The date of the latest version is shown above. There are now a few photos of oscilloscope traces for this circuit showing various aspects of performance. Thanks to Dimitri Danyuk there are also simulations of the output stage which give a useful indication of the level of linearity achieved plus a possible improvement. A sketch of a suggested board layout is included, but is not tested.
In addition to demonstrating the latest version of the feedforward error correcting output stage this is also a design exercise to demonstrate that it is possible to make an amplifier completely asymmetric, but still with good performance. Symmetry is widely believed to be essential, or at least a worthwhile aim, but the present design has no balanced differential stages, no complementary pair output power devices, a single voltage power supply, and an output stage where one half operates in class-A while the other half is in class-AB. Symmetry may, at best, reduce second and other even harmonic distortion, but this is the least unpleasant form, and odd order harmonics and crossover are far more important. Initial tests show that all distortion, even second harmonic, is low in this design.
The version shown here uses 12 transistors and relies on a single fuse for overload protection. More complex designs could possibly give better results, but the aim was to see how well a fairly simple version could be made to perform. The supply voltage can be about 58V for a 35watt version, e.g. using a 40V rms supply transformer. The capacitor coupled output protects the speakers from output transistor failure or anything else causing a large shift in dc output voltage, while any adverse effects of the capacitor are reduced by including it in the overall feedback loop. A rough calculation suggests that a shorted output will not damage the power transistors before the fuse blows, but I have no intention of testing this, though I did manage to blow a 3A-T fuse during testing with a 100kHz square wave into a 4uF load, (a bad idea!) and the amplifier was still ok after fitting a new fuse.
Thermal protection is now shown in the diagram to limit heatsink temperature to 70 deg.C using a normally open thermal switch. This is bolted to the heatsink, and in effect switches off the amplifier at 70 deg. The cut off takes a few seconds, and the output distorts during this time, but switching back on is faster. Putting the switch in the driver stage current source means a normally open switch can be used, so the contacts have no effect during normal operation. Reliability is improved, because only a small current is switched, instead of the alternative of switching the 55V power line.
T type anti surge fuses typically can survive at 10 times their rated current for about 100 msec, and at 5 times the rated current for 1 sec, so a 2A supply line fuse in an amplifier capable of supplying 8A peak output current into a 3ohm load is not the serious limit to output power that it may seem. Even for a continuous 8A peak sine wave output the average supply current will be about 2.5A, and a 2A-T fuse can pass this curent almost indefinitely. The protection from a fuse is clearly inferior to that from full load-line protection, but the resulting simplicity and ability to drive difficult loads without protection circuits adding distortion are some advantage.
The BF469 (NPN), BF423 (PNP) and BF470 (PNP) are used because they have very low base-collector capacitance. The 2SC2547 has very low audio frequency noise figure, under 0.5dB, at the 0.2mA collector current and 10k source resistance used here. (The high gain version 2SC2547E is better if available). The 10k resistance adds less than 2 microvolts rms input noise over a 20kHz bandwidth, which should be inaudible with most speakers. All resistors used were 1% tolerance metal film apart from the 4R7 values in the output network.
The supply line rejection is very good, the input and driver stages not having any connection to the supply other than the bias current supply for the 2V7 zener. A single supply voltage is used, so only one large smoothing capacitor is needed in the power supply, 15000uF 63V being typical. There may be some advantage in using several smaller values in parallel, the impedance often being smaller and current rating higher for the same cost. The same advantage is gained by using a 4400uF output capacitor made from two 2200uF in parallel, though the original reason for this was because the 4700uF capacitors I could obtain were too tall to fit in the case I wanted to use.
