Hi Charles,
Another annoying fact is the ML figures their schematics are a deep art and hide them from view by not releasing any, even for service. What I've seen lately is a love affair with uP's in an amplifier, and reasonably normal circuitry. I just had a run in with them last year.
Is the circle above the "A" label in the Sansui diagram a CCS? This would make sense if it were.
Hi Andy, Edmond,
I must agree with Charles and others. I have always had better performance with less complexity using BJTs over Mosfets. I have been working with output stages using no feedback and the same voltage amplification stage(s). The Mosfet types always need a lot of corrective feedback and still fail to sound as good. I'm keeping an open mind, but experience is telling me that BJT output stages are more linear. I've even converted a few Mosfet designs to use BJT's and ended up with better sound.
I received Doug Self's book not long ago and have read the first chapter only. In general, my experience agrees with what he has written so far.
-Chris
Ah, okay. Thanks for clearing that up. It was / is still Nelson's IP, wasn't it?
That is really my main point. Most things I see as new are simply rehashes of earlier ideas. No harm in giving credit where credit is due. I also think that learning about the history of audio electronics is valuable.
Another annoying fact is the ML figures their schematics are a deep art and hide them from view by not releasing any, even for service. What I've seen lately is a love affair with uP's in an amplifier, and reasonably normal circuitry. I just had a run in with them last year.
Is the circle above the "A" label in the Sansui diagram a CCS? This would make sense if it were.
Hi Andy, Edmond,
I must agree with Charles and others. I have always had better performance with less complexity using BJTs over Mosfets. I have been working with output stages using no feedback and the same voltage amplification stage(s). The Mosfet types always need a lot of corrective feedback and still fail to sound as good. I'm keeping an open mind, but experience is telling me that BJT output stages are more linear. I've even converted a few Mosfet designs to use BJT's and ended up with better sound.
I received Doug Self's book not long ago and have read the first chapter only. In general, my experience agrees with what he has written so far.
-Chris
anatech said:
Hi Chris,
Well, I wasn't actually trying to make the argument that MOSFETs are more linear than BJTs - they aren't. However, if you look at Self's analysis, you'd probably come to the conclusion that not only are they less linear than BJTs in this application (true), but that they are truly awful and horribly nonlinear (false).
One might reach that latter conclusion based on his graphs of the incremental gains. But the point that Edmond and I are trying to make is that his data indicating truly awful behavior are based on a faulty SPICE analysis. When more accurate SPICE models are used (that take into account weak inversion), the results are much different.
Hi Andy,
I've only completed the first chapter of his book. I also would give these parts the benefit of the doubt on that point until I have read that section and experimented more. Who knows what I could be doing wrong here? I'