Sander,
Yes, I agree LTSpice is quite a nice computer game (I call it QSZ -- Quest for Seven Zeros
). BTW, are there models for the Sanken RET's around?
Find attached the Aol bode plot (no load), again seems to fit well. Textbook perfect 60deg+ phase margin with Cc=33pF... but it will oscillate with CL=7nF, that's the point where it starts in the .tran and then I see 0 deg margin, so I broke the loops at the right point. Was a little tricky, because of the parallel loops. I had to tweak Cc balance a little -- PNP side -20% -- to get rid of a ripple in the phase.
I'm sick today and have a day off, could't sleep the whole night... and have a bag full of tweaking ideas:
- cascode the bias transistors to beef up gain and high-Z node impedance
- use laterals for outputs because they have more linear gm and spread the diff voltage needed for the casodes
- sense output current with a wheatstone bridge and use it for a little positive feedback to lower output impedance (could even be made adjustable, also with increased Rout for speakers designed for that)
- use double pole compensation and/or a detuned T network to stabilize the excess gain
- ...
(some of that I did quicktest already)
- Klaus
Yes, I agree LTSpice is quite a nice computer game (I call it QSZ -- Quest for Seven Zeros
). BTW, are there models for the Sanken RET's around?Find attached the Aol bode plot (no load), again seems to fit well. Textbook perfect 60deg+ phase margin with Cc=33pF... but it will oscillate with CL=7nF, that's the point where it starts in the .tran and then I see 0 deg margin, so I broke the loops at the right point. Was a little tricky, because of the parallel loops. I had to tweak Cc balance a little -- PNP side -20% -- to get rid of a ripple in the phase.
I'm sick today and have a day off, could't sleep the whole night... and have a bag full of tweaking ideas:
- cascode the bias transistors to beef up gain and high-Z node impedance
- use laterals for outputs because they have more linear gm and spread the diff voltage needed for the casodes
- sense output current with a wheatstone bridge and use it for a little positive feedback to lower output impedance (could even be made adjustable, also with increased Rout for speakers designed for that)
- use double pole compensation and/or a detuned T network to stabilize the excess gain
- ...
(some of that I did quicktest already)
- Klaus
Attachments
Klaus,
No, or at least I have yet to find them, I used different libs and edited them a little to get close, I think OrCad libs came the closest. The output stage was then subsequently tested on the bench and final values for the components were selected.
I'm sorry to hear that, but that of course means you can tinker with LTspice all day long
Obviously there is room for improvement still. But we set out to reach a certain performance, and when we comfortably reached that level we did not bother to optimize further. As-is the ExtremA provides THD levels that are a far cry from the vast majority of commercial amplifiers, but for Halcro, but those amplifiers come at a hefty pricetag.
Ps. If you want to talk LTspice a little more, ping me at ssassen[at]hardwareanalysis[dot]com.
Cheers,
Sander Sassen
http://www.hardwareanalysis.com
No, or at least I have yet to find them, I used different libs and edited them a little to get close, I think OrCad libs came the closest. The output stage was then subsequently tested on the bench and final values for the components were selected.
I'm sorry to hear that, but that of course means you can tinker with LTspice all day long
Obviously there is room for improvement still. But we set out to reach a certain performance, and when we comfortably reached that level we did not bother to optimize further. As-is the ExtremA provides THD levels that are a far cry from the vast majority of commercial amplifiers, but for Halcro, but those amplifiers come at a hefty pricetag.
Ps. If you want to talk LTspice a little more, ping me at ssassen[at]hardwareanalysis[dot]com.
Cheers,
Sander Sassen
http://www.hardwareanalysis.com
Sander,
Your amp is without doubt a very impressive statement WRT to distortion performance. As far as understand it uses heavy global feedback as an element to get there (I didn't input the front-end so far). My intend, besides playing the "computer game" for fun and self-education, is to design an OS that is "perfect" on its own and then use it with a simple front-end without GFB, maybe just as simple as an excellent step-up transformer for the VAS, with optimized drive. That would possibly "ruin" the perfomance, but I tend to prefer the feedback handling those uncorrelated speaker currents be as localized as possible, plus using the "poper feedback priniple"; that is, use it only to linearize large-signal errors of an otherwise already very linear stage and try to linearize the "actuator control signal" as much as possible before closing the loop (if that is the correct term in english).