Why use capacitor coupling to the speaker output? In a previous commercial design I used a direct coupled output with a relay for speaker protection switching. I really hated that output relay! The only fault ever to happen to that design was relay failure. Many years ago most transistor power amplifiers used output coupling capacitors, and then there was much less danger of speaker damage, e.g. if one of the power transistors became shorted. Output capacitors add a little distortion, (D.Self, Electronics World Sept 1997 p718, shows a measurement at 40W into 8ohms via a 'standard' 6800uF of 0.0025% mostly third harmonic at mid frequencies, rising above this only below 20Hz. More recent measurements by C.Bateman found significant second harmonic and demonstrated that there is an optimum bias voltage at which distortion is a minimum, though this voltage appears to be too low to be convenient for the present application). They also reduce damping factor at bass frequencies, have a small series resistance and are generally thought to be a "bad thing". This present design reduces these problems of a coupling capacitor by including it in the overall feedback loop.
One common criticism of electrolytic capacitors is that some of the energy absorbed is only released after a time delay. One writer even suggested that this would have an effect similar to reverberation. If a simple capacitor could store an audio signal and play it back at a later time, that would certainly be astonishing, but fortunately the truth is much simpler. An example sometimes mentioned involves a charged capacitor being discharged close to zero volts, and then left open circuit, when the voltage slowly drifts upwards again. (The effect is sometimes called 'dielectric absorption', 'dielectric relaxation', or 'soakage'). The effect can be modelled approximately by just two 'ideal' capacitors connected together in parallel via a very high value resistor. If capacitor C1 is charged the other, C2, will slowly charge through the resistor, then if C1 is discharged momentarily by a small resistor, the other capacitor will still be charged, and after the discharge resistor is removed some of this charge will flow back to C1 and its output voltage will rise. There is nothing non-linear involved in this model, and the second capacitor will only 'remember' the long term average dc voltage. (I have added an article about capacitor distortion, in which I demonstrate that adding dielectric absorption to a coupling capacitor has little effect on gain error, but actually reduces phase error.) The effect can be important in some other applications, e.g. 'follow and hold' circuits. Dielectric absorption is associated with dielectrics with polar molecules, in which the molecules become aligned in response to an electric field. This is an inherently non-linear process, once a molecule is fully aligned it can make no further contribution. The observed effect is just low level third harmonic distortion, but steps can be taken to reduce this distortion such as connection to a high impedance load, or using a very high value capacitance to keep voltage drop low, or using the capacitor inside a feedback loop. There are expensive 'audiophile' electrolytic capacitors available, some of which do actually have lower third harmonic distortion than standard types (according to measurements by Douglas Self), but the precautions mentioned here are generally sufficient to make these unnecessary.
Using capacitor coupling the output offset voltage is no longer important, and so there is no need for a differential pair input stage which is usually used to balance out the offset. Anyone still worried about slew rate limiting and t.i.d. will be pleased to find that the high frequency compensation is applied via 2.7pF back to the base of the first input stage transistor, and there is no differential input stage prior to the compensation as in most conventional designs which can limit or distort at high slew rates, and these effects will be absent.
The use of a parallel feedback 'virtual earth' circuit means that this design is inverting, and some small subjective effect has been claimed for signal inversion. This can be avoided if required simply by reversing the speaker connections. The feedback arrangement avoids common-mode input stage distortion, and gives good protection against radio frequency interference, the input passing through a low-pass filter, then via 5.1 kohms to the low impedance input. The 47pF capacitor is an essential part of the input stage frequency compensation for loop stability, but also helps reduce interference. The high impedance feedback from the output stage will help avoid interference pickup from the speaker cables getting into the input stage, as will the output inductor, included for stability with capacitive loads.
