Hello everyone.
After the recent thread below, there has been interest in seeing how the circuit can be improved. This thread was created so that those not interested in changing the circuit can continue in the old thread.
http://www.diyaudio.com/forums/solid-state/174468-very-best-amplifier-i-have-ever-heard.html
This thread is to discuss improvements in performance, stability, and reliability.
Among the issues which were brought up:
1: The circuit can be unstable
2: Output stage bias has no clear adjustment
3: The topology can be improved
4: The frontend CCS is strange, it is hard to know what it was designed for. A more conventional CCS will be less confusing and more flexible.
5: Frontend rails may have sub-optimal power supply
6: Transistors used don't have adequate voltage specs for the rails
To start, I will offer my thoughts. These are not final, only to give you something to think about. I have attached several files. The first schematic is the original Goldmund amp. The second is the one containing the mods I describe here. The third attachment contains LTSpice simulations of them (where no models could be found for the original transistors, substitutions were attempted). These mods aren't my final word, I vomited them into the simulator at light speed, for lack of time. How to use the simulations: To change test frequency, change parameter "freq" - to change number of warm-up cycles, change parameter "dlycyc" - to change simulated number of cycles, change parameter "numcyc". When performing an FFT, select the same number of points as in the parameter "FFT". If the .MODEL statements are annoying, right-click them and select "invisible". They will reappear the next time you open the file.
Stability
Simulated, the frontend is actually VERY stable, showing textbook curves and well-defined behavior. The issues seem to lie in the output stage.
Output Stage Issues
Firstly, the output FETs have no emitter resistors. Since devices vary widely, and no equalizing force is present, the FETs probably won't share the load equally unless matched by hand.
Secondly, there are not trimmers included in the circuit for output stage adjustment, only one for offset adjustment. We do not know how - or if - the output stage was adjusted at the factory. Adding an adjustment for offset would go a long ways towards making the design more consistent and reliable between builds.
My suggested options for adjustment:
1: Use a trimmer for R20 first, and set bias to a reasonable value. Then turn off unit, measure trimmer and replace with a close value resistor.
2: Install a trimmer permanently, in such a way that a wiper disconnect will not destroy the output stage.
Topology Improvements
The general topology of the Goldmund amp has been seen many times over in the past. Many members of this forum have worked to refine it in their own time over the years, and I am sure many changes might be made.
While we may be tempted to go all-out, this would likely result in the amp being hardly recognizeable in the end. If this is so, why not simply create a whole new amplifier? In this light I would prefer to keep the design "close" to the original. After all, it is my observation that after a few powerful mods have been made, the leftover mods decrease in the ratio of performance gain to price. There is one thing that I suspect is important to the amp's sonic performance. I will copy a post I made:
Secondly, there are two modifications to the VAS which increase performance dramatically.
The transistors of the second LTP don't see an equal Vce, and so aren't balanced, even if they have the same quiescent. Without the second LTP being balanced, the current mirror is more or less for show. MikeB's Symasym cascodes this transistor to ground, and this balances the LTP at no signal. However the best way is to make the Vce's match. In my schematic I do this be cascoding the left transistor to the right's collector. This decreases distortion by a factor of 10, to .003%!
The second mod is considered after seeing Early affect on the lower VAS's linearity. At hundreds of volts Vce variation, it is not a good current mirror. One more transistor helps with this by shielding the mirror from large voltage swings. This brings THD down by a fraction, to .0025%.
CCS
The CCS was made in an odd way. I am not sure whether the designers' concern was noise or temperature stability, but one transistor is turned upside-down and it's B-C junction used as a diode. The behavior of this configuration will depend a lot on the specific transistor used, the originals being the BC182B. The Zener used is a 6V type, and IIRC zeners around this area have the lowest temperature coefficient. If we flip the odd transistor back the "right" way, tempco is fairly low among CCS's, given the two transistors are thermally coupled. This is the way I recommend. The output impedance of this configuration is not very high. However, because the gain of the amp is so high, the voltage variation across it is very low. Because of this, the contribution of the CCS to the overall performance is swamped by larger issues in other parts of the circuit. I don't believe going beyond this will reap great benefits, as far as the numbers go.
