Class i and siblings

Status
This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.
This triple configuration does not appear to be stable, in any non-trivial (>5nF) capacitive load. See attachment, load is 10nF/step. For output swings toward the rails, I would expect the stability margins to be even worse. This is for zero source impedance, which is not necessary realistic - it is almost impossible to fully analyze the stability of this triple configuration in isolation from the preceding transimpedance gain stage.

Unfortunately, a Zobel cell, output inductance, driver BC cap are here not helping.
 

Attachments

  • stability.jpg
    stability.jpg
    172.7 KB · Views: 409
Last edited:
Fodity

Hi Edmond
Dr. Bernard M. Oliver (Barney) did not mention any optimum point in his article (http://www.hpl.hp.com/hpjournal/pdfs/IssuePDFs/1971-02.pdf). He proposes as optimal the whole interval from 13 to 26 millivolts across one resistor (p.13), so any figure is right.
But only for the primitive case, in which - as necessary for organizing Class AB nonlinearities - are utilized nonlinearities of the output devices. These nonlinearities are very inconvenient (susceptible to thermal shocks, sluggish, asymmetric etc.). And they are not fod, i.e. free of (crossover) distortion.
So some people like you and Kendall eliminate them by high amount of NFB thru powerful drivers.
Then you should create (synthesize, tailor) your own nonlinearities instead, preferably fod. Only this property (let’s call it fodity) must be thermally stable, precise, independent of frequency etc.
So Oliver’s analysis and Oliver’s criteria are not applicable. You’ve already noted this in #28.
And if, for example, basics of Kendall’s nonlinearities dictate that (like in most cases with bjt’s) Iq must grow with temperature, one cannot try to stabilize it.
And what about fodity? This is simply a “main condition” from #113: difference between the two nonlinear functions Ip(Vin) and In(Vin) must result in a linear function Gm*Vin.
For your simulation in #115 you define Vin as the mean voltage of the top and bottom driver input.
But it's no use in this case, like “Barny Oliver with all his wisdom”, because no one is interesting about what happens deep inside the circuitry. Seems that more useful would be
Gm=del(I(R37))/del(V(in)-V(out)), where (in) is the left node of R1 or R3.
For Class-I theoretically calculated Gm is exactly equal to 1/R24 (#3 in this thread), you can verify it by simulation. And what about ABII?
Cheers Mir
 
VT & Gm

Hi Edmond
Dr. Bernard M. Oliver (Barney) did not mention any optimum point in his article (http://www.hpl.hp.com/hpjournal/pdfs/IssuePDFs/1971-02.pdf). He proposes as optimal the whole interval from 13 to 26 millivolts across one resistor (p.13), so any figure is right.

Hi Mir,

Perhaps I'm wrong, but I thought there's general agreement that Vq = VT is optimal. Although Oliver didn't say this explicitly, his fig.6 (see 1st pic. below) suggest that this is indeed the case (go*R=1).
On the other hand, the result of some of my simulations reveals that the optimal Vq depends on the signal level. I found optima between 14 and 38mV (@ room temp.)
Furthermore, I also got some contradicting result: On page 28 I mentioned some increase of the distortion when Vq=VT, while on another occasion I found a little improvement, though this was with a slightly different auto-bias circuit.
So, I think there's no need to strictly adhere to the Vq=VT rule (as you said already).

But only for the primitive case, in which - as necessary for organizing Class AB nonlinearities - are utilized nonlinearities of the output devices. These nonlinearities are very inconvenient (susceptible to thermal shocks, sluggish, asymmetric etc.). And they are not fod, i.e. free of (crossover) distortion.
So some people like you and Kendall eliminate them by high amount of NFB thru powerful drivers.
Then you should create (synthesize, tailor) your own nonlinearities instead, preferably fod. Only this property (let’s call it fodity) must be thermally stable, precise, independent of frequency etc.
So Oliver’s analysis and Oliver’s criteria are not applicable. You’ve already noted this in #28.
And if, for example, basics of Kendall’s nonlinearities dictate that (like in most cases with bjt’s) Iq must grow with temperature, one cannot try to stabilize it.
And what about fodity? This is simply a “main condition” from #113: difference between the two nonlinear functions Ip(Vin) and In(Vin) must result in a linear function Gm*Vin.
For your simulation in #115 you define Vin as the mean voltage of the top and bottom driver input.
But it's no use in this case, like “Barny Oliver with all his wisdom”, because no one is interesting about what happens deep inside the circuitry.

IMHO, it's most relevant to look at the transfer function relative to this 'deep inside input'. The point is that any nonlinearity between this input and the output directly translate to the overall performance. Besides, as the phase compensation stuff is connected to these driver inputs, the overall performance at higher frequencies is even more determined by this transfer function.

Seems that more useful would be Gm=del(I(R37))/del(V(in)-V(out)), where (in) is the left node of R1 or R3.

