Gain structure in a cascaded-gain microphone preamplifier

Hi all,

In designing my discrete mic pre I've run into some difficulty in determining the gain structure, and I'd like to ask for your input.

First, this is what most common microphone preamplifiers look like:



The core is a variable-gain amplifier, where the gain is usually varied by changing the amplifier's feedback resistor network. While this does work, one of the problems with this approach is that it sounds differently in different gain settings. This is because the change in feedback also (generally) impacts the amplifier's closed-loop frequency response and thus its distortion signature. In addition to this the variable part of the feedback network is necessarily connected to rather sensitive nodes in the amplifier, and the wiring to a gain switch or pot can easily pick up noise.

A cascaded-gain mic pre avoids this by having fixed-gain amplifier blocks with variable attenuators in between:



(Note that I don't claim this to be a new approach. Some of the oldest amplifiers worked like this, and this method has been discussed here a few times in the past, like in posts by pstamler, bcarso and Samuel Groner in this very interesting thread, and here by PRR).

But now the question is: what should the gain be for the amplifier blocks, and what attenuation range do we need for the attenuators?

We don't have enough information yet to be able to say anything. What should the gain range be? For the time being I've picked 20-60db. Other than that we want the usual goodies: high dynamic range, low noise, particularly in low-gain configurations.

I've already picked a few numbers out of the air. I want to be able to use common low-noise SMD transistors, most of those have a maximum VBE of ~45V. So call the maximum output swing of each amplifier stage +/-20V, or 14VRMS, or 23dBV. Most common transformerless mic pre recipes are differential on the inside, let's keep the entire pre differential, just for laughs. Differential swing is double the single-ended swing, or 29dBV.

Now, from my own experience and from what others have posted here, it would be nice if the pre could handle occasional 1VRMS (ie 0dBV) signals. To not clip for such levels the gain of the first stage should be no more than 29dBV-0dBV=29dB. Then again, noise theory says that if it has sufficient gain the first stage will dominate noise performance, so gain shouldn't be very much lower than this.

The other two (fixed) gain blocks are both preceded by a variable attenuator. There is little point in having a minimum attenuation higher than 0dB, as that would be wasting gain in the highest overall gain settings. On the other end of the scale the maximum attenuation should at least be equal to the gain of the following gain block. If it would be any lower, then for strong signals the second (or third) stages could clip even when the first one doesn't, and that would only waste dynamic range.

My current plan looks like this:



The final attenuator scales the output level of the third stage back to normal studio levels. As said before the maximum unclipped output of the gain blocks is 29dBV or a little over 31dBu into 600R; the 6dB attenuator leaves a more manageable 25dBu. Full gain range is 20-60dB, as specced.

How does this look so far?

One of the issues left to be decided is the attenuator impedance. Too high and it will kill low-gain noise performance, too low and the attenuators will function as space heaters. While I have sim-tested the gain blocks with a 200R differential load my current plan is to have the attenuators be 600R. At that resistance a sine that's just below clipping will have each attenuator dissipating just over a Watt, and the total noise figure in the lowest gain setting would be 4-5dB. This can be improved somewhat by implementing Samuel Groner's suggestion to switch out the third gain block for gains of 40dB and lower, but that would probably affect the sound of the entire chain somewhat.

JDB.
[Of course, the output attenuator could just as well be a 4:1 transformer]

I need to spend some time reading and digesting your post but I've often thought about a similar setup.

You might not need to necessarily have equal gain stages and attenuation stages chained together.  I would think you would want more gain at the front and less gain as you move through your stages BUT I would increase the headroom through each stage so that you don't saturate the next one at full gain from the previous one.  

I've noticed this on a couple designs, especially the Hamptone type pre design that floats around here.  Ive found that it's crunchy sound actually comes from the first stage, which is a fixed gain, saturating the second stage(also fixed gain) when the volume/gain pot(voltage divider between stages) was set close to all the way up.  While most would say that I simply turned the volume too high, I found the preamp hard to use since I was losing the dynamic range that I needed to do various things like input instruments.

I got around this by finding the compression point of the first gain stage and then using this as my reference point for input to the second stage.  I then increased the voltage rails and then decreased the gain of the first stage back to where it would be without the increased rails.  This at least keeps the second stage fully usable in most cases I find myself using this preamp in.