The five MJE and MJL transistors are all mounted on the same heatsink bracket, and the MJE15030 temperature compensation transistor is immediately next to the MJE15031 to keep their temperatures close. Having reduced the output resistors to 0.5 ohms instead of the 1 ohms used in the original 20watt design it was expected the quiescent current stability would be poorer, but in fact it was better, there being typically less than 5mA drift when set to 100mA. The use of a fairly high power MJE15030 as the compensation transistor ensures that it has good thermal coupling to the heatsink and the same thermal time-constant as the transistor it needs to track. Lower power transistors have the disadvantage that their base-emitter voltage would be too large at the 10mA collector current, and even with the 1k adjustment pot removed the quiescent current would be too high, and a base to collector resistor would be needed, which would reduce loop gain in the output sub-amplifier. The 1k control shown should be set to its maximum resistance when first switching on, and then reducing to increase quiescent current to a suitable value, e.g. 100mA. The quiescent current only depends on the base-emitter voltage of one of the output transistors, unlike conventional class-B circuits where at least two transistors have an effect, and so the compensation transistor only needs to detect the temperature of one output transistor, in this case the MJE15031. Conventional circuits of course need an accurate setting of current for minimum crossover distortion, whereas the present design has low distortion at any medium or high current.
Further reduction of the output resistors may be worthwhile, but the need for accurate values is a problem. 1% tolerance high power resistors under 1 ohm are hard to find, and the 0.25 ohm values were each made from four 1 ohm 0.7 watt 1% metal film resistors in parallel, and the 0.5 ohm value from two of these 0.25 ohm arrangements in series. This makes a total of 16 individual resistors, which takes up a fairly large space on the circuit board, and even more parallel resistors would be inconvenient in this respect. The present arrangement is adequate for use with normal music, but when testing with high signal levels into excessively low impedance loads the resistors began to smoke before the 3A fuse came to their rescue. This did them no obvious harm, but their power rating is based on normal music signals rather than prolonged full level square waves into low impedance. To be on the safe side using these resistors the recommended fuse is now 2A (T type anti-surge). Even with low impedance speakers fuse failure is unlikely to be a frequent problem. Incidentally, the need for accurate matching of the output resistors is common to most class-B or class-AB amplifiers, and any difference in values can increase distortion.
Why only 35 watts?
A survey previously run on this website showed that only a minority of those responding were interested in a 100 watt version ( 7 out of 40 ). Also there is the problem that for a 100 watt rating even with parallel pairs of 200 watt output devices they are being pushed well beyond their safe limit even with a mimimum load of 3 ohms, ( taking reactive components into account and derating for a maximum heatsink temperature of 75 deg.C). Some speakers ( fortunately very few ) have impedance falling to 2 ohms or less, so 3 ohms is not an unreasonable aim. Using single output devices 35 watts seems a reasonable maximum specification. Anyone with 2ohm speakers should preferably sell them and buy something more sensible, but failing that the supply voltage should then be limited to 50 volts (e.g. from a 35V rms transformer).
If only speakers with a less demanding impedance, never less than 6 ohms, are to be used and more power is required it would be possible to exceed 100 watts into 8 ohms by using two of the 35 watt boards in a bridge amplifier, although it was not designed with this application in mind. An alternative (which has not been tried yet) is to increase the supply voltage, e.g. to 80V. The transistors specified are adequate for over 100V, but the capacitor voltage ratings must be chosen correctly, and resistor power ratings increased where necessary. For the 35W version the electrolytics need to be 63V, apart from the three 10uF which can be 6.3V or more.
The 1.5uF input capacitor should be a non-polarised type, e.g. polyester, which I used myself, or lower distortion but higher priced types such as polyphenylene sulphide. The 2p7 and 5p6 can be 'low-k NPO' ceramic rated at 100V or more. This type are surprisingly easy to damage with excessive soldering temperature, and some sort of heat-shunt should be used. (I recently saw a copy of distortion measurements published by C.Bateman which confirms the excellent results from this type). There are few alternatives available at 2.7pF. (Silvered mica are one possibility, but these are found to have a relatively high dielectric absorption, and as mentioned above may therefore also have higher non-linearity. Ref: Circuits, Systems & Standards, 'Understanding Capacitor Soakage to Optimise Analogue Systems' by Robert A. Pease, National Semiconductors, reprinted in Electronics World, Oct.1992 p832-835). The transistor capacitances in this part of the circuit are definitely non-linear, and of far greater concern, and should be as small as posible. For this reason the BF469/470/423 are specified, having 1.8pF max Ccb at 30V. The BF423 is a TO92 device, but the higher power BF470 could be used here also.