- Klaus
Your amp is without doubt a very impressive statement WRT to distortion performance. As far as understand it uses heavy global feedback as an element to get there (I didn't input the front-end so far). My intend, besides playing the "computer game" for fun and self-education, is to design an OS that is "perfect" on its own and then use it with a simple front-end without GFB, maybe just as simple as an excellent step-up transformer for the VAS, with optimized drive. That would possibly "ruin" the perfomance, but I tend to prefer the feedback handling those uncorrelated speaker currents be as localized as possible, plus using the "poper feedback priniple"; that is, use it only to linearize large-signal errors of an otherwise already very linear stage and try to linearize the "actuator control signal" as much as possible before closing the loop (if that is the correct term in english).
- Klaus
Klaus,
No, no, that's just pointing fingers as if this is a revival of the early '80s designs that had horrible performance, this amplifier does *not* subscribe to the notion that feedback is a tool to lower THD. Actually there's a modest amount of global feedback used with 26dB of gain (20x).
Your simulations have already shown that output stage THD alone is nothing short of spectacular, hence you can rest assured it is teamed up with a front stage that is nothing less spectacular in terms of performance.
Cheers,
Sander Sassen
http://www.hardwareanalysis.com
No, no, that's just pointing fingers as if this is a revival of the early '80s designs that had horrible performance, this amplifier does *not* subscribe to the notion that feedback is a tool to lower THD. Actually there's a modest amount of global feedback used with 26dB of gain (20x).
Your simulations have already shown that output stage THD alone is nothing short of spectacular, hence you can rest assured it is teamed up with a front stage that is nothing less spectacular in terms of performance.
Cheers,
Sander Sassen
http://www.hardwareanalysis.com
Hi Sander,
This might be a misunderstanding here. I'm aware that the OPS is a precision buffer and doesn't need additional corrective global FB. But, what I interpreting from the schematic (and I might be wrong, of course), the front-end itself is open-loop and has some 60dB open-loop gain (@DC), so there is ~40dB of corrective feedback for the front end distortion (which is low, being class A) as Acl is ~20dB. Isn't the front structure "undegenerated LTP --> folded cascode --> buffers, fully differential", thus has the LTP gain? So, this 40dB (order of magnitude) of feedback I interpreted as "heavy global FB", albeit wisely used, to large-signal linearize something already linear and quasi-distortion-free at small signal. One could as well tap the feedback off at the input of the OS, only loosing the decrease in output impedance but not with any significant distortion penalty.
- Klaus
This might be a misunderstanding here. I'm aware that the OPS is a precision buffer and doesn't need additional corrective global FB. But, what I interpreting from the schematic (and I might be wrong, of course), the front-end itself is open-loop and has some 60dB open-loop gain (@DC), so there is ~40dB of corrective feedback for the front end distortion (which is low, being class A) as Acl is ~20dB. Isn't the front structure "undegenerated LTP --> folded cascode --> buffers, fully differential", thus has the LTP gain? So, this 40dB (order of magnitude) of feedback I interpreted as "heavy global FB", albeit wisely used, to large-signal linearize something already linear and quasi-distortion-free at small signal. One could as well tap the feedback off at the input of the OS, only loosing the decrease in output impedance but not with any significant distortion penalty.
- Klaus
Let's assume that distortion from metallfilmresistors are real, which resistors carry ac current on the amp PCB of ExtremA?
The resistors I have bought are standard metallfilms that are magnetic with +/-50ppm/c tempco and I know there are types that are non-magnetic and also has better tempco, like +/-15ppm/c.
Or should I just solder on and stop being anal?
/Peter
The resistors I have bought are standard metallfilms that are magnetic with +/-50ppm/c tempco and I know there are types that are non-magnetic and also has better tempco, like +/-15ppm/c.
Or should I just solder on and stop being anal?
/Peter
Peter,
Yes, I gave you a few alternatives which you all deemed too expensive, so just use what you've got or pick one of the more expensive alternatives.
Cheers,
Sander Sassen
http://www.hardwareanalysis.com
Yes, I gave you a few alternatives which you all deemed too expensive, so just use what you've got or pick one of the more expensive alternatives.
Cheers,
Sander Sassen
http://www.hardwareanalysis.com
Peter,
Of course, obviously those are in the feedback network, hence 2K2 and 47K.
Cheers,
Sander Sassen
http://www.hardwareanalysis.com
Of course, obviously those are in the feedback network, hence 2K2 and 47K.
Cheers,
Sander Sassen
http://www.hardwareanalysis.com
Great looking design, I'd build that instead of a Pass / Krell clone.
