Current feedback single-ended input stage + push-pull VAS?

[antonis]

Perhaps a little crude

It fits the bill for 2955/3055..

[fryingpan]

But still, it (the original design) does soft clip. (Apparently). I wouldn't ascribe it to the output devices, MOSFETs, even lateral ones, don't clip "soft". They may produce more harmonics (worse linearity) but they still clip hard. That is also true of JFETs, by the way.

I also wouldn't ascribe it to the symmetry throughout the design. If the VAS is soft clipping, it would be soft clipping regardless of the dual input (because what happens for the top cascode happens for the bottom one as well). The symmetry in the input stage only serves (in theory) to combat non-linearities (which makes sense in the original design because, after all, it is supposed to be a hi-fi amp).

[PRR]

advance the state of tone.

Arrghk! I carefully made a paste-up and forgot to post it!

Bandwidth: the amplifier may be near-flat well past 20kHz.... 'my' Ampeg VT-40 went on and on (but I saw no reason to squeal 50kcps in 50 year old transformers already strained by a fire).

I'm sure most of us know the standard g-pickups resonate in the 3kHz-5kHz area and then drop-off 12dB/oct, and any large cone speaker that touches 3kHz-5kHz will drop VERY fast past there. One popular and super-typical plot (post the image this time, fool!):

[fryingpan]

Yeah, and I mean, output 1 (top schematic) can feed inputs 1 and 2? (It's not a differential stage, I presume they should be fed the same signal).

[Rob Strand]

Yeah, and I mean, output 1 (top schematic) can feed inputs 1 and 2? (It's not a differential stage, I presume they should be fed the same signal).

This might help show how to connect the cascode.    However, as you can see there's no soft clipping.

The first schematic I replace the R+R+C boostrap current source with a BJT current source.
The second I add a second pseudo-cascode VAS.

Schematic:


Waveforms:

But still, it (the original design) does soft clip. (Apparently). I wouldn't ascribe it to the output devices, MOSFETs, even lateral ones, don't clip "soft". They may produce more harmonics (worse linearity) but they still clip hard. That is also true of JFETs, by the way.

True, laterals and verticals are different.   The question is where is the soft clipping coming from.  At this point I can only think the MOSFETs.   

[fryingpan]

Yeah, and I mean, output 1 (top schematic) can feed inputs 1 and 2? (It's not a differential stage, I presume they should be fed the same signal).

This might help show how to connect the cascode.    However, as you can see there's no soft clipping.

The first schematic I replace the R+R+C boostrap current source with a BJT current source.
The second I add a second pseudo-cascode VAS.

Schematic:


Waveforms:

But still, it (the original design) does soft clip. (Apparently). I wouldn't ascribe it to the output devices, MOSFETs, even lateral ones, don't clip "soft". They may produce more harmonics (worse linearity) but they still clip hard. That is also true of JFETs, by the way.

True, laterals and verticals are different.  The question is where is the soft clipping coming from.  At this point I can only think the MOSFETs. 

What I'm thinking is: in a MOSFET amp, you can drive the MOSFETs directly from the VAS. Maybe the drivers are compensating for the reduced drive by the VAS as the cascode bias shifts? Or maybe there is too much global feedback?

Also, again, why do you have to use the upper cascode as a mere current source? Why not feed the same input as in the lower cascode?

[Rob Strand]

What I'm thinking is: in a MOSFET amp, you can drive the MOSFETs directly from the VAS. Maybe the drivers are compensating for the reduced drive by the VAS as the cascode bias shifts? Or maybe there is too much global feedback?

Also, again, why do you have to use the upper cascode as a mere current source? Why not feed the same input as in the lower cascode?

From the first post (and the Slone schematic)  I thought you wanted to take the simpler Elliot sound products (ESP) with the single input transistor and graft the Slone cascode onto the VAS circuit.  The problem there is the Slone circuit is fully symmetrical and the ESP is not.   *If* were to try to add two Slone cascodes to the ESP circuit then circuit I just posted is how to do it.   Each polarity swing has a Slone cascode which saturates during clipping. (Just for the record while what I have drawn works, I don't see it doing the ultimate job of soft clipping.  I wouldn't build anything until we see some soft clipping.)

From your comments it seems the picture you had in your mind was not a single input transistor with current feedback at the emitter but two such input circuits, one for each polarity.   So a structure more along the lines of the circuit Clint Eastwood posted earlier.  In that case the positive side VAS and cascode would end-up being the mirror image of the negative VAS and cascode in my last circuit.  Slone's schematic you posted earlier already shows that mirror image VAS+cascode.

You posted this earlier

I don't understand why the single output of the input stage cannot be fed into both "signal" transistors in the VAS. Is it a matter of current?

I might see your dilemma.   If you take Slone's schematic it has two VAS circuits, one for each polarity.    The problem trying to feed both VAS transistors from a single input transistor is it's not possible to DC couple the collector of the input transistor to the upper VAS transistor ; a voltage inversion is also required.    The DC voltage at upper VAS transistor is sitting near the +ve rail, whereas the collector of the input transistor is near the negative rail - which is fine for the lower VAS transistor.