Frontend Rails
The frontend rails are constructed using a voltage doubler taken from a 60VAC supply. This results in ~120V rails. OS has described this as a "trainwreck"
Transistor Safety
For the frontend, rails are nearly +-120V. This means that at max power/clipping a transistor may see near 240V. Our transistors should be rated higher than this.
An apt alternative to using high-voltage transistors is to cascode. This way we can use common, familiar parts. In my mod, the MPSA93/43 are replaced by the suitably spec'd MPSA92/42, and output drivers are cascoded.
*catches breath*
Okay, I'm ready.
- keantoken
After the recent thread below, there has been interest in seeing how the circuit can be improved. This thread was created so that those not interested in changing the circuit can continue in the old thread.
http://www.diyaudio.com/forums/solid-state/174468-very-best-amplifier-i-have-ever-heard.html
This thread is to discuss improvements in performance, stability, and reliability.
Among the issues which were brought up:
1: The circuit can be unstable
2: Output stage bias has no clear adjustment
3: The topology can be improved
4: The frontend CCS is strange, it is hard to know what it was designed for. A more conventional CCS will be less confusing and more flexible.
5: Frontend rails may have sub-optimal power supply
6: Transistors used don't have adequate voltage specs for the rails
To start, I will offer my thoughts. These are not final, only to give you something to think about. I have attached several files. The first schematic is the original Goldmund amp. The second is the one containing the mods I describe here. The third attachment contains LTSpice simulations of them (where no models could be found for the original transistors, substitutions were attempted). These mods aren't my final word, I vomited them into the simulator at light speed, for lack of time. How to use the simulations: To change test frequency, change parameter "freq" - to change number of warm-up cycles, change parameter "dlycyc" - to change simulated number of cycles, change parameter "numcyc". When performing an FFT, select the same number of points as in the parameter "FFT". If the .MODEL statements are annoying, right-click them and select "invisible". They will reappear the next time you open the file.
Stability
Simulated, the frontend is actually VERY stable, showing textbook curves and well-defined behavior. The issues seem to lie in the output stage.
I'm never sure where to start when I consider designing an FET amp. Only special amps can drive hundreds of nF's of capacitance gracefully (the amplifier sees hundreds of nF if the capacitance is moved up to the driver stage as in an FET amp), plus there's a bunch of self-resonance and oscillation gotchas to watch out for. It must be hard for a simulator to accurately predict the behavior of a real FET amp.
The high degeneration on the differentials means very low phase distortion, which really helps stability when there is minimal compensation. That is, if you can account for the FETs' nonlinear capacitance. But I have no experience in this realm.
- keantoken
This circuit is also not unstable, barring any disastrous layout choices.
Open loop gain is a conservative 53db [according to simulation], with impressively well-defined HF behavior, thanks to the heavy degeneration. Phase margin is 75 degrees, which is not bad at all.
- keantoken
Jam, it is not unstable, I don't know why you would think that. Without the heavy (75 ohms!) degeneration on both LTPs it might be. See my post:
http://www.diyaudio.com/forums/soli...mplifier-i-have-ever-heard-9.html#post2328755
- keantoken
Hi Keantoken,
I agree that heavy degeneration of the diff. or VAS would help buy there are more factors at play here. We are not talking sims here as they are only a guide.
Question to be asked is how this heavy degeneration could affect the sound?
Regards,
Jam
Ok, I'm not an engineer and have no qualifications, and there's no reason for anyone to listen to me, but in simulation the circuit has very well-defined gain behavior. After billions of hours simulating, I am dead sure that this is almost solely because of the high degeneration resistors, which swamp nonlinear Gm and therefore linearize phase/pulse/HF response. Considering this, the frontend should be VERY stable, it's the power stage that is problematic. Am I right? Does this mean that the oscillation is more a problem with the FETs self-oscillating? It would take a very highly reactive network, like the high-Q stuff that comes from active components near oscillation, to throw the frontend off. Am I right?
- keantoken
Keantoken,
You are on the right track. My main problens were with the output stage. I have built the same amp with different layouts with some being stable and others not. All I am saying it test the board out first before releasing it.