This would be the paper it is based on http://www.grimmaudio.com/whitepapers/balanced building blocks.pdf
Has some resemblance to a Pass super symmetric design I think.
This would be the paper it is based on http://www.grimmaudio.com/whitepapers/balanced building blocks.pdf
Has some resemblance to a Pass super symmetric design I think.
From my point of view I'd say Dave is right (but the cascode is NOT the key point)... a fully differential op-amp is "super-symmetric" by design, blips of distortion (Pass lingo) appearing at only one output are reduced by feedback at both outputs because of the coupling at LTP emitters (that's the "X-thing"). That's why TI seems to pay royalties to Mr.Pass for the OPA1632 (at least that's what I read from another EE).
- Klaus
(PS.@Sander: I'm happily simming variations of the OS... some things work great, others don't. I'll post the more reasonable variations as soon as I dare ;-)
- Klaus
(PS.@Sander: I'm happily simming variations of the OS... some things work great, others don't. I'll post the more reasonable variations as soon as I dare ;-)
Dave said:
Referring to the E-A as a bjt lookalike of the PL-X with triple darlington outputs does not do either just.
Neither does referring to a folded cascode input stage, you might just as well give reference to the BT line stage input.
Sure there are some simularities to the PL modular gain stages 1 to 4, but i could show you some other input coupled folded cascode schematics (which pre-date the X-amps) that lead to the same assumption, which i'm not.
An inductor between the input stage emitters doesn't make it a SuSy.
I do agree that the E-A is at the same staircase level as the PL's, it's not a hacked together bridged 80s fast high NFB opamp design.
The reason it's somewhat sad that every hick has his attention focussed on NP handouts and this one goes by fairly unnoticed, same goes for Mr Stuart's and Mr Ovidiu's rather jolly poweramp
(hope you don't mind that i used 190MHz Phily BD139/140-16 for the drivers, shame i can't put my 90MHz NEC A1141/C2681 in the E-A output stages)
This is starting to go a bit off topic I think, wouldn't want to draw attention away from this great design this thread is intended to show off.
However Jacco I think you are missing the point. I far as I know the BT (BlowTorch) line stage does not have balanced inputs or outputs (hence in not a fully differential amplifer) nor does it have (again as far as I know) global feedback loops. Hence it obviously can not given reference to here regardless of it having or not having a folded cascode input stage.
However Jacco I think you are missing the point. I far as I know the BT (BlowTorch) line stage does not have balanced inputs or outputs (hence in not a fully differential amplifer) nor does it have (again as far as I know) global feedback loops. Hence it obviously can not given reference to here regardless of it having or not having a folded cascode input stage.
Hi,
Some possible refinements to the ExtremA's output stage (simulations only, please take them as general ideas, not as well-proven practical tweaks for that particular amp! They aren't, nor do they seem to be neccesary for it anyway).
Starting with the minor things, I lowered the emitter resistors to 0.1R, so the output impedance is 2.5 times lower, 0.05 ohms vs. 0.125 ohms.This will also make the "gm halving" effects more benign when we enter class B. To compensate for the lower voltage drop accross them, I level shifted the sense voltages with a Schottky diode drop. This didn't destabilize the loop thermally, bias drops almost identically with temperature as in the original OPS, dominated by the gain transistor's Vbe drop. The current sources allow fine-trim of bias/offset but have to be of very good quality, as the diode drop is outside the loop. Alternatively one could offset the emitter potentials of the gain transistors but I found that a bit too complex and little awkward.
Second, I added AF shunts accross the base stoppers, there is no need for a base resistance at frequencies which are way below the Hfe roll-off (where impedance transform of the output trannies starts) and below the resonances of any parasitic tank circuits. This improved linearity to some amount, maybe not worth the effort/risk.
But the most effective change was the addition of an error cross-feed network which I added after looking at the indiviual error terms and finding that they partly cancel when summed together. While it also improved linearity in class A a little bit, it had a (totally unexpected) dramatical influence of the distortion when A/B transitions occur. Improvements are on the order of 30dB (factor 30)! The value of 1.1k was empirically found and is a bit critical (better to be on the high side of it than on the low side). I couldn't find any tendency of instability due to that addition (not having simmed all possible load conditions, of course), the step response is practically identical, indicating similar stability margins. At a first glance it deemed necessary to double up the dominant pole compensation cap because the driving point AC resistance is now about half its original value, but investigation with the step response didn't confirm this. From a small signal standpoint there are two different feedback signals fed into the bases which might account for that somwhat unintuitive observation... there might be other reasons that I overlooked in my crude "student-level design analysis" (and as a proof for this, I didn't manage to get the loop gain probing to work, with the cross coupling).