I suppose this is fairly relative to how you plan on using this design but I think it gives slightly more control.

Of course, the output attenuator could just as well be a 4:1 transformer.

Click to expand...

I guess you mean 2:1. In fact the output pad is where I see most problems. If you want low output Z the amplifier load quickly gets nasty again. I'd rather lower the gain of the last stage and forget the pad.

A minimum gain of 20 dB will make many users apply external input pads--I think you should include a 20 dB input pad on the PCB.

Samuel

I also wonder if someone couldn't try a parallel approach with gain stages instead of a serial chain.  Kind of like a summing network.

> the second (or third) stages could clip even when the first one doesn't

You could cascade power rails. First stage 30V, second stage 300V. Now there is "never" a reason to drop 20dB between stages to avoid internal overload. (You may need it to avoid smoking your tape deck.)

Older tube PA amps often had the first stage on +90V, +380V on the final driver. The input*gain would never need over 90V supply, though the real reason was high ripple rejection in low-cost (small-C) rail filter.

IMHO, the output pad JUST to tame the level is stupid. (Output pads are valuable to buffer strange loads.) You are saying 2X the voltage, 4X the heat, and throwing most of it away.

Studio mike preamp gain "can" be less than zero dB. Hot mikes on insane percussion into consumer-level inputs needs loss. Stupid. But we can't omit the preamp because it has the Phantom and stuff. Yes, we could use a simpler scheme, but we need a preamp for harpsichord, and why have multiple boxes?

OTOH, at some point you must admit that one box can't do everything.

There is not an obvious way to gain-change from very high through zero to loss while also giving lowest noise. You have to plot your *actual* problem and pick your choices. In many studio cases, low-noise is irrelevant when you have 1V at the mike. We drop in an input pad, leave gain in the preamp, "add noise" which nobody can possibly hear.

20dB first stage gain "can" dominate the noise structure in most real cases. 26dB does not seem like much more, but at 1V signal it gets above the easy limit of penny transistors.

Note that in any non-dangerous case where you have >1V signals, you find a Condenser mike with an internal amplifier. The actual "first stage" is inside the mike. And the output hiss is often higher than a dynamic. You could (Germans did) design a condenser input as a high-gain "Line" input, rather than a lowest-noise ribbon/dynamic input.

Get some 2,000V transmitting triodes. They can be run at high Gm, low thermal hiss. Fix-gain of 60dB, 1,000V output. Put the fader here (wear rubber shoes). Buffer with small unity-gain line amp.

>>>> OTOH, at some point you must admit that one box can't do everything.

what if try "divide and conquer" approach?

you have 2 gain + output stage.

you could use 2 gain stages for a quiet/normal levels
you could use only the 2nd gain stage for extra-hot mikes or an instrument input.

you could use some fancy switches on the frontplate

your marketing dept. could write some orgasmic stuff on NOVEL APPROACH and EXTRA VERSATILITY and who cares what else...

you could also have a dual output stage and use them as "doubled" outputs for extra multing tricks, when using both gain stages on a single mic on a quiet source

when switched into "dual" mode, your pre could transform into a dual channel pre for two mikes on hot source like drums or guitar

you could also have a dual DI
or a hot mic + a DI channel

and a fancy front panel with smart graphics. and pretty switches.

Svart said:

You might not need to necessarily have equal gain stages and attenuation stages chained together.  I would think you would want more gain at the front and less gain as you move through your stages BUT I would increase the headroom through each stage so that you don't saturate the next one at full gain from the previous one.

Click to expand...

Assume the second and further gain stages are the same amplifier but with different feedback networks. For the same amount of combined gain you get the lowest level of third and higher order distortion if the gain is shared equally between the stages, as nonlinearity increases faster in the amp with the highest gain (and thus the least feedback). In other words: IP3 decreases faster than gain increases. Noisewise I fully agree that when you set a given gain level you want less attenuation in the first attenuator than in the second one.

Samuel Groner said:

Of course, the output attenuator could just as well be a 4:1 transformer.

Click to expand...

I guess you mean 2:1.

Click to expand...

Right. I actually meant to say 2400:600.

Samuel Groner said:

A minimum gain of 20 dB will make many users apply external input pads--I think you should include a 20 dB input pad on the PCB.

Click to expand...

Would that be to increase the input clipping point higher than the +5dBu where it is now, or simply to get a less-hot signal at their recorder? In the latter case it might make more sense to expand the range of the second attenuator.