[Rob Strand]

I played around a bit to force more soft clipping by changing the divider to the base of the Slone Cascode transistor.   Referring to my last schematic I reduced the divider on the negative rail base resistor (R29) from 4k7 to 1k.   That makes the cascode transistor cut off sooner , ie. when the output is further from the supply rail.

The waveform of the negative swing is:



You can see we have traded quite a bit of output swing to get a small amount of rounding of the edges.  Nowhere near the softness shown in Slone's waveforms.

The same idea on the positive rail doesn't work.   It seems the emitter resistor on the current source messes things up.

[fryingpan]

So, essentially, the idea is that you should DC couple all the power amp stages, otherwise you may be making an oscillator?

[teemuk]

But still, it (the original design) does soft clip. (Apparently). I wouldn't ascribe it to the output devices, MOSFETs, even lateral ones, don't clip "soft". They may produce more harmonics (worse linearity) but they still clip hard. That is also true of JFETs, by the way.

I wouldn't ascribe it to the devices either.

Although MOSFETs in something like a linear CMOS amp can actually clip very very softly. And they probably won't as source follower outputs for a closed loop amp circuit with high OLG and lots of NFB. ...And then, even generic BJTs will also soft clip as open loop differential so...

...It's really all about the circuit and obviously Slone cascode does a thing or two to enhance more gradual transformation into non-linearity and clipping. Yet, I have a hunch that it is a "hi-fi" circuit, and as such the size of "soft" knee region is always a compromise between output power and headroom before audible distortion kicks in. Therefore they probably did not go over the top with the softness.

"Soft clipping" can sometimes be a somewhat vague term as well. Many manufacturers actually appreciate more "graceful" clipping that may have a bit rough edges but does not have any other ill effects like "rail sticking", spurious oscillations etc. Soft clipping in practice can even be too "muddy", certainly things like tube amps advertised for it never clip as softly as CMOS linear amps, and in turn those CMOS linear amps are often described too "mushy" and lacking headroom.
There's probably a preferable balance between hardness and softness but folks are often too keen on getting caught to buzzwords and mental impressions like hard clipping=bad, soft clipping=good.

[fryingpan]

But still, it (the original design) does soft clip. (Apparently). I wouldn't ascribe it to the output devices, MOSFETs, even lateral ones, don't clip "soft". They may produce more harmonics (worse linearity) but they still clip hard. That is also true of JFETs, by the way.

I wouldn't ascribe it to the devices either.

Although MOSFETs in something like a linear CMOS amp can actually clip very very softly. And they probably won't as source follower outputs for a closed loop amp circuit with high OLG and lots of NFB. ...And then, even generic BJTs will also soft clip as open loop differential so...

...It's really all about the circuit and obviously Slone cascode does a thing or two to enhance more gradual transformation into non-linearity and clipping. Yet, I have a hunch that it is a "hi-fi" circuit, and as such the size of "soft" knee region is always a compromise between output power and headroom before audible distortion kicks in. Therefore they probably did not go over the top with the softness.

"Soft clipping" can sometimes be a somewhat vague term as well. Many manufacturers actually appreciate more "graceful" clipping that may have a bit rough edges but does not have any other ill effects like "rail sticking", spurious oscillations etc. Soft clipping in practice can even be too "muddy", certainly things like tube amps advertised for it never clip as softly as CMOS linear amps, and in turn those CMOS linear amps are often described too "mushy" and lacking headroom.
There's probably a preferable balance between hardness and softness but folks are often too keen on getting caught to buzzwords and mental impressions like hard clipping=bad, soft clipping=good.

I don't like the CMOS sound either. I'm not after super-soft clipping, but some softening of the knee, even just up there near the power rails (let's say within 3dB of the maximum peak) could prove beneficial.

[Rob Strand]

So, essentially, the idea is that you should DC couple all the power amp stages, otherwise you may be making an oscillator?

The normal way to do things is with DC coupling.   However, you can get AC coupling to work is specific cases.  If you choose a funky idea you have to work harder to make it work.   (Even the Elliot sound circuit you posted has a little quirk.  He splits the feedback resistor into a DC path and an DC path whereas the common form of the circuit uses a single feedback resistor.)

I wouldn't ascribe it to the devices either.

Although MOSFETs in something like a linear CMOS amp can actually clip very very softly. And they probably won't as source follower outputs for a closed loop amp circuit with high OLG and lots of NFB. ...And then, even generic BJTs will also soft clip as open loop differential so...

