This is one of the amplifier front ends for the modular audio amplifier system. This front end is heavily inspired by Bob Cordell's MOSFET amplifier with error correction and his excellent book about designing audio power amplifiers. I learned a lot from his book and tried to apply that in this design, plus some rather unusual design decisions on my own.
- Extremely high power supply rejection
- Power supply invariant operation with wide operating range
- Input over-voltage protection
- Negative feedback loop stays intact during heavy clipping
- Adaptive clipping by sensing the output stage power supply
The schematic looks more complicated than it actually is. The first page shows all circuits that are not in the signal path. The second page is basically the front end of Bob Cordell's amplifier with some nice extras added.
The front end requires two floating 15V supplies in order to work properly. Without extra supplies for the front end, voltage headroom is pretty limited. The floating supplies are added on top of the power amplifier supply rails and are then filtered by capacitance multipliers. Bob's design already features high and symmetric PSRR so extra filters are not absolutely necessary, but additional filters do no harm and improve PSRR significantly.
The schematic shown here is suitable to build amplifiers of any power rating. It will run fine with output stage power supply rails as low as +/-12V and as high as +/-85V. For a very high power design, I would change some resistor values to reduce dissipation. For supplies over roughly 60V, all 18kΩ resistors (R45, R46, R107-R110, R161 and R162) should be roughly 27kΩ. Bias is almost supply voltage invariant because the CCS used in this design are extremely stable. Heat sinks are optional and can be left away for low supply voltage operation.
The rail voltage reference circuits derive a filtered reference mixed from both supply rails, which is used as reference for the second LTP cascodes and the feedback Baker clamp circuitry. This adaptive reference in combination with the feedback baker clamps ensure that the front end and subsequent stages cannot be driven into saturation. Even in case of severe overdrive, the front end remains linear with the feedback loop intact.
Three constant current sources bias the front end. Two of the CCS drive the long tailed pairs and one is just for the cascode reference of the first LTP. The CCS design is based on my previous investigation of constant current sources for audio applications.
I added a CCS to the second LTP because this makes the second LTP standing current independent from the first LTP current. The CCS used for the first LTP are very stable, but having the ability to set both currents independent from each other makes experiments much easier. Also, the output is now current limited, which might be seen as advantage or disadvantage dependent on point of view. I see current limiting as an advantage because it helps to prevent destruction in case of some potentially destructive events.
The input filter offers many options: C3 is a huge film capacitor, which is my preference. As an alternative, a bipolar electrolytic capacitor (C4) can be installed with a small film capacitor (C5) in parallel optionally. Resistor R3 ensures that no DC builds up. R2 and C2 may be used to properly terminate the audio input cable. They should only be installed in case a long cable connects to the amplifier input directly and be omitted inc case any pre-amp or audio input transformer is used. R4, R5 and C6, C7 form a second order low pass filter with a corner frequency of 100kHz. Most amplifiers use a first order filter here, but I find that a second order filter has better performance. Resistors R4+R5+R6 sets the input impedance. I chose 20kΩ as a middle course between 10kΩ often used for professional audio interfaces and 50kΩ for consumer audio electronics. R1 and C1 are the usual measures to break ground loops and may be replaced by a direct connection instead.
The amplifier input protection I developed earlier, is implemented here. This is supposed to be a robust amplifier building block. In case no protection is required or desired, the components can be left away easily.
The input LTP allows to use CFPs instead of single transistors. This improves linearity of the LTP. I like to build it this way since more inherent linearity prior to the application of negative feedback is a good idea in my opinion. Also, the LTP is heavily degenerated to further increase linearity. A tail current of 8mA ensures high transconductance is maintained. Since the LTP transistors are cascoded, a broad range of transistors can be used. Transistors Early effect is also reduced by cascoding. In case the CFP should not perform as desired, Q23 and Q24 can be removed and R13 and R14 be bridged in order to fall back to a normal LTP design.
One minor downside of the CFPs here is that this costs a tiny bit of phase margin. Negative feedback becomes less effective at higher frequency, while the inherent linearity is still there.
Hugh from AKSA confirmed that CFPs inside LTPs work well an he has successfully done that before in his amplifier designs. The CFPs linearize the gm of the active devices so that the horizontal S curve is flattened. This is especially effective in for high amplitude signals. So after I took the decision to use CFPs, I learned that this was a good idea and why.
The CFP arrangement also reduces memory distortion. I recommend to study the work Pierre Lacombe has presented.
I added transistor Q13 as floating cascode into the second LTP mainly for distributing power dissipation in case of high voltage application. Q13 could be removed and C-E shorted in case this is not desired. However, cascoding the cascode makes the load more symmetric, which I believe is a good decision.
The feedback baker clamps are explained in Bob's book and I can't explain this topic nearly as well so I refer to the book at this point. The basic idea is to avoid saturation of the amplifier by diverting excessive signal swing beyond a certain threshold back to the inverting input of the amplifier where it combines with the rest of the (clipped) signal. This way the feedback loop stays intact even under heavy clipping conditions. Saturation and recovery issues of transistors are therefore avoided.
At the time of writing, some of the transistors in the schematic are obsolete. This does not really matter since the design does not seem to depend on specific transistor types too much.
All BC547/BC557 transistors stand for just any complimentary small signal transistors that may be available.
All BC550C/BC560C stand for low noise transistors can be substituted by any complimentary high hFE and high fT low noise transistors that may be available.
The SC3503/SA1381 transistors are unique, but the design also simulates well using the TTA004B and TTC004B transistors. Using those transistors instead results in a tiny loss of phase margin. Also, the SOA of the TTA/TTC transistors is much lower and this limits the possible supply rail voltage. I use both types in the schematic to indicate which ones are preferred to have low Cob like the SC3503/SA1381 feature.
The MJE340/MJE350 transistors were chosen due to their excellent SOA.
In general, for low power designs, lower power transistors could be used that are easier to find having good specifications. The components in the schematic and their operating conditions shown here are suitable for amplifiers up to several hundreds of Watts output power.
Miller input compensation is one of the best compensation techniques for this kind of front end. I wanted to keep much of the Cordell spirit alive and so I kept it. I found getting the amplifier stable very challenging and it took a good deal of simulation and experimentation to obtain a stable amplifier.
The difficulty is that six components are involved in this compensation technique. This is the most complex compensation technique known to me.
Simulation is only partially helpful because this results in a vast variety of setups that appear stable in simulation, but oscillate terribly on the bench. In case the compensation is off, the amp shows sustained several MHz and high amplitude oscillation in idle condition and any other condition. I smoked several components until I finally succeeded.
When iterating compensation on the bench, the question is where to start at all. Bob's book (first edition) does not explain how to calculate or iterate Miller input compensation. I found some guidelines Bob has shared on DIY Audio that I copied here for reference with component names adjusted to my schematic so that it makes sense in this context:
Let's assume I've chosen a closed loop gain of 20 and a gain crossover frequency of 2MHz. This means that I want the forward gain of the combined input stage and VAS to be 20 at 1MHz, assuming that the output stage has unity gain (if the output stage had a gain of 2, the target would be 10 instead). Having chosen the feedback shunt resistor (R64 + R65) to be 215 ohms, I now know that the reactance of C81 must be about 19 times 215 Ohms at 1 MHz, or about 4k. Thus C81 becomes 20pF.
I then want to put in a stabilizing zero at a frequency above the gain crossover frequency. This is done with R81 at 680 ohms. This gives me a zero at about 11 MHz.
Next I analyze and stabilize the inner loop that is closed by C81 and R81. This can be a high-bandwidth loop with a gain crossover beyond 10 MHz, since only small-signal fast transistors are in the loop. This fast loop is stabilized by C80 and R80. This is very light lag compensation.
Note that there is not a lot of interaction at the base of Q2 because the 215 ohm resistor is small compared to the impedances of C81 and R81.
The advantage of the scheme is that it rolls off the high frequency forward path gain by applying feedback around the input stage, increasing its dynamic range, reducing its distortion, and not making it work harder at high frequencies (the way Miller compensation does). It only has to work harder at very high frequencies where C80 comes in, but this is pretty far out.
The ULGFs of the major and minor loops are so far apart, perhaps a decade, that the major loop feedback contribution to the minor loop gain of the MIC is pretty much inconsequential in the frequency range of the minor loop ULGF.
Compensating the MIC minor loop can be tricky because there are so many possible combinations of the resistance and capacitance values. However, a simple rule-of thumb approach can be used to converge on a circuit fairly quickly. There are three R-C pieces to the design. The first is C81, the MIC dominant pole compensation capacitor against feedback resistor R60 to R63 and feedback shunt R64 + R65 (ignore R81 for now and let it be zero, as it often can be). The global ULGF is set by making C81s reactance equal to R60-R63 at the desired global ULGF.
The second R-C piece is the IPS shunt lag-lead load consisting of C80 and R80. This is the primary compensation network for the MIC loop. For now, you can assume R80=0Ω.
The third piece is the VAS R-C load consisting of C82 and R82. In some designs you can get away without it, but I strongly recommend it. It establishes a predictable and dampening load for the VAS at high frequencies. It also controls what proportion of the VAS signal current flows back through C81. Indeed, for most designs you can arbitrarily set C82 to 50pF. If you make it too big, you will impair the achievable slew rate of the VAS. Now, select R82 to set the time constant of C82*R82 to be the same as the time constant of C81 and the resistance it has in series with it. The division of VAS signal current between the two paths (through C81 and C82) will now be largely flat with frequency.