(I can see that a pad might still be useful. For now I'm trying to get the core of the mic pre nailed down, further down the line I have a stackable relay board planned to handle pad, polarity, phantom and the attenuator networks. I want to keep the heart of the mic pre affordable and easy to build, and offer the relay board as an optional extra. Still for the purists who want proper simultaneous switching of both attenuators a fistful of relays may be cheaper (and easier to wire!) than an appropriate Elma or Grayhill switch)

PRR said:

Note that in any non-dangerous case where you have >1V signals, you find a Condenser mike with an internal amplifier. The actual "first stage" is inside the mike. And the output hiss is often higher than a dynamic. You could (Germans did) design a condenser input as a high-gain "Line" input, rather than a lowest-noise ribbon/dynamic input.

Click to expand...

Yes. I can see that a user who gets a +5dBu input to clip might not care that much about having a single-digit noise figure. Still, counting on my fingers it looks like this pre should manage ~100dB dynamic range even with a 20dB pad engaged (versus just over 120dB at the lowest gain setting).

PRR said:

IMHO, the output pad JUST to tame the level is stupid. (Output pads are valuable to buffer strange loads.)

Click to expand...

Samuel Groner said:

In fact the output pad is where I see most problems. If you want low output Z the amplifier load quickly gets nasty again. I'd rather lower the gain of the last stage and forget the pad.

Click to expand...

I didn't give the full story on the output pad (PRR hints at some of the other bits). Getting the level down is only one reason for having it. The amp needs buildout resistors, as it's stable for capacitive loads up to a few nF but not much more. While I've added current limiting on my in-progress schematics, it would be nice if nothing were smoked when either or both outputs are shorted. Finally it would be nice if loading doesn't impact the amp's performance too much.

All this is satisfied by having a simple 100Ω-200Ω-100Ω U-pad on the output. Differential output Z is 100Ω, not very different from when using regular buildout resistors. The loading is less than half the design load for the output, output stage dissipation stays within 10% of what a plain 600Ω-load draws, and when one of the outputs is grounded by being connected to a single-ended input the system still works.

(Having said that, the attenuators are all off-board, either on front-panel rotary controls, on the aforementioned relay board or on the output connector for the output pad. Don't like 'em? Don't populate 'em! At this point I'm merely trying to figure out if there is a useful configuration possible for the board that I'm laying out, not trying to dictate the One True Way)

Thanks,

JDB.

EDIT: Forgot to say, but all things being equal a 6dB output pad also gives better low-gain noise performance than either increasing the attenuator in front of the last amp or reducing the last amp's gain.

> a 6dB output pad also has better noise performance than either increasing the attenuator in front of the last amp or reducing the last amp's gain.

For any significant overall gain, the noise is in the input, not the output.

Say the input hisses like 100 ohms. And your output pad hisses like 100 ohms, with 6dB loss. If you have more than 6dB gain between, the final hiss is mostly input hiss.

Yeah, if you put in real values like 150 ohms mike, 30 ohm first stage parasitic, 101 ohm pad output, the answer is a bit different. And there are <0dB gain situations. But I don't necessarily see this as a Good Idea.

There are other reasons to pad. And a lot of 10-Watt p-p tube amps rated +32dBm (1.5W) output because they pad. Then you can hang any darn thing on the output and "get sound". Miles of phone line, all your recorder inputs in parallel, a dozen radio-network links. But I'd say such oomph is an OUTput function, and a mike amp ought to be optimized for INput functions, and not be expected to pull trucks.

One rule of thumb or generality about low noise design worth mentioning is system S/N is generally dominated by the first gain stage's S/N. Cascading multiple gain stages may improve performance when using gain blocks with limited open loop gain bandwidth but there are popular topologies where open loop gain increases along with increased closed loop gain. 

I don't suggest good noise performance can't be accomplished this way, only that you need to be more diligent about the noise floor in the stage(s) following initial gain stage, etc.

JR

Forgot to say, but all things being equal a 6dB output pad also gives better low-gain noise performance than either increasing the attenuator in front of the last amp or reducing the last amp's gain.

Click to expand...

