1. The problems with bipolar junction transistors
The improved class-B power amplifier circuit still has a few problems, some of which become more severe if we require a higher output power than the simple 20 watt example. The stability margin of even that design was not very good when driving capacitive loads, and also the distortion was only measured at low and medium output levels. The purpose was merely to show the effectiveness in reducing crossover distortion. At higher outputs the distortion at 20kHz did increase, as would be expected, but the reason for this was not investigated in detail. The crossover distortion was reduced below the noise level, and that was all the design was intended to demonstrate. Actually the level at which 20kHz distortion was measured is the maximum level likely to be encountered at that frequency in normal music signals, so distortion at much higher levels is not necessarily important.
At high power levels there is an additional problem. Suppose we still use the same output stage circuit as the 20watt design, only increasing device ratings and supply voltage. The power output devices in a 100 watt amplifier may need to drive over 10 amps into a typical '8 ohm' speaker with impedance falling below 4 ohms at some frequencies. The 'worst case' current gain of a typical device may fall below 10 at this current, and the driver transistor then needs to supply 1 amp base current. If this driver transistor has a current gain of 50 at 1 amp collector current it then needs a base current of 20mA, supplied by the collector current of the input transistor of the triple. This current must flow through the feedback resistors of the output triple, which for a pair of 100 ohm resistors gives a voltage drop of 1V. This requires an additional 1V input to the output stage, which in the negative direction will certainly be enough to switch off the 'class-A' error amplifier. The error correction mechanism then becomes inoperative at high output currents, and distortion from the error amplifier switching off becomes likely. Choosing higher gain devices and reducing the feedback resistors would help but may still not be enough to avoid this problem.
2. A different answer with power mosfets.
One obvious way to reduce current requirements in the driver devices is to use power mosfets. These have a much lower input current requirement than bipolar junction transistors (BJTs). Even at 20kHz where input capacitance becomes significant, the input current is unlikely to exceed 1mA. This improvement by a factor of 1000 seems to solve the problem. A possible output sub-amplifier is shown next:
This can be improved if required by replacing R2 and R1 by current sources. The local negative feedback loop in effect contains only the mosfet and one of the driver transistors, the other acting mostly as an emitter follower buffer stage. There is an added advantage that provided the two driver transistors are in good thermal contact their variations in base-emitter voltages with temperature will cancel and there may be no need for further temperature compensation.
So, thinking this to be the perfect output stage, I started to design a complete amplifier using the 'improved class-B' technique to eliminate crossover distortion.
Then something occurred to me which should have been obvious from the start. If a local feedback loop using just two or three transistors is sufficient to linearise a single mosfet in one of the sub-amplifiers, then surely the same sort of loop can be applied to a complete class-B mosfet output stage and achieve the same result. An example of what I mean:
This is not intended as a final design, just an example to illustrate a point. Q1 to Q7 form a normal input and driver stage, similar to those recommended by Douglas Self. Usually a mosfet output stage would be driven direct from a driver stage such as Q7 and overall negative feedback used to linearise it. In the case of BJT devices the output stage must use two or three devices in each output sub-amplifier to reduce the necessary driver stage current to reasonable levels, and local feedback in each of the two sub-amplifiers. The temptation is to do the same with mosfets, but in fact greater linearity can be achieved by using a single unity gain output triple, but with the two mosfets together, biased to a quiescent current of about 100mA, as the output 'device' driven by Q8 and Q9. The reason why this is more linear than the single mosfet output triple such as Fig.1 is because the low mosfet mutual conductance at 100mA relative to higher currents is doubled because two devices are conducting at the same time and both contributing gain. The output stage (Q8 to Q11) in Fig.2 is already more linear than the sort of class-A error amplifier which a mosfet version of the 'improved class-B' circuit would use, so it seems an unavoidable conclusion that a mosfet version would be entirely pointless. The question remains whether a circuit similar to Fig.2 will be as good as or better than the previous 20watt BJT design. Is crossover distortion still below the noise level? Well, probably Fig.2 would not be so good, but improvements are possible, and this is a starting point for my next design, to be finished someday, not necessarily soon. There is possibly nothing entirely new or original about this sort of design, only the realisation that what seems a good idea for BJT circuits is not necessarily good for mosfets, and in this case something far simpler and more conventional turns out to be better.
The big question is whether the greater negative feedback possible with mosfets is sufficient to achieve the same linearity as BJTs at their own maximum feedback. Opinions seem to be divided on this issue. My own experience of mosfet amplifiers includes producing a commercial amplifier board many years ago. I have used a pair of these in a 75 watt per channel amplifier for the last 18 years with the only problem in that time a faulty speaker switching relay. Tests revealed no square wave ringing with the usual capacitive test loads in spite of heavy negative feedback, in sharp contrast to the 20watt BJT design which produced a worrying level of ringing with only moderate feedback. (The ringing was observed prior to the output inductor to try to ensure it was amplifier ringing rather than just the effect of the inductor resonating with the load capacitor, which is relatively harmless.) This gives some hope that the heavy local output stage feedback suggested here may prove to be effective in a mosfet design
To summarise: If a local feedback loop is sufficient to linearise a single mosfet so that it can be used as the class-A error amplifier in my previous design, then the same sort of loop can be applied to a pair of mosfets to produce a complete output stage with the same or better linearity, and no need for accurate component balance, thermal compensation etc.
For anyone wanting to experiment with Fig.2, I again must warn that this is not a final design, and will not necessarily work in exactly this form. For example there is little to limit the current through some of the transistors when the amplifier is over-driven, the gate to source voltages of the mosfets need to be limited to about 12V, supply line fuses should be added to protect against power device failure and a speaker protection relay is a good idea. In the commercial design mentioned earlier a large proportion of the development time was spent checking and controlling what happened if just one of the supply line fuses is open-circuit. A number of diodes and resistors had to be added to ensure no electrolytics became seriously reverse biased and no devices took too much current. I have heard reports of fuse failure causing serious damage in some equipment the fuse was supposed to be protecting. This sort of consideration is of as much importance as designing for low distortion.
Just to illustrate the additional level of complexity needed for a general purpose commercial design, here is the circuit diagram of the 'Renardson Electronics MPA80 Mosfet Power Amplifier' from 1984. This started out as just the basic circuit from the Hitachi Application Note.