Audio amplifier output stage comparison


There are countless options for amplifier power output stages. Most often emitter follower output stages, either as dual or triple emitter follower configurations are used. Diamond buffer output stages are used in operational amplifiers and headphone amplifiers a lot, but I haven't seen this circuit used for power amplifiers often and wonder why. I tend to swim against the (main-) stream and this is why not following the trend of using an emitter follower output stage seems appealing to me. In this article, I will compare the dual and triple emitter follower output stages with two and three stage variants of a Diamond output stage based on circuit simulation. Circuit simulation is only an approximation and results depend on model accuracy a lot, but may indicate real life performance to some extend.


Below schematics show the circuits I set up for investigation. Note that this variant of the Diamond Buffer requires elevated supply rails for the drivers due to the current sources I chose. The emitter follower stages do not require elevated supply rails, but have elevated supplies nonetheless in order to have mostly equal conditions. The current sources powering the Diamond Buffer stem from an earlier investigation of high performance current sources. The current sources used in this example require slightly over 4V to work, so the supplies are elevated by 5V for some margin. Both output stages are set up with roughly 100mA standing current in the driver stage and 1A in each power transistor. Given a +/- 30V supply, this results in a powerful class A output stage for 8Ω (purely resistive) loads. This excludes any crossover distortion from the results, but also hides handling of such. Suitability for class AB operation is another topic. Power transistor base resistor is 2.2Ω and emitter resistor is 0.22Ω.

Two stage emitter follower

Three stage emitter follower

Two stage Diamond buffer

Three stage Diamond buffer

DC bias

Stable DC bias is less important for class A output stages, like for the setup in this example. It matters most for output stages biased into class AB. I probably will investigate this case as well, but in another article. The main benefit of class AB bias is much higher efficiency and higher output power, which comes at the expense of higher distortion and, more important, this type of distortion is crossover distortion, which is entirely unrelated to the audio signal and therefore most objectionable.

The bias spreader of the emitter followers and the three stage Diamond buffer looks unusual, but is actually nothing special - just a complimentary bias spreader. In some cases I added extra diodes held at constant temperature to lower the compensation factor while increasing the DC bias current. The typical bias spreader (like shown for example in figure 14.10 of Bob Cordell's book) that compensates six VBE, totally overcompensates in simulation. Although that does not mean that this will happen in real life as well, some options to tweak compensation are nice as this may help to get the bias stable both in simulation and reality. There are literally countless possibilities to get the emitter follower DC bias stable. The decision is not only which bias spreader arrangement to chose, but also, which transistors to put on common or separate heat sinks and whether or not to sense the heat sink temperature and so on. This cannot really be assessed in simulation, but needs a lot of fine-tuning on the bench. Setting DC bias of the dual emitter follower is done by varying R144, for the triple emitter follower by R52 and for the three stage Diamond by R117 in this example. This resistor also sets the compensation factor, which complicates setting the bias. This is why I introduced D5 and D6 for the triple emitter follower: To elevate the bias into class A without affecting the compensation factor.

The DC bias of the Diamond buffer is set by the driver transistors, which need to be on the same heat sink as the power transistors in order to sense temperature. Using cascoded current sources, the cascode transistors can be put on the same heat sink. Compensation factor is much lower and it is yet to see whether this is sufficient in real life. Power transistor DC bias depends on resistors R11 and R12, as well as the standing current of the driver stage. For setting the typical class AB bias current, probably compromises in driver stage standing current needs to be made.

The first two stages of the three stage Diamond buffer in theory should cancel their temperature coefficient in case Q65 and Q40 as well as Q66 and Q39 are thermally coupled. The bias spreader senses the third stages transistors for compensation. However, this does not work for me in simulation. This does not mean it does not work in general or in real life. This just indicates that this might be complicated. Maybe my simulation setup is just wrong.


Resistance of the input voltage source for the stability analysis was stepped with 10Ω, 100Ω, 1kΩ and 10kΩ.

Dual emitter follower

This is the dual emitter follower stability uncompensated, i.e. with R1 and C1 shunt compensation removed:

The dual emitter follower is a pretty stable from AC point of view even without any compensation applied.

Note that the compensation is far from perfect, but illustrates the effect. In order to optimize compensation, the input voltage source impedance needs to be considered. All output stages have the same compensation network applied. In case of the dual emitter follower, shunt compensation does not yield any significant improvement.

Triple emitter follower

This is the triple emitter follower stability uncompensated, i.e. with R1 and C1 shunt compensation removed:

The plot shows that the uncompensated triple emitter follower is outright unstable with any impedance of the signal voltage source.

This is the effect of the shunt compensation network on stability:

The shunt compensation makes a huge difference and now stability has improved dramatically.

Two stage Diamond buffer

This is the diamond buffer stability uncompensated, i.e. with R2 and C2 shunt compensation removed:

The Diamond buffer shows only slight instability when left uncompensated.