Some published designs use the highly regarded Toshiba 2SC3281 and 2SA1302 output power transistors, but apparently these were discontinued by Toshiba a few years ago. They are still widely advertised, but there are cheap versions made in China by a different manufacturer and with an inferior performance, and worse still, there are reports of the widespread sale of counterfeit devices marked with the Toshiba name. These devices will certainly not survive at anywhere near the full voltage or power rating. To see a report on the counterfeit trade goto: Elliot Sound Products.
The plastic packaged Motorola MJL21194 used here have a power rating of 200 Watts and 16 Amp maximum current. Although the current-gain-bandwidth product appears to be higher for the Toshiba devices, a reduction by a factor of 30 or more for collector current increasing from 3 Amps to 10 Amps does not inspire confidence. The MJL21194 appears to have a far lower specification, but it is better maintained, and seems to be better than the Toshiba device at 10 Amps. The current gain falls by only about 20% from 100mA up to 8A, and the MJE devices have virtually flat current gain over their operating range.
One potential disadvantage of not using an output relay is that switch on and off thumps from either the amplifier itself or the signal source used are not so easily avoided, although they can be minimised by suitable choice of smoothing capacitors etc. The version shown above had only a short 1 volt low frequency pulse at the output when switched on, and with a speaker connected this was audible, but unobtrusive. There was no noticeable switch-off thump, but of course any output signal began to distort as the supply voltage fell towards zero. (A 300VA 35V rms transformer and 27,000uF smoothing capacitor were used in the power supply for testing purposes.)
The input / driver stage feedback loop was originally unstable until a capacitor was added from input base to earth, at first 100pF, now reduced to 47pF (a low-k NPO ceramic or a polystyrene type). Input and output capacitance are always present in inverting driver stages, and both contribute a phase lag. The feedback compensation capacitor has an effect called 'pole splitting' in which the two open loop phase lags, initially effective at similar frequencies, are replaced in the closed loop by one starting at low frequency, sometimes just a few Hz, and another at very high frequency, such as 10 MHz. It doesn't always work as well as that, but for the present circuit the intention is that the 2p7 feedback capacitor makes the stage into a simple integrator at high frequencies as far as the overall feedback is concerned, while the 47pF keeps the input stage local loop to a reasonable 6.6MHz unity gain frequency. The input stage closed loop second pole is calculated to be at 6.9MHz, 20 times the overall feedback loop unity gain frequency and therefore, in theory at least, harmless.
The amplifier is stable with any load tried, including pure capacitance from zero to 6uF. Stability is often specified with a 2uF load in parallel with 8 ohms, but stability into small values of capacitance must also be ensured because of the occasional misguided attempt to match cable impedance to speaker impedance, which could add significant capacitance. (The reputation for destroying amplifiers gained by high capacitance cables may also have something to do with a decreased separation between conductors and consequent increased risk of shorting if the cable is stood on or stressed in some other way.)
Low level high frequency oscillation was still seen on just part of a square wave test signal, and this was tracked down to the top half of the output stage. The use of three transistors in a feedback loop can be a problem, but in this case it is simple to add loop compensation from base to emitter of the bias stabilisation transistor which forms part of the output triple. At first 220pF was thought to be adequate, but small traces of instability could still be induced with difficult loads, and so this was increased to 680pF, which still leaves adequate output triple loop gain at high frequencies.