Jam
Nagys,
You are totally off base. Your assumption that if use the same parts the amp will be stable. Well, what about the board material, layout, component spacing and location and list goes on, after all we are talkingf about a high speed circuit here.
[...]
Jam
Output Stage Issues
Firstly, the output FETs have no emitter resistors. Since devices vary widely, and no equalizing force is present, the FETs probably won't share the load equally unless matched by hand.
Secondly, there are not trimmers included in the circuit for output stage adjustment, only one for offset adjustment. We do not know how - or if - the output stage was adjusted at the factory. Adding an adjustment for offset would go a long ways towards making the design more consistent and reliable between builds.
My suggested options for adjustment:
1: Use a trimmer for R20 first, and set bias to a reasonable value. Then turn off unit, measure trimmer and replace with a close value resistor.
2: Install a trimmer permanently, in such a way that a wiper disconnect will not destroy the output stage.
Topology Improvements
The general topology of the Goldmund amp has been seen many times over in the past. Many members of this forum have worked to refine it in their own time over the years, and I am sure many changes might be made.
While we may be tempted to go all-out, this would likely result in the amp being hardly recognizeable in the end. If this is so, why not simply create a whole new amplifier? In this light I would prefer to keep the design "close" to the original. After all, it is my observation that after a few powerful mods have been made, the leftover mods decrease in the ratio of performance gain to price. There is one thing that I suspect is important to the amp's sonic performance. I will copy a post I made:
Wait, I didn't see your last response Jam.
I stated in my last response my observations on the the effect of the degeneration on phase behavior. While they decrease open-loop gain, they linearize phase behavior, at least for the frontend [by swamping nonlinear Gm]. This leads me to think they would improve imaging and soundstage. How they affect other aspects of the sound, I am not sure (hell, I'm not sure of any of this though). Due to the lowered OLG, the distortions of the output stage will be emphasized. This may not seem to be a big problem if the 2SK1058/J49 give benign distortions. If the FET parasitic capacitances are too nonlinear, they may ruin the phase behavior when played loudly.
Of course, it may be more difficult to tell how this behavior will change after the crossover threshold is reached.
- keantoken
Secondly, there are two modifications to the VAS which increase performance dramatically.
The transistors of the second LTP don't see an equal Vce, and so aren't balanced, even if they have the same quiescent. Without the second LTP being balanced, the current mirror is more or less for show. MikeB's Symasym cascodes this transistor to ground, and this balances the LTP at no signal. However the best way is to make the Vce's match. In my schematic I do this be cascoding the left transistor to the right's collector. This decreases distortion by a factor of 10, to .003%!
The second mod is considered after seeing Early affect on the lower VAS's linearity. At hundreds of volts Vce variation, it is not a good current mirror. One more transistor helps with this by shielding the mirror from large voltage swings. This brings THD down by a fraction, to .0025%.
CCS
The CCS was made in an odd way. I am not sure whether the designers' concern was noise or temperature stability, but one transistor is turned upside-down and it's B-C junction used as a diode. The behavior of this configuration will depend a lot on the specific transistor used, the originals being the BC182B. The Zener used is a 6V type, and IIRC zeners around this area have the lowest temperature coefficient. If we flip the odd transistor back the "right" way, tempco is fairly low among CCS's, given the two transistors are thermally coupled. This is the way I recommend. The output impedance of this configuration is not very high. However, because the gain of the amp is so high, the voltage variation across it is very low. Because of this, the contribution of the CCS to the overall performance is swamped by larger issues in other parts of the circuit. I don't believe going beyond this will reap great benefits, as far as the numbers go.
Frontend Rails
The frontend rails are constructed using a voltage doubler taken from a 60VAC supply. This results in ~120V rails. OS has described this as a "trainwreck"
Transistor Safety
For the frontend, rails are nearly +-120V. This means that at max power/clipping a transistor may see near 240V. Our transistors should be rated higher than this.
An apt alternative to using high-voltage transistors is to cascode. This way we can use common, familiar parts. In my mod, the MPSA93/43 are replaced by the suitably spec'd MPSA92/42, and output drivers are cascoded.
*catches breath*
Okay, I'm ready.
- keantoken
Attachments
Last edited: