This amplifier evolved from the previous 6 transistor mosfet amplifier. That design was already so good there seemed no reason to look for improvements, but the high open-loop distortion at 20kHz and the resulting triangular 'error voltage' extracted suggested that the non-linear mosfet capacitance was responsible for almost all the distortion, and although the closed-loop distortion remains below 0.01% this could easily be reduced further by driving the mosfets from a lower impedance just by adding an emitter-follower. There is an added advantage that the high frequency compensation can now be supplied by a fixed capacitor, 220pF, instead of the mosfet capacitances, and then a series resistor, 100R, can reduce the phase lag from this capacitor at the frequency where other phase lags become significant, improving the phase margin.
The input resistors add a little noise. An 11k2 total series resistance at 20 deg.C. with a 20kHz bandwidth will add thermal noise at about 1.9uV r.m.s. The input transistor is a very low noise type, and should only worsen the figure by about 0.5dB. With a 1V r.m.s. input for full output this would give a signal to noise ratio of about 114dB, which is perfectly adequate. Listening with an ear close to the speaker (my MS-20, sensitivity 87dB/Watt) I could hear no noise.
The available output voltage swing is reduced slightly by the cascode driver stage. This could be avoided by a fairly simple bootstrapping circuit, but the improvement in maximum output may be just a fraction of a dB, depending on the supply voltage used, so too small to notice. The temperature coefficient of the output stage d.c. operating level can have a further small effect on available output, which again may be too small to be worth worrying about. Any drift will be very slow, and with capacitor coupled output this is not a problem. My own amplifier runs only slightly warm with a 60V supply, and the drift in operating point between cold and normal operating temperature gives less than a 1V variation from the optimum level for symmetric clipping. This would reduce maximum unclipped sine wave power by only 0.4dB. I am reluctant to add temperature compensation partly because I want to keep the design simple. To avoid additional components just replacing the 47k d.c. feedback resistor with a PTC thermistor seems obvious, but a quick calculation suggests problems with self-heating. A transistor in the d.c. feedback network has been suggested to me, which may work well, but adding a nonlinear component in the feedback network is something I prefer to avoid. My current favourite is a circuit using the LM19 temperature detector (1V5 at 25 deg.C, -11mV/deg.C, 3 pin TO-92, cost 35p.). In practice, unless a stabilised supply is used the positive supply voltage is variable, and so even a perfectly constant 'mid-point' voltage will not ensure accurate symmetrical clipping, so it is a questionable 'improvement'.
Ideally the output stage operating voltage should be adjusted using the 10k preset by checking for symmetrical clipping using an oscilloscope, but failing this it should be good enough to set the voltage with no signal to about half the supply voltage with the amplifier at its normal operating temperature. The negative output swing is reduced by the cascode stage, while the positive swing is reduced by the falling supply voltage at high output levels, so hopefully these effects will be approximately equal, giving something not too far from symmetry. Music signals are mostly asymmetric, giving greater clipping on one polarity or the other, so there is no point trying to be exact about this. If clipping occurs regularly you need a higher power amplifier anyway.
There is excellent supply rejection. The input stages have no direct link to the supply, and the high feedback will greatly reduce any breakthrough in later parts of the circuit. The distortion observed appeared to have no supply related components above the noise level. The 3V6 zener has a fairly high impedance at 1mA, typically 400R, and some theoretical improvement in supply rejection could be gained by using a ZRC330 3V3 voltage reference, or alternatively two red leds in series.
For best results with either mosfet amplifier design I suggest avoiding capacitive loads above 4uF and avoiding high input signal source impedance (3k or less should be no problem, if driven direct from a volume control this could be 10k or 5k log). It is important to distinguish between 'stability' and 'unconditional stability'. The fact that an amplifier fails to oscillate with any reasonable load is no guarantee of stability because the loop gain changes near to clipping, so that the unity gain frequency, where oscillation can occur, can shift downwards, and to ensure this has no unpleasant consequences it is a good idea to try to keep the phase lag round the feedback loop under 180deg at any frequency up to and a little beyond the loop unity gain frequency. An amplifier's internal open-loop output impedance may be significantly inductive at some frequencies, and with a capacitor load the resulting phase shift can in principle already approach 180deg without any additional effect from compensation capacitors etc. The effect of load impedance can be a serious problem for high feedback designs simply because the load to be used is unknown and therefore could cause phase shift at any frequency. The output inductor is about 0.4uH (13 turns 1.2mm enamelled copper wire with inside diameter 1cm and winding length 16mm) and is intended to reduce the effects of capacitive loads at high frequencies, but can do little to help in the audio frequency range.
The high frequency compensation uses a capacitor from the driver stage output to earth, which has the advantage of being minimum phase, whereas the more common 'Miller' feedback capacitor method is not, and in a typical circuit can add extra high frequency phase shift for a given attenuation because of feedforward through the capacitor, which is usually a small effect, but can be more serious with low stage gain, or near clipping where the stage gain falls, which as mentioned above is where stability can already be a problem. The question of whether the driver stage or the output stage clips first is critical for this effect. The feedforward effect was mentioned by Baxandall in his 1978 Wireless World series.
The shunt feedback inverting circuit avoids common-mode input distortion, which would not be reduced by the feedback. The 10pF capacitor in parallel with the 220k feedback resistor would normally be a problem because any radio frequency interference picked up by the speaker or its cable could be fed back to the input stage with little attenuation. A capacitor is often added in this position to improve stability by cancelling some of the phase lag round the feedback loop, but here it has an entirely different purpose, which is to more accurately define the feedback level at high frequencies where the feedback through a 220k resistor is not easily predictable because of stray capacitance. The 10pF and the 390pF from input base to earth form a potential divider which will attenuate any interference picked up by a factor of 40.
Leaving out the output inductor can result in an extra open-loop 80deg phase shift at the unity gain frequency, which we certainly want to avoid. This is demonstrated, plus additional information about loop stability, interference rejection, and board layout HERE. Stability could still be improved further, but the version shown here has been built and tested and found to have no obvious problems. I have a report of instability with small capacitive loads, which I had missed, and there is now a change in one compensation component, and there is an optional 100n capacitor added across the output so that the series resonance frequency with the output inductor is kept well below the unity gain frequency.
Apart from the following distortion measurements the performance is similar to the MJR-6 amplifier, and the same design and construction information applies.
Distortion measurement.Harmonic distortion, measured at 300mV rms input.
The 20Hz figures are not very accurate, maybe plus or minus 10dB.
Intermodulation from input 20kHz plus 21kHz, each at 150mV rms, the 1kHz distortion component was measured at -109dB relative to the output from 300mV.
Distortion: Input 1kHz at 300mV, distortion just appears above the noise level. Second harmonic measured at -116dB (0.00016%). Third harmonic at -129dB (0.000035%).
Distortion: Input 7.5kHz at 300mV. Although this looks a little spiky the only audio frequency component is the 15kHz second harmonic, measured at -106dB (0.0005%).
Distortion: Input 20kHz at 300mV. The nulling is relatively poor and the harmonics are beyond the range of the pc spectrum analyser. With the undistorted component nulled by about 80dB we still have a fairly good sine wave, so the distortion is certainly at a very low level. The distortion at 20kHz is of limited interest, all harmonics being well beyond the audible range, but comparison with the almost triangular 20kHz 'distortion' for the previous design without the improved nulling achieved here shows what a great improvement has been made to the high frequency linearity. A very approximate estimate for the distortion is "somewhere around 0.001%". The reason for the poor nulling appears to be that the open-loop phase shift is a little over 90deg at 20kHz and therefore a higher order phase correction circuit is needed to accurately null the remaining 20kHz component.
NOTE: The distortion extraction method used for these tests described here actually rejects noise added by the input resistors, which are the greatest source of noise in this design. This noise component is indistinguishable from input signal noise, and so is nulled along with all other components of the input. The distortion traces are therefore clearer than would be the case with more conventional measurement techniques. Improving test methods can make the results look worse, while publishing results with high noise content can suggest that distortion is 'below the noise level'.
Further improvements.It would be fairly easy to reduce the distortion further. Most of that remaining is caused by the output stage nonlinearity, and there are several ways to improve this. One is to increase quiescent current to reduce the variations in gm, another is to use several mosfets in parallel to reduce the output impedance, so that variations have less effect on the output signal.
There are higher power versions of the lateral mosfets (e.g. BUZ900DP and BUZ905DP) which appear to be just two of the low power types connected in parallel in the same package, so with quiescent current increased to 150 - 200mA and supply voltage to 95V this could be an easy way of both reducing distortion and increasing maximum power output to around 100W, provided the heatsink is also improved.
Photos of stages in the construction. This is a version I built myself, just to use for my computer sound output, and designed for low cost, so the appearance is not great, but it works well.
UPDATES: 25.May 2005
1. The 7R5 across the output inductor could be reduced to 1ohm to give better damping of resonances with capacitive loads, as demonstrated in the loop stability link above. This has other effects, such as increasing the input capacitance of the output stage because of reduced bootstrapping effect with low impedance loads, which may be a problem at high frequencies, so I left these as 7R5 in my own amplifier. This is one area needing further thought.
2. The quiescent current adjustment pot has been reduced to 220R, having found that the typical value required is around 100R. If the wiper of the original 1k became open-circuit the output stage current would become dangerously high. A parallel resistor added to the 1k pot would be an alternative.
3. For improved lifetime the use of electrolytic capacitors rated at maximum temperature 105deg.C is suggested. The lifetime is said to double for every 10deg below the rated maximum, so a 4 to 1 improvement could be gained compared to the 85deg.C types, all else being equal.
4. The 10uF across the 10k pot could be a non-polarised electrolytic, which it has been reported add less distortion than polarised types. The other electrolytics in the circuit have their distortion reduced to insignificant levels by the negative feedback, and for these the standard polarised types are adequate.
5. The lateral mosfets specified are more expensive and difficult to obtain in some countries. The UK supplier I suggested can apparently supply to other countries, but the delivery charge and any import duties could make this less attractive. Exicon types ECX10P16 (=2SJ162) and ECX10N16 (=2SK1058) appear to be equivalent lateral mosfets with similar specifications, but I have not tried these.