Now just set the value of C80 (I just do it by simulation) to achieve the desired minor-loop ULGF. I usually choose about 10MHz. Optionally add some resistance at R80 to optimize phase and gain margin of the minor loop. Resistance can be added at R81 to add some phase lead to optimize the global gain and phase margin, if desired. If resistance is added at R80, one may wish to appropriately revise the value of R82. Note that even with R81 = 0Ω, the non-inverting amplifier structure of the MIC loop prevents its gain from going below 0dB, so a zero is created there no matter what.
With R81=1k, the forward open-loop gain of the compensated amplifier cannot go below +6dB (i.e., X2) in theory. If the closed loop gain of the amplifier is 20, and ULGF is 1MHz, the zero will occur where the closed loop gain will have fallen to 6dB at 6dB/octave. This will be one decade above 1MHz, i.e., 10MHz.
If there is peaking in one of the local loops, it can cause peaking in the overall open loop gain, and this will inevitably result in peaking in the closed loop gain at frequencies above the ULGF.
I found that all six components are required and need to be optimized. In case any component is not optimized, the amplifier will oscillate full bore.
Unfortunately, the first revision of the PCBs lacked some features that I found out to be necessary during evaluation of the prototype on the bench.
Most important, I forgot to add the shunt compensation to the node where the feedback Baker clamp diverts excess signal to. Thus, R83 and C83 were missing in the first revision. Luckily, adding R83 and C83 does not make compensation even more complicated because the values of R82 and C82 should be copied. R70 is required to isolate the shunt compensation from the rest of the feedback network.
I added R51 to R54, which are required for loop stability during clipping. The mechanism behind remains unknown to me. I just figured out they are required and need to have unusual high value. Note that C113 and C114 were also additions in revision two and just act as a filter together with the resistors connected to this node. C113 and C114 are optional, but should never be installed without R51 to R54.
I changed the voltage offset generation of the CCS cascodes from LEDs to resistors due to concerns about stability of the cascodes. Resistors are cheaper anyway and high impedance drive of the cascodes is better for stability in any case. The first revision using LEDs did not show any such instability though.
Another small upgrade was to change the Zener diode used as floating cascode reference in the input LTP. Now LEDs D10 to D12 generate the reference and have lower current noise than a Zener diode. Those LEDs were saved in the CCS earlier so it was a zero cost upgrade.
Simulation of power supply rejection shows inherently good and symmetric rejection of both rails. The addition of R-C filters of 10Ω and 82µF for the front end supplies help to avoid loss of power supply rejection at higher frequency. This shows how much effect on performance a few cheap components can have. With the active filters, power supply rejection is further improved considerably. Grunge from the power supplies does not contribute to most genres of music so I prefer to minimize this as far as possible. Most recordings unfortunately feature enough annoying background noise already.
Bob wrote an entire chapter in his book about "Clipping control and civilized amplifier behavior", where he also explains the feedback baker clamps as one possible solution. I find the idea behind feedback backer clamps very clever and since I haven't seen anybody implementing this before, I need to give this concept a chance.
During clipping, a lot of very nasty things happen to an amplifier unless counter-measures are implemented. First, parts of the circuit are driven into saturation, which just means that they are no longer functional. Apart from this, the negative feedback loop tries to combat the extreme deviation of the output signal from the input signal and uses all the crazy high open loop gain to do this. So the control loop gets out of control entirely trying to get a partially non-functional amplifier to attempt the impossible. Clearly, this is not going to end well. I added some illustrations depicting what might happen in such situations.
Feedback baker clamps return the clipped part of the signal back to the inverting input of the amplifier and so the feedback loop actually does not clip, but stays intact and retains control of the amplifier. The output signal waveform is clipped however and together with sensing the output stage supply voltage and dynamic adjustment of the clipping threshold, saturation of the output stage can be avoided as well. In above examples, the output stage supply voltage is +/-30V and clipping occurs with roughly 2V margin to the power supply rails. The overall result is a well-behaved amplifier that tolerates overdrive without freaking out.
Note that Miller input compensation has sub-optimal clipping recovery behavior and avoiding saturation is a big advantage in this case.
Here is an example configuration together with the folded driver module and the four pairs BJT output stage mounted to one of the radiators of the chassis. Inter-module connections are mostly complete, only missing the negative feedback connection between front end and driver module. The power supply would be connected on the upper right hand side to the output stage. The loudspeaker would be connected to the lower right hand side of the output stage and two auxiliary floating power supplies would be connected to the front end PCB. Audio input is on the upper left hand side at the place where the BNC adapter occupies the SMA connector.
Testing the module is ongoing. I succeeded to engineer and improvise the revision 2 updates on the revision 1 PCBs and managed to reach the main objectives. Listening test is pending.