That's a trick we use to fudge noise figure through gain blocks..   ;D

I think you will agree with me that with this arrangement there are a lot of possibilities for bad combinations that could lead to interstage clipping or less-than-optimum S/N.
The advantage with the single-stage one-knob configuration is that there is almost no room for operator mistake.
The almost universal answer to cascaded arrangement is to make the gain switchable using a multi-wafer rotary switch or logic-controlled relays. I suspect that's what you'll have to do in the end.
Then, why not considering a partial bypass arrangement à la Neve 1084/Schlumberger UPS, where an ultra low noise preamp with ca. 20dB gain is inserted before a more "pedestrian" pre for the highest gain positions?
It is tempting to find a "one size fits all" type of circuitry, but when you look at all the (often conflicting) constraints, I reckon it's the only approach that guarntess best performance optimization in all circumstances. Now I don't say it's the most practical, or cost-effective...but you seem to be keen on performance.

http://www.gordonaudio.com/

IIRC there was a thread here about preamp design kind of like this thread some time ago.

I think you will agree with me that with this arrangement there are a lot of possibilities for bad combinations that could lead to interstage clipping or less-than-optimum S/N. The advantage with the single-stage one-knob configuration is that there is almost no room for operator mistake.

Click to expand...

He want's to gang the two variable pads.

IIRC there was a thread here about preamp design kind of like this thread some time ago.

Click to expand...

Indeed, linked in the first post of this thread .

Would that be to increase the input clipping point higher than the +5dBu where it is now, or simply to get a less-hot signal at their recorder?

Click to expand...

Probably a combination of both in many cases, but more often the later.

The core is a variable-gain amplifier, where the gain is usually varied by changing the amplifier's feedback resistor network. While this does work, one of the problems with this approach is that it sounds differently in different gain settings. This is because the change in feedback also (generally) impacts the amplifier's closed-loop frequency response and thus its distortion signature.

Click to expand...

I agree with JR that this is only a half-truth. As he mentioned a standard current-feedback topology does pretty well with ~constant loop gain over a 40 dB range, and then there are distortion mechanisms which are not inversely proportional to loop gain. And with your approach the distortion behaviour ain't consistent either as the first two stages are operated at different levels with varying gain.

Samuel

Samuel Groner said:

I think you will agree with me that with this arrangement there are a lot of possibilities for bad combinations that could lead to interstage clipping or less-than-optimum S/N. The advantage with the single-stage one-knob configuration is that there is almost no room for operator mistake.

Click to expand...

He want's to gang the two variable pads.
Samuel

Click to expand...

I haven't seen that expressed clearly...
Anyway, the issue is: should he lose the possibility to run the first stage at 40+ dB gain, which is the key to achieve a really nice NF?
I reckon 3 stages is too much and refusing the use of variable feedback for adjusting gain is counter-productive. And using an output attenuator is not the nicest thing either in terms of interfacing with the outside world.

should he lose the possibility to run the first stage at 40+ dB gain, which is the key to achieve a really nice NF?

Click to expand...

Noise figure is really defined as the amount of noise that the amplifier *adds* to the signal, not how much it amplifies the noise floor or other noise components present on the input signal.  Generally the noise figure is dominated by the input portion of an amplifier(like a front end FET or BJT) because the subsequent stages simply amplify that noise as if it were also part of the signal.

We've sort of adapted the concept of Noise Figure here because, like an opamp/gain block, this has multiple stages where noise introduction into a certain stage will cascade through the rest and JDB's design needs it's lowest noise stage first.

I personally don't like variable feedbacks because they can present problems when you use them as they can cause changing input impedance, especially in current feedback networks.

I personally don't see the problem with integrating a small pad on the output of the amp since as JDB stated, you need buildout resistors for protection already and you can incorporate the use of a pi-pad to keep a known minimum load on the output stage and adjust the impedance seen to the outside world through this if you see fit to.  In RF I would be doing this to adjust for reflections and line impedance and I've seen this happen at low speeds too regardless of who says that audio frequencies don't need good terminations..

If you can guarantee that your stages are quieter than you need then I don't see a problem actually using the lowest gain section first.  With less gain/feedback you can definitely keep noise currents down, possibly voltage noise as well if you work the section right IMHO.

Svart said:

should he lose the possibility to run the first stage at 40+ dB gain, which is the key to achieve a really nice NF?

Click to expand...