With values adjusted so that with 4 ohms load the amp is just on the edge to class B mode (~200mA margin) I get the following THD20 results (half-bridge):
38ppm (old), 8ppm (new)
At 1/10th power (well inside class A margins):
3.5ppm (old), 0.5ppm (new)
Thus, improvements somewhere below a factor of ten, nothing to write home about (and considering it's all virtual anyway, as of yet).
Now, driving 3 ohms (thus about 1/4 period in class B mode), THD20:
0.3% (old), 0.013% (new)
2 ohms (running 11A peak, close to absolute current limit):
0.8% (old), 0.02% (new). This is a factor of 40 improvement. Using only this mod on the original OS also gives factors in the range of 10...20.
In the attached response I plotted output voltage slope (1st derivate), one can see that the crossover distortion is almost completely absent (red--old compared to blue--new), so the resulting distortion must come from gradual changes in the waveform, sporting only low order components and the .four output confirmed this, the original is dominated by H7 (probably the most offensive harmonic), my variation shows mainly H3. Further, in the graph I also plotted the currents through the upper emitter resistor, my version (green) shows a more benign waveform with less HF content/ringing (purple--old).
One last change is to tap off the feedback for the global loop not directly at the output but from a second averaging point between the output's emitters. This will place the ohmic output resistance outside of the loop, but as I'm not a friend of those extreme damping factors (usually rendered useless with cable and contact resistances), I feel a Zout of 0.1 ohms (bridged mode) should be low enough for the majority of speakers. The benefit of this is that the output current completely "disappears" from the main FB loop, it sees only that little error term (some ppm while in class A) which it may correct. So, the global loop merely sets the output voltage while the current is handled only by the local loop in the output stage. The individual tasks of the loops don't get mixed up, the global loop handles only the voltage, sets the gain and linearizes the front-end with the help of feedback, while the OS takes care of the current portion of the cake alone.
And the Zout is purely passive, so to say. This is not the same as to add some resistance to an output whose impedance is controlled/lowered by global feedback, the current term still enters the main loop, albeit to a lesser extent.
This surely is a matter of personal taste but somehow I feel that it might be of benefit if a global loop is freed from having to handle the (quasi non-correlated) current terms, since to apply a voltage at the output and then "fire-and-forget" about the current draw is IMHO what makes an ideal amplifier in the first place, after all (like non-FB designs do, at the cost of increased distortion). To my knowledge this also is an old and known concept, probably implemented in some comercial amps. With a very good OS one could even abandon overall feedback and could close the front-end loop just behind the VAS.
Regards, Klaus
Some possible refinements to the ExtremA's output stage (simulations only, please take them as general ideas, not as well-proven practical tweaks for that particular amp! They aren't, nor do they seem to be neccesary for it anyway).
Starting with the minor things, I lowered the emitter resistors to 0.1R, so the output impedance is 2.5 times lower, 0.05 ohms vs. 0.125 ohms.This will also make the "gm halving" effects more benign when we enter class B. To compensate for the lower voltage drop accross them, I level shifted the sense voltages with a Schottky diode drop. This didn't destabilize the loop thermally, bias drops almost identically with temperature as in the original OPS, dominated by the gain transistor's Vbe drop. The current sources allow fine-trim of bias/offset but have to be of very good quality, as the diode drop is outside the loop. Alternatively one could offset the emitter potentials of the gain transistors but I found that a bit too complex and little awkward.
Second, I added AF shunts accross the base stoppers, there is no need for a base resistance at frequencies which are way below the Hfe roll-off (where impedance transform of the output trannies starts) and below the resonances of any parasitic tank circuits. This improved linearity to some amount, maybe not worth the effort/risk.
But the most effective change was the addition of an error cross-feed network which I added after looking at the indiviual error terms and finding that they partly cancel when summed together. While it also improved linearity in class A a little bit, it had a (totally unexpected) dramatical influence of the distortion when A/B transitions occur. Improvements are on the order of 30dB (factor 30)! The value of 1.1k was empirically found and is a bit critical (better to be on the high side of it than on the low side). I couldn't find any tendency of instability due to that addition (not having simmed all possible load conditions, of course), the step response is practically identical, indicating similar stability margins. At a first glance it deemed necessary to double up the dominant pole compensation cap because the driving point AC resistance is now about half its original value, but investigation with the step response didn't confirm this. From a small signal standpoint there are two different feedback signals fed into the bases which might account for that somwhat unintuitive observation... there might be other reasons that I overlooked in my crude "student-level design analysis" (and as a proof for this, I didn't manage to get the loop gain probing to work, with the cross coupling).
