it's the exact reason I've given...

Samuel
 
Actually the wilson mirror works pretty well without precise matching. But its coolest feature is base current compensation (compared to the usual 2 Q circuit), so in that sense matching makes that work better. As PRR pointed out recently, transistors are so intra-batch consistent these days that you can usually rely on a same-date-code decent match in most parameters.

BTW, there are some duals that are fairly cheap from Philips and Diodes Inc. that pluck adjacent chips off the wafer and package together. Not as tight as some but damn good for cheap.

The drive of Q8 from the mirror input is not having a large effect on first glance---maybe there is a subtle reason for it---actually now that I ponder there is: it constrains the drain voltages to be equal by making Q4 operate with about zero Vcb. That makes the drift in Voffset smaller. Cool.

I'm not sure what you mean by a normal VA stage---do you mean like a single or quasi-darlington stage (also with Miller comp)? If so, the reason is probably the above about drain voltage equality.

Brad
 
Also regarding the voltage rating: the collectors of Q8 and Q9 will be swinging a bit beyond the ratings of the 4403/4401 (although most of those Q's would shrug off ~53V with reasonable Z in the base circuit).

Q7 has a fairly high Z open loop in the base circuit and sees practically rail-rail voltages, so it's not a bad decision to use the higher V part there, and then you certainly want the same part for Q8 to match. One possible enhancement BTW would be to provide some low-a.c.-Z voltage drop in the collector of Q7 to equalize its dissipation compared to Q8. This would tend to reduce the diff pair offset V and thus better adjust the Q1a and b drain voltages. Probably not worth the effort, but it's a thought.

You could approach high voltages in the input stage but only running the thing at very low (~unity) gains. I wouldn't use this as a voltage follower probably---I suspect it's not unity-gain stable with that 75pF, but that's just a guess, and the input common-mode range is a little restricted, though not by very much with those low-pinchoff FETs. Maybe I'll throw this into sim at some point and see what the predicted gain/phase is.

All in all, a very nice and clever design with attention to details.
 
I wouldn't use this as a voltage follower probably--I suspect it's not unity-gain stable with that 75pF, but that's just a guess.

In the manual he says that the two lowest gain postitions "set the gain of the amplifier at unity.". I don't think he uses a pad in front of the amp, so this would imply unity gain stability.

Any hints for the advantages of the two-diff.-gain-stages over the standard architecture? I knew that architecture from the 5532/5534, but there it is used for nested feedback which includes the output stage. That seems not to be the case here, if I looked right.

Samuel
 
I will take him at his word regards stability---I guess there's enough degeneration from the emitter R's in the second stage, and the output Z of that voltage gain node is limited by the Zobel on the output. It is nice that it only requires 75pF as that speaks to decent slew rate.

Advantages...hmmmm. One, besides the drain voltage equality is that you have push-pull drive to the output stage, hence approximate equality of + and - slew rate. When you are far from slew rate limiting (i.e., probably most of the time) the equality shouldn't have much effect. But it's nice to have. I remember Keith Johnson opining that the ear was very sensitive to asymmetrical slew rate.

Diff stages have an inherent lack of even-order distortion. But the single Q or quasi-Darlington with Miller C has local negative feedback which Self argues iirc is an advantage.

Linsley-Hood and Borbely like to do their compensation in the early stage(s) which can help to make the amp not be slew-rate limited (the large- and small-signal behavior essentially identical). We know from feedback theory that there is a small noise penalty to be paid for this at high frequencies, but these are probably way outside the audio band.

Many ways to skin the cat.
 
> I'm stuck with Q7 and Q8; at first they look like another diff. pair; but I would expect Q8 to be connected to Q9 for this.

Q8 IS "connected to Q9". Q12 Q13 make a short-circuit with constant ~5Vbe voltage drop. If you didn't have that output stage, you would just tie Q8 and Q9 collectors together and call it the output. But we want more "oomph" than Q8 Q9 give, and Q14-Q17 need about 3V of bias, so we stick the 3V between Q8 Q9 collector.

> the advantages of the two-diff.-gain-stages over the standard architecture?

Symmetry. Opinions vary on this point, but this way does work well.

It needs 2 gain stages because FETs, even these hot FETs, don't give as much gain as a BJT. You can do pretty good audio with one BJT volt-amp stage, but super audio wants more. TL071 gets by with a low-gain FET input and most gain taken in a BJT stage, but we should not ask large gain of a TL071. 990 gets 50dB-60dB gain in the BJT second stage, and gets 20dB-30dB more in an input BJT.

I am puzzled by the way Q8 gets nearly no base drive. Sure, Q7 pokes Q8's emitter pretty good. In a 2-second thought, I see a way that looks more symmetrical, while still forcing good balance. Probably if I thought a few hours or days I'd find a flaw or tradeoff with that idea, since Dan didn't didn't do it that way.

> I suspect it's not unity-gain stable with that 75pF, but that's just a guess

Taking 100Ω across the input cathodes and 75pFd feedback, it aims for unity-gain at 21MHz. It zeros to gain of 2 above 10MHz, but we can suspect that "something" (probably output stage*) is losing gain and gaining phase above 10MHz. If nothing else goes soft before 20MHz, it looks unity-gain stable.

The slew rate is, what, 3mA in 75pFd? 40V/uS? 80V/uS if slammed all the way to cutoff. Even on 28V rails, that would be ample for audio.

Noise resistance is above 100Ω. Not a good choice behind a 1:1 transformer. With 1:2 iron, NF will be over 1dB, maybe over 2dB with typical transformer loss. Very very good, plenty-good for any practical work, but not record-setting. OTOH, the FET input could run much higher iron ratios and get NF awful close to zero.