Noise figure is really defined as the amount of noise that the amplifier *adds* to the signal, not how much it amplifies the noise floor or other noise components present on the input signal.  Generally the noise figure is dominated by the input portion of an amplifier(like a front end FET or BJT) because the subsequent stages simply amplify that noise as if it were also part of the signal.

Click to expand...

If you have a problem with the concept of noise figure, let me put it differently; a transformerless mic pre has to have a very low gain setting resistor if you want to achieve a good noise level. For a nominal 200ohms source impedance, achieving -128dBu EIN, the total parasitic resistance must be lower than 40ohms; that include 2x Rbb' and the resistance of the feedback network seen by the input emitters. In practical terms, it means the gain adjust resistance must be ca. 20ohms. If you use opamps with standard output capability (5534 type) that imposes a gain of 30dB (2x 300ohms FB resistors and 20 ohms emitter-to-emitter). You need 36dB gain to achieve -129dBu EIN.

I personally don't like variable feedbacks because they can present problems when you use them as they can cause changing input impedance, especially in current feedback networks.

Click to expand...

It's true but these variations are two orders of magnitude below the nominal impedance; I don't see how they could be of any practical significance.

I personally don't see the problem with integrating a small pad on the output of the amp since as JDB stated, you need buildout resistors for protection already and you can incorporate the use of a pi-pad to keep a known minimum load on the output stage and adjust the impedance seen to the outside world through this if you see fit to. 

Click to expand...

I would rather reduce the gain in an active manner, by increasing feedback. The changes in sonic signature created by this NFB increase is probably of the same order as using an attenuator with a non-controlled outside world impedance. It is extremely easy to achieve 100ohms output impedance from an active output. A 6dB attenuator presenting 100ohms output impedance to the outside world AND relatively stable load to the output stage WHATEVER the load is, is not very simple and presents challenges to the ouput stage in terms of power.

In RF I would be doing this to adjust for reflections and line impedance and I've seen this happen at low speeds too regardless of who says that audio frequencies don't need good terminations..

Click to expand...

Practically, there is no control of the final termination, it can be anything from 600R to 20k, purely resistive or reactive...Whatever solution, pattenuator or just protection resistors, will be a compromise.

If you can guarantee that your stages are quieter than you need then I don't see a problem actually using the lowest gain section first. 

Click to expand...

That would mean that all stages should have at least ca. 36dB gain.

With less gain/feedback you can definitely keep noise currents down, possibly voltage noise as well if you work the section right IMHO.

Click to expand...

Less current usually means more noise voltage.

abbey road d enfer said:

[...] a transformerless mic pre has to have a very low gain setting resistor if you want to achieve a good noise level.

Click to expand...

The proposed first stage has a fixed 33R gain-setting resistors and 330R feedback resistors; estimated noise figure rel 200R is between 0.9dB and 1.8dB depending on whether you use hard-to-get low noise transistors (2SB737 or similar) or run-of-the-mill 2N4403s. The first stage can handle lower gain/feedback resistors, but 33R/330R looks like a reasonable balance between noise and dissipation-induced distortion.

I wonder if many of us regularly run into situations where a 0.9dB NF works fine but a 1.8dB NF would be totally unacceptable.

abbey road d enfer said:

[...] A 6dB attenuator presenting 100ohms output impedance to the outside world AND relatively stable load to the output stage WHATEVER the load is, is not very simple and presents challenges to the ouput stage in terms of power.

Click to expand...

I'm not sure what you mean here. The proposed second/third stage was designed to drive 200R differential loads with minimal distortion, I fail to see a scenario where a load connected through a 100R-200R-100R U-pad would make the output stage unhappy (if you do know of such a scenario, please let me know now while I still have a chance to improve the proto and its layout).

JDB.
[sleep now, noise graphs tomorrow]

> there are a lot of possibilities for bad combinations that could lead to interstage clipping or less-than-optimum S/N.

Some people can't handle one knob, yes.

> He wants to gang the two variable pads.

But maybe not the normal ganging.

Mark-up the diagram as G(=gain cell) and P(=pad/pot).

G1-P1-G2-P2-G3(P3)

For max gain both P are no-loss.

For less gain, turn-down P2.

But if it is lower than 20dB, G2 could clip before G3. So stop P2 at 20dB loss.

But allow the knob to keep turning, only now it leaves P2 at 20dB, and turns P1 from zero to 20dB loss.