Agreed, that's also very useful, in particular when comparing Class-I with AB2.

For Class-I theoretically calculated Gm is exactly equal to 1/R24 (#3 in this thread), you can verify it by simulation. And what about ABII?
Cheers Mir

Well, more precisely, for positive signals it's indeed equal to 1/R24. However.......... for negative signals it's equal to 1/R25 and you know of course what that means. Perhaps, this is the weakest point of Class-I, at least to me. I would never build this OPS.

Regarding AB2, in post 115, I made again a stupid mistake. I wrote: "So the incremental transconductance becomes: gm = del( I(R37) / del (0.5*( V(B1) + V(B2) ) - V(out) )"
That's okay. Then I wrote: "To my own surprise, gm is almost perfectly constant. At Vout =0 there is a tiny dip of about 2.7%. No gm doubling at all."
That's BS. No wonder I was surprised. Have a closer look at the expression for gm: Del(I(R37))/Del(0.5*(V(b1)+V(b2))-V(out)) . Do you see the error?

Fig.2, green curve, shows the correct 'internal' gm of the amp on page 41. It's far from constant; it varies from 3 to 4.6 Siemens. I also tried to plot the overall gm, but at some output levels, it's becomes + or - infinity (weird!). So I've plot the inverse gm, i.e. the incremental Zo, which varies from -0.13mOhm to +1mOhm. See the blue curve. The reason for this weird behavior seems to be the cross-connection of R23 and R24. If I replace them by CCSes, Zo varies from +0.3mOhm to +0.9mOhm, in about the same manner as in fig.3.

The 3rd pic. shows the 'internal' (green) and overall gm (blue) of AB2 with a MOSFET OPS. This one doesn't have the resistor cross-connection in the driver stage. So gm behaves 'normal'. Purple is the incremental Zo.

BTW, what do you mean by 'fod'? 'Free of distortion', right?

Cheers,
E.
 

Attachments

  • AB2-MOS-GM.png
    AB2-MOS-GM.png
    21.7 KB · Views: 134
  • AB2-BJT-PP2-GM.png
    AB2-BJT-PP2-GM.png
    19.1 KB · Views: 364
  • Oliver-Xover.jpg
    Oliver-Xover.jpg
    47.1 KB · Views: 382
Last edited:
Hi Damir,

What if you connect that cap directly to the output, that is, before the Zobel network? Also, what happens if you add parasitic inductances of the leads and traces, in particular to the OPS trannies?

Cheers,
E.

Hi Edmond,
I don't know how much capacitance is realistic to have before Zobel network but with 100nF it is not stable, with 10nF it is. Next test is with parasitic inductances(10uH I think that is realistic) and 1nF before Zobel and 2uF after and no problem there.
Damir
 

Attachments

  • EDMOND-OPS-1b-100n-b.jpg
    EDMOND-OPS-1b-100n-b.jpg
    151.5 KB · Views: 183
  • EDMOND-OPS-1b-100n-c.jpg
    EDMOND-OPS-1b-100n-c.jpg
    156.4 KB · Views: 170
Hi Edmond,
I don't know how much capacitance is realistic to have before Zobel network but with 100nF it is not stable, with 10nF it is.

Hi Damir,

Most amps don't like 100nF (even Bob's HEC amp). OTOH, 10nF is a bit low.

Next test is with parasitic inductances(10uH I think that is realistic) and 1nF before Zobel and 2uF after and no problem there.
Damir

10uH?No! 10 to 50nH in series with the legs of the OP devices (~10nH/cm).

Cheers,
E.
 
For whatever reasons, you seem to get better results in simulation (optimistic device models?). Do yourself a favor and repet the experiment with the amp output biased towards the rails. You may be surprised by the results.

Also do yourself a favor and determine the phase margins of the output triple. A sine transient analysis cannot tell between 60 and 5 degrees of phase margin.


Note that the Zobel and output inductors are there to protect the amp against worse case loads, think of them as safety margins. A quality amp should be stable without these.

10nF load is not that much, think of electrostatic speakers.

But then, as long as your speakers allow, I guess you can live with a marginally stable amp. Lots of people do, without even knowing about.
 
Disabled Account
Joined 2008
For whatever reasons, you seem to get better results in simulation (optimistic device models?). Do yourself a favor and repet the experiment with the amp output biased towards the rails. You may be surprised by the results.

Also do yourself a favor and determine the phase margins of the output triple. A sine transient analysis cannot tell between 60 and 5 degrees of phase margin.


Note that the Zobel and output inductors are there to protect the amp against worse case loads, think of them as safety margins. A quality amp should be stable without these.

10nF load is not that much, think of electrostatic speakers.

But then, as long as your speakers allow, I guess you can live with a marginally stable amp. Lots of people do, without even knowing about.


I agree with everything Walter is saying.

syn08 has done a very good research on triple EF stability problems, a good thing to read, you can find it here.