No output protection except smoke-release in the 20Ω resistors.

> This is a high-performance design

The Old Grump says: for 17 transistors (and 7 diodes!), it better be good. There's gotta be over 5 bucks of silicon in that thing!
 
It is not stable in the simulator or on the breadboard even at gain of 2 on my bench. Perhaps on a proper PCB, in a nice case, it behaves better. Usually darlington type outputs have worse phase margin than a simple output after an emmiter follower buffered VAS. Local feedback through he miller capacitor can be applied better as it is done in the 990 topology.
I think Dan has mentioned using a load isolator (inductor+resistor) to eliminate high frequency warble on the output.
 
Thanks for all contributions!

I think Dan has mentioned using a load isolator (inductor+resistor) to eliminate high frequency warble on the output.

There is a load isolator on the schematic, no?

OTOH, the FET input could run much higher iron ratios and get NF awful close to zero.

I think he uses a JT-13K7-A, so this would give the "awful close to zero"!

BTW, do have "emitter" resistors (in the diff. pair) the same noise effect with JFETs as with bipolar stuff?

Samuel
 
[quote author="bcarso"]Tamas, have you tried to see what value of C21 and R21 mght make it stable?[/quote]

Brad,

Yes, I did play around with it, but I do not recall what values yielded better stability, sorry. I deemed the design a bit complex and lost interest. One of these days I will return to it and play with it some more.

Right now I am chest deep in a Rod Elliot P66 type preamp using a direct coupled discrete opamp instead of the TL071.

Cheers,
Tamas
 
Samuel, yes.

Tamas, understood! I have some "real"work I have to get on too ;-)...

Meanwhile I was intrigued enough to verify your results in sim. I have an alternate comp scheme that looks promising while preserving the distinctive features of the design, but I want to understand why the existing one misbehaves. The thing is somewhat subtler than it looks at first, since there are two paths for the comp, one immediately around Q8 and a more global one through the input current mirror and Q7/second current mirror.

I have simmed with a low-Z feedback network setting the closed-loop gain so as to minimize input capacitance effects. I should do a new FET model to be more representative of the 2SK389, but I think, given the source resistors, that's not very critical. Infact if I up those a bit to account for the lower gm of each half of the 389 vs my SK170 model it should get very close.

Brad
 
series R-C across input FET drains
I don't understand enough to see why lowering the impedance at high freqs in this place should lower the gain of the amp? You want to place the RC circuit parallel to R12/R13, right? Is this not the exact opposite to what Jensen did with the inductors in the 990?

Samuel
 
Samuel, Deane did his inductor trick to allow a lower equivalent Z in the emitters of the bipolars at low frequencies, transitioning to higher Z at highs. Otherwise, for stability, the second stage rolloff has to be earlier and the slew rate much reduced. Jensen did this for audio band noise performance, to not spoil the superb e sub n of the LM394 pair. The FETs, otoh, already have an equivalent internal R (the reciprocal of their transconductance) and then Dan adds some external as well (which as PRR points out does degrade the noise performance some).

The comp I suggest (which does not claim to be optimal btw, just unity-gain stable---I've been working on a refinement but it is tricky) puts the primary rolloff in the input stage, with a zero after a while that flattens out the rolloff. Then the second stage starts to roll off there and, because of the much smaller C load is not the limitation on slew rate. This is sort of the Borbely/Linsley-Hood approach, and as I mentioned does have a noise drawback at high frequencies.

What one strives for, for fast settling times, is a uniform 6dB/octave rolloff. The suggested comp does not achieve this but at least steers the phase shift away from the rocks to avoid oscillation.

I'm less in love with this design as I play with it more, but there are some slight alterations that would help its dynamic performance a lot. And I already mentioned one that would help the already good d.c./low freq. performance.

Brad
 
Geez, I missed this whole discussion. Yep, it's over complicated, squirrelly, but sounds good.

The whole thing was done by trial and error and enormous beer consumption through the winter of 1991.

When I was totally clueless, by the way.

But one day, it just sounded right, and stayed stable throughout. There are a couple of thousand channels out there by now. No special screwing around to make them work.

By the way, this preamp is essentially discontinued, if anybody is interested I could make gerbers of the board layout and parts list available.

Teach me more about stability, eh?? Math is my sore point...
 
[quote author="Dan Kennedy"]Geez, I missed this whole discussion. Yep, it's over complicated, squirrelly, but sounds good.[/quote]
Me too, I was in the UK at the time...[quote author="Dan Kennedy"]By the way, this preamp is essentially discontinued, if anybody is interested I could make gerbers of the board layout and parts list available.[/quote]
Rock ON!!! :thumb:
[quote author="Dan Kennedy"]Teach me more about stability, eh?? Math is my sore point...[/quote]
You've got a great set of ears, that's all I need to know! :thumb:

Keith
 
Heh. I don't think it is overly complicated, and if that's what you got with trial and error and beer, think of what some informed design with a first-rate and properly aged proprietary red would do :grin:

And I agree with Keith: the ears have it.
 
Thanks Dan for joining in!

To be honest, I did not understand the circuit in it's last detail - I hope I can ask my Qs as this thread got warmed up? What is the function of Q12/Q13? Is this a bias generator for the output stage or another voltage amp stage? Or maybe just a unity gain output driver?

What about the feedback network? Recommended value of the feedback resistor, any cap in parallel with it? At which minimum gain does this thing run?

if anybody is interested I could make gerbers of the board layout and parts list available.
I'm sure that absolutely nobody would be interested in this! :grin: Seriously, I'm sure people here would be very happy with you offer! Thanks!

Samuel
 

Latest posts

Back
Top