There is probably an elegant mechanism to do this, perhaps with slip-clutches. Conceptually, you could glue a slide-pot to the knob of another slide-pot. If each were 4" throw, you must move 8" to cover the range. But I see problems making the correct pot turn/slide as you go down and then back up.

For the values he proposes, a simple ganged pot could work... it's only 20dB per stage. When you get into RF/IF stages you may have 40dB gain per stage, and need to lose most of that on strong local stations. "Better" TV and sonar systems do stagger the gain control. They fade the last stage, but leave it hot enough to handle large signal. Then they fade the middle stage, before fading the input stage.

I do agree that "change of sound" IS inevitable due to signal-level differences.

> If you can guarantee that your stages are quieter than you need

That never happens. There is always some situation which can't be "quieter than you need".

> I don't see a problem actually using the lowest gain section first.

In a low-signal amplifier, the first stage is LEAST likely to clip. I don't see why it has to be 36dB; often 10dB-15dB is ample to dominate the next stage noise. But if you need a 36dB stage, seems to me it makes sense as the first smallest-signal stage.

And yes, my 1V percussion plus 36dB pops the tops off your dime transistors. There is NOT a universal structure.

Rule of thumb for lesser systems; see where it leads you.

The sensitivity at "the gain knob" should be 50mV to 500mV. Put enough gain after so that this levels just-clips. Put enough gain before so that your weakest source's peak will make this level.

What does this mean? Take 100mV target. Assume your post-pot amp has 1uV hiss (10K pot and 5532). Here we have 100dB S/N. Traditionally this was ample. If not, you can pick a higher target or find a lower-hiss post-amp.

Taking 100mV target and 1mV minimum source, we need gain of 100 before the gain kno.

Now look at maximum source. This is the problem: far-mike ribbon on harpsichord is VERY different from close condenser on angry percussion. In the old days, I would rarely get hit with over 30mV. Gain of 100 put this at 3V. A 12V rail can cover this easily.

Work the other way. Say you have 1V mike signal. Say your final product does not need over 100dB S/N. Your allowed noise level relative to 1V is 10uV. A "good" input is 0.2uV. It is fully valid to insert a 50:1 pad. A 10:1 pad is certainly not wrong when you have 1V banging on your XLR.

There's many ways to do it. You should investigate the ways less traveled. But my intuition says this way has issues which may be less happy than other ways.

I assume you are allergic to transformers. They change some trade-offs. For example, a 1:4 ratio brings a 150 ohm source up to a happy impedance for a low-cost transistor. OTOH, the same iron turned around 4:1 can turn a too-hot 1V signal into a more practical (for low-noise input) 250mV signal. Yes, the 4:1 connection whill not be lowest-noise. But we already showed that noise is not a big deal when you have 1V signals. Using this trick, 1mV to 1V signals can be brought into the 4mV to 250mV range, a much less stressful proposition. (True, you can't smoothly slide through the entire range of gain, you must stop and flip/switch the tranny; we accept such a complication with our 20dB pads.)

If you have a problem with the concept of noise figure, let me put it differently

Click to expand...

?

I was explaining to you what noise figure is..

Noise figure is simply the negative change in SNR through a device regardless of external feedback.

NF=SNR(in)/SNR(out)

What you are talking about is entirely different and is made up of a number of different noise sources acting as one.

Less current usually means more noise voltage.

Click to expand...

Yes, usually, but if that was always the case then we wouldn't be seeing low voltage and current noise amplifiers like I work with everyday.  What I simply meant is that JDB could very well play with his design to be both low current noise and low voltage noise but I don't think either will be necessarily large enough to worry about in this design.


I suppose you could also do hybrid stages of VFA/CFA to optimize noise.  

ah but there are millions of novel ideas just waiting to be found out there.  

JDB, did you have an exact idea of where you wanted to end up with this design or is this more of a "lets see where it goes" kind of thing?

these variations are two orders of magnitude below the nominal impedance; I don't see how they could be of any practical significance.

Click to expand...

For some reason I thought we were talking about current feedback amps.  Anyway, try changing the feedback on a current feedback opamp and see how it likes it.  It doesn't.  You start changing it's transimpedance curve but it's not quite like a voltage feedback opamp's GBP since most CFAs are designed to work optimally around a very finite feedback value and you could easily burst into oscillation on one end or severely cut your bandwidth on the other.  Either way you will no longer have a linear system over gain.