...It's really all about the circuit and obviously Slone cascode does a thing or two to enhance more gradual transformation into non-linearity and clipping. Yet, I have a hunch that it is a "hi-fi" circuit, and as such the size of "soft" knee region is always a compromise between output power and headroom before audible distortion kicks in. Therefore they probably did not go over the top with the softness

The bottom line is the idea of the soft VAS doesn't work in general.  That's demonstrated in my BJT simulation.  The softening isn't inherent in the Slone-cascode alone.   To get the softening in Slone's waveforms you need the Slone-cascode (no doubt there) but it needs to be working with the MOSFET characteristic, basically forcing the MOSFETs into a higher resistance region as the amplifier approaches clipping.  The MOSFETs need more swing on the gate than BJTs and also go into Triode mode as VDS reduces. 

[Rob Strand]

As a quick test I pulled some 2SK1058/2SJ162 models from the web and  replaced the output stage on the ESP circuit with the MOSFETs.   Result no soft clipping.    If I change the divider on the cascode base like I did before I see a small amount.    If I also raise the load impedance it becomes more soft.     None of the waveforms are soft like the Slone waveforms.

[Rob Strand]

In order to reduce any confounding results from the amp topology I change the target amplifier to a dual supply symmetrical design which is DC coupled.  That also prevents the tilt from the output cap affecting how we see things.  I didn't force the N and P spice models to be exact equivalents, so there's some small asymmetry from the models - that's going to occur in real devices anyway.

I whipped up two amps: one with a MOSFET output stage, one with a BJT output stage.
For each case I change the divider resistor (to VCC or GND) on the cascode to be 2k2 (my exaggerated softness) and 4k7 (Slone's divide by 2)

Schematic:



Waveform (scaled sort of like Slones oscilloscope pic).



Waveform zoom in at peak:



So we can see some softeness.  Not quite like Slone's oscilloscope but close enough to show the main features.   Keep in mind the BJT design can swing more so the depth of clipping is effectively lower.   So the cascode is doing something and the output devices make a difference.

Just to drive the point home the example amplifiers in this post are not intended to be built they are only suitable for spice.   Some of the transistors are pushed voltage-wise (spice doesn't care).   The cap on the feedback path is probably best left out so there's a DC path for any current mismatch.  Things like that need to be looked at more closely for a real design.

[fryingpan]

So what is actually causing the soft clipping? Maybe a combination of the cascode VAS into MOSFETs plus current limiting or whatever?

[Rob Strand]

So what is actually causing the soft clipping? Maybe a combination of the cascode VAS into MOSFETs plus current limiting or whatever?

Ultimately the voltage divider on the Slone cascode sets the soft limit.   That's the thing "doing it".   However, there's a little more to it, that's why the device also has an effect.

It's a bit tricky to understand:      When the output voltage swing is high, the output current is high.   The output devices have a transconductance which is, in loose terms, how much gate/base voltage is required to achieve a certain output current.   Lower transconductances devices need more swing on the input for same output current.     The increased input voltage means the cascode VAS can cut-off the device easier.

The other day I actually did an experiment where I replaced the output devices with transconductance devices.   If the transconductance was low I could get MOSFET looking soft-clipping and when the transconductance was high it was more like the BJT case.   I was pretty convinced that's why the output devices have an effect as well.

So the hanging question is why the simple circuit didn't show the clipping.   I didn't get that far but I suspect the lower output current means the input to the gate/base are low and the VAS can't reach the clipping knee.

[fryingpan]

Anyway, looking at both your schematic and Elliott's designs, I don't think there is a clear need for a DC path. Technically, since the emitter(s) of the input stage are connected to the load (and to the output devices as well), there is a clear path for DC currents to flow. Also, there could be a relative improvement by switching the common base devices in the cascode with MOSFETs (or FETs, if the voltages the devices see are low enough), since they should have a gentler transition into their linear (ohmic) region compared to BJTs. The design probably calls for higher voltages for the input stage and driver to eke out more power from the output devices, assuming such a thing is actually desired.

[antonis]

there could be a relative improvement by switching the common base devices in the cascode with MOSFETs (or FETs, if the voltages the devices see are low enough), since they should have a gentler transition into their linear (ohmic) region compared to BJTs.

Is I_DSS*[ 1- (V_GS/V_P) ]² more linear than I_S*[ exp (V_BE/V_T) -1 ]..??

[Rob Strand]

Is I_DSS*[ 1- (V_GS/V_P) ]² more linear than I_S*[ exp (V_BE/V_T) -1 ]..??

It's not so much the form of the linearity but the voltages required for a given output current.

Here's the results where the output devices are replaced by transconductance elements.  These are totally linear.  A smaller gm just means you need more input voltage for the same output current.

These have 4k7+2k2 on the cascode bias.

Schematic:



Waveform:




Zoomed:



The top two waveforms (red, cyan) are almost identical to the BJT and the second from the bottom (blue) almost identical to the MOSFET.

[fryingpan]

there could be a relative improvement by switching the common base devices in the cascode with MOSFETs (or FETs, if the voltages the devices see are low enough), since they should have a gentler transition into their linear (ohmic) region compared to BJTs.

Is I_DSS*[ 1- (V_GS/V_P) ]² more linear than I_S*[ exp (V_BE/V_T) -1 ]..??

Well... one derivative is linear, the other is an exponential.