With values adjusted so that with 4 ohms load the amp is just on the edge to class B mode (~200mA margin) I get the following THD20 results (half-bridge):
38ppm (old), 8ppm (new)
At 1/10th power (well inside class A margins):
3.5ppm (old), 0.5ppm (new)
Thus, improvements somewhere below a factor of ten, nothing to write home about (and considering it's all virtual anyway, as of yet).
Now, driving 3 ohms (thus about 1/4 period in class B mode), THD20:
0.3% (old), 0.013% (new)
2 ohms (running 11A peak, close to absolute current limit):
0.8% (old), 0.02% (new). This is a factor of 40 improvement. Using only this mod on the original OS also gives factors in the range of 10...20.
In the attached response I plotted output voltage slope (1st derivate), one can see that the crossover distortion is almost completely absent (red--old compared to blue--new), so the resulting distortion must come from gradual changes in the waveform, sporting only low order components and the .four output confirmed this, the original is dominated by H7 (probably the most offensive harmonic), my variation shows mainly H3. Further, in the graph I also plotted the currents through the upper emitter resistor, my version (green) shows a more benign waveform with less HF content/ringing (purple--old).
One last change is to tap off the feedback for the global loop not directly at the output but from a second averaging point between the output's emitters. This will place the ohmic output resistance outside of the loop, but as I'm not a friend of those extreme damping factors (usually rendered useless with cable and contact resistances), I feel a Zout of 0.1 ohms (bridged mode) should be low enough for the majority of speakers. The benefit of this is that the output current completely "disappears" from the main FB loop, it sees only that little error term (some ppm while in class A) which it may correct. So, the global loop merely sets the output voltage while the current is handled only by the local loop in the output stage. The individual tasks of the loops don't get mixed up, the global loop handles only the voltage, sets the gain and linearizes the front-end with the help of feedback, while the OS takes care of the current portion of the cake alone.
And the Zout is purely passive, so to say. This is not the same as to add some resistance to an output whose impedance is controlled/lowered by global feedback, the current term still enters the main loop, albeit to a lesser extent.
This surely is a matter of personal taste but somehow I feel that it might be of benefit if a global loop is freed from having to handle the (quasi non-correlated) current terms, since to apply a voltage at the output and then "fire-and-forget" about the current draw is IMHO what makes an ideal amplifier in the first place, after all (like non-FB designs do, at the cost of increased distortion). To my knowledge this also is an old and known concept, probably implemented in some comercial amps. With a very good OS one could even abandon overall feedback and could close the front-end loop just behind the VAS.
Regards, Klaus
Attachments
I'm almost done with populating the PCB's. I choosed to use standard metalfilms for most positions but will try to source better varieties for the most critical positions.
A couple of resistors and the input BJT's are what is left to mount on the board and here comes the question;
When matching the inputs 550C input transistors what procedure should I use?
Is it enough to use a DMM that measures Hfe or do I need to set up a circuit that bias the BJT's to similar current as they will see in the ExtremA circuit?
If the latter, one point is enough or should I measure Hfe att several currents/biasing points?
Anyone feel free to answer, thanks in advance!
/Peter
A couple of resistors and the input BJT's are what is left to mount on the board and here comes the question;
When matching the inputs 550C input transistors what procedure should I use?
Is it enough to use a DMM that measures Hfe or do I need to set up a circuit that bias the BJT's to similar current as they will see in the ExtremA circuit?
If the latter, one point is enough or should I measure Hfe att several currents/biasing points?
Anyone feel free to answer, thanks in advance!
/Peter
Peter,
A simple DMM with Hfe function will do. Matching within 5% is fine, it isn't critical anyway, the biggest effect you'll see from a mismatch is a higher DC offset voltage which you'll have to use the trimmer for. There's also a slight increase in 2nd harmonics, but that's covered completely by the noise floor.
Best regards,
Sander Sassen
http://www.hardwareanalysis.com
A simple DMM with Hfe function will do. Matching within 5% is fine, it isn't critical anyway, the biggest effect you'll see from a mismatch is a higher DC offset voltage which you'll have to use the trimmer for. There's also a slight increase in 2nd harmonics, but that's covered completely by the noise floor.
Best regards,
Sander Sassen
http://www.hardwareanalysis.com
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.