This is the effect of the shunt compensation network on stability:

Much better now.

Three stage Diamond buffer

This is the diamond buffer stability uncompensated, i.e. with R2 and C2 shunt compensation removed:

This configuration is even worse than the triple emitter follower and somehow this was expected.

This is the effect of the shunt compensation network on stability:

Unlike the triple emitter follower, adding the shunt compensation does not improve stability into the usable range.

Triple emitter follower optimized compensation

While the triple emitter follower is praised for excellent performance all the time, almost nobody really explains how to stabilize this beast. As shown before, some shunt compensation does improve the situation, but there is further room for improvement. Since shunt compensation has helped ahead of the first stage, it is somewhat obvious to apply this to the second stage as well. Here, it is advisable to shunt to the collector instead since this also mitigates the base-collector capacitance, which is dependent on VCB. So this both acts as a snubber to reduce any ringing of CCB and both swamps CCB, which makes the situation more predictable. This is the trick to get the triple emitter follower stable on paper. Further tricks are low value resistors in the middle stage collectors in order to tame feedback via the power supply and adding R-C shunts to the third stage base to collector as well. The shunts in the third stage should be minimized. Values around 47pF and 100R are pretty good to start with. Shown on the schematic and important in reality is also the R-C shunt at the output. This does not really affect stability in simulation with resistive loads.

Below is an example of an optimized three stage Diamond buffer, which showed worst instability earlier:

Now the gain peak at roughly 8MHz is gone. The second stage shunt compensation addresses this gain peak, while the shunt compensation ahead of the first stage causes the roll-off in the lower frequency range.

Emitter follower stability problems explained

So far, only small signal analysis of the output stages was discussed. This is based on many simplifications and therefore inaccurate. In reality, the situation is much more complex. Emitter follower stages have many inherent stability problems. Here is what I learned and collected during my research:

Once the signal approaches the supply rails, the COB increases dramatically.

In output stages biased into class AB, the current through the transistors changes with signal amplitude and rises once the signal swings close to the power supply rails. This causes FT to drop.

Wiring inductance easily forms oscillator circuits with parasitic capacitance. Resistors help to spoil the Q-factor and this is one reason why resistors at the base (or gate) are important. However, for BJT, the base resistor should not be larger than ten times the emitter resistor value.

Emitter followers are destabilized by any capacitive loading at the output. This is why there is an inductor at the output of an amplifier: To avoid capacitive loading of the output stage.

Emitter followers output impedance increases with frequency, but they don't like to be driven by such an output impedance. Thus, chaining emitter follower stages inherently leads to problems.

Emitter follower stages are impedance converters. A single emitter follower turns resistive source impedance to inductive output impedance. A dual emitter follower turns the resistive source impedance to inductive2, which is unstable even with the lowest capacitive load. Placing a capacitor at the input converts the input impedance to capacitive and returns the output impedance to inductive1.

The previously shown stabilization of the triple emitter follower was achieved by slowing down the second stage adding base to collector capacitance that works in conjunction with the base resistance. Another way would be to slow down the last stage (by selecting slow transistors for example) and in turn speed up the second stage by paralleling multiple fast transistors like the SC3503 / SA14381 pair instead of the rather slow MJE15032 / MJE15033 pair.


Assessing distortion in simulation is difficult, but the result is good enough for a rough estimate. Below table shows distortion at 1kHz near clipping at 30V supply rails with 8Ω resistor load.

Input resistance Dual emitter follower Triple emitter follower Two stage Diamond buffer Three stage Diamond buffer
10Ω 0.002% 0.002% 0.006% 0.002%
100Ω 0.004% 0.002% 0.008% 0.005%
1kΩ 0.025% 0.003% 0.028% 0.002%
10kΩ 0.194% 0.009% 0.173% 0.005%

It is obvious that the triple stage output stages performance is mostly independent from input resistance while the two stage OPS shows a strong dependence. At 10kΩ input resistance this goes so far that visible attenuation of the two stage OPS output signal can be observed. This shows that the input impedance of the three stage OPS is much higher.

The numbers from simulation do not make much sense to me because I expected distortion being higher some orders of magnitude. I assume that the models and setup used for simulation are somewhat unrealistic. With distortion numbers that low, I wonder whether it could be a good idea to run the OPS in class A and exclude it from the feedback loop to just run it open loop instead. However, an open loop OPS has low damping factor, which might play well with a certain set of loudspeakers or not. Since most amplifiers have high damping factor, most loudspeakers are likely optimized to play well with those instead.


There is a trade-off between performance and stability. The three stage OPS offer more performance while showing potential stability issues. In theory, the Diamond buffer arrangement should have many benefits, but this does not show in simulation. I had a bias when starting the investigation that there must be something magic and superior to the Diamond configuration. The name alone implies something precious. This did not hold true in simulation, but I will for sure build a Diamond output stage to see and hear how it performs in reality.