BTW: Personally I never use Zobel, output inductors nor input filter when I do stability simulations. Those has nothing to do with the "internal" stability of the amplifier.

Cheers
Stein
 
Class-I

Hi Edmond
[snip]
For Class-I theoretically calculated Gm is exactly equal to 1/R24 (#3 in this thread), you can verify it by simulation. [...]
Cheers Mir

Hi Mir,

However, practically, with perfectly match transistors and a highly optimized circuit, Zo still varies by about 0.5mOhm. This is of the same order as with AB2.
Regarding R24 & R25, you will need four terminal (Kelvin) high precision current sense resistors. Even 0.5% types do increase the distortion by tens of dB. :sad:

Cheers,
E.
 
Hi Stein,

>BTW: Personally I never use Zobel, output inductors ...

Agreed.

> nor input filter when I do stability simulations....

This statement needs some explanation: not every amp likes a full power square wave, not only because one of more stages get saturated (or completely turned off), but also it over-stresses the drivers. This is most apparent in case of a MOSFET OPS (because of high input capacitances). It really makes sens to reduce the slew rate from say 500V/us to 50V/us, as you can replace the relative large (and sluggish!) driver trannies by smaller (and faster!) ones. The best way to reduce the slew rate is by means of a second order LP filter, as a 1st order is far less effective. This is because at the onset of a transition the SR is still pretty high.
Perhaps you still remember that Andy C. has already explained this on the other forum.

A better way to test an amp on stability (ringing) without an input filter was proposed by Glen K. (AFAIK). Instead of a large square wave he uses a large LF sine wave and superimposed on this a small HF square wave. With this setup you can examine the step response at various output levels -also near the supply voltage, very important!- without needlessly over-stressing the amp.

Cheers,
E.
 
Last edited:
gm Class-I

Hi Edmond
[snip]
For Class-I theoretically calculated Gm is exactly equal to 1/R24 (#3 in this thread), you can verify it by simulation. And what about ABII?
Cheers Mir

Hi Mir,

I forgot to drop a plot, sorry. Here it is for the Class-I circuit (from page 3, without R2).
Red: emitter currents of the OP transistors.
Green: the 'internal' gm of the OPS (driver + OP trannies). Notice that the sliding bias mechanism doesn't keep it constant.
Blue: Gm of the whole amp.
Purple: Zo = 1/gm. Indeed, almost perfectly equal to R24, ~0.33 Ohms.
Cyan: Zo - 0.33. Now the true nature of this circuit is revealed: After subtraction RE from Zo, this value appears to be far from constant.
One might say that this circuit is equivalent to an amp with a large amount of NFB together with a series resistor (RE) connected to the output.
Needless to say that we shouldn't get fooled by a seemingly nice and constant Zo or gm. It's the constant value of RE that attributed to this 'constantness' (assumed that both RE's are exactly equal !).
Why is Zo equal to RE? Because the FB is taken from the emitters of the OP trannies instead of the output proper. See 2nd figure. Looks pretty insane, don't you think so?

>And what about ABII?

In AB2 the (differential) FB is taken directly from the output (as it should be). As a result, Zo is far lower. Theoretically a few uOmhs, practically around 1mOmm.

Cheers,
E.
 

Attachments

  • Kendall-GM.png
    Kendall-GM.png
    22.1 KB · Views: 373
  • StupidAmp.png
    StupidAmp.png
    7.4 KB · Views: 371
Last edited:
....
The cross-connection reduces the feed-through of switching currents (of the sense trannies) into the main signal path. As a result, less induced distortion from these trannies. However, it has hardy any effect on the Ip*In product.
.....
E.

But there still small current continuity gap feed from input signal when crossing, you could improve it if and make cross conduction or another additional circuitry on it that able to create negative(opposite) current gap that naturally created by class B or AB circuitry. And looks like cross connection need to work in class AB to do this task.
Of course not necessary, because less than 0.005% THD couldn't be heard.
All you need to improve is stability.
 
D2S

It's possible to linearize the 'internal' gm of a class-AB OPS. See:
Michael Williams, "Making a linear difference to square law fets", EW+WW, Jan. 1994, pp.82-84.
It's about the so called D2S amp*, which relies on the fact that the difference of two equal quadratic functions returns a linear function.

Cheer,
E.

*No, it doesn't run on Deuterium Sulfide. ;)
 
Last edited:
Hi, Edmond, any link? Ok I'll try with googing search.

Hi Edmond,
I don't know how much capacitance is realistic to have before Zobel network but with 100nF it is not stable, with 10nF it is. Next test is with parasitic inductances(10uH I think that is realistic) and 1nF before Zobel and 2uF after and no problem there.
Damir

Use this test, instead. It is at least more realistic and easier. Both in test is using AB2.
 

Attachments

  • stb vs unstb.PNG
    stb vs unstb.PNG
    19.2 KB · Views: 365
Status
This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.