A small instability effect for about 5 usec while coming out of clipping was more difficult to cure. Calculations show that the 'pole splitting' mentioned above gives a second pole at about 6.9MHz, and the input stage loop unity gain frequency is 6.6MHz, but the instability was around 1MHz. The overall feedback from the output stage should have only a little effect at 1MHz, the 2p7 having more effect than the overall feedback at this frequency. Under normal operation there should be, and is, no stability problem. But, very close to clipping the gain of the input stage falls, and the second pole changes frequency and may add extra phase shift at 1MHz. This alone should not be enough, but one clue is that adding a small capacitor in parallel with the original single 200k feedback resistor gave continuous oscillation for even very small capacitance value. A 200k resistor is unlikely to be purely resistive at 1MHz, there may be enough parallel capacitance to be the source of the trouble. The single 200k is now replaced by two 47k plus a 20k to give the same closed loop gain but a lower impedance feedback network. The dc feedback also had to be changed, and this is now shown connected to the input capacitor instead of the input base, which was actually done from the start in the constructed version, so this is a correction rather than an update. Now the ringing effect has been reduced to insignificance by this resistor change, plus adding 5p6 to the common base stage, and adding 100nF across the output terminals to filter out any remaining low level high frequency. Although there would be no audible effect it is undesirable to have any high frequency signal reaching the output and potentially interfering with other equipment. At sufficiently high clipping levels other effects are found, but the distortion by this time is about 50%, and not relevant to normal audio operation where even moderate clipping should be avoided.
A 1uF input capacitor gave measured response -1dB at 25Hz, and the 1nF filter capacitor gives -1dB at 35kHz, but these both depend on signal source impedance, which was 600 ohms in these tests. Extending the response far beyond these frequencies has little point and will merely allow through interference from outside the audio range. Increasing the 1uF to 1.5uF gives -1dB at 16Hz and may be worthwhile if the speakers used have a very extended bass response. Personally I would also increase the 1nF to 1.5nF to give -1dB at 23kHz on the grounds that this almost certainly has no audible effect and extra interference suppression is more worthwhile.
The use of an output coupling capacitor usually gives poor bass damping factor, but this is avoided by including the capacitor in the overall feedback loop. There is then another danger which was pointed out to me, that low frequency loop stability needs to be checked, and there is a possibility of negative output resistance at some frequencies. Measuring output impedance as a function of frequency did reveal a very small negative resistance at 10 Hz, but only about -0.015 ohms, which is no problem, and in practice will merely cancel some of the speaker cable resistance. The effect of negative resistance is that adding a load increases the output voltage slightly. The damping factor at 15Hz was high at 650, reducing to 300 at 50 Hz upwards. The result of feedback loop low frequency phase shifts appears to have helped give an excellent bass output impedance. (By accident more than by design). Typical speaker impedances have an inductive component at low frequencies, below their resonance frequency, and this will add further phase shifts, but the inclusion of a 22uF capacitor in the dc feedback path shifts the phase in the opposite direction, which will help the stability margin, as does the effect of the input capacitor.
The output inductor value was not yet calculated or measured, but proved adequate for stability into capacitive loads. It has 13 turns in a single layer of 18swg (1.25mm dia) enamelled copper wire with inside diameter 1cm and length 2cm. It is air cored and must not have any other components near its ends.
This is still just an experimental version, not a final design. Anyone wanting to try the present circuit should remember that layout is as important as circuit design, and it is a good idea to take the earth connections via three separate wires to a common earth point so that low current, high current and non-linear current all go through different routes.
Inevitably some will ask what this amplifier sounds like, but this is purely a theory plus measurement based design. If I measured the distortion using equipment which added distortion of its own ten times greater than the amplifier my results would rightly be dismissed as meaningless, and yet typical speakers have distortion hundreds, or even a thousand times greater than that of the amplifier, so drawing conclusions about the amplifier from listening tests may be even more meaningless. One aim of the design has been to reduce crossover distortion, which is a type not produced by speakers, so this could still in principle be audible and its reduction may therefore be worthwhile.
My next project is a much simpler and more conventional design suitable for less experienced constructors rather than the more experimental feedforward designs. Again current fashions are disregarded and theory and measurement used as the guiding principles rather than the many conflicting and unsupportable claims about what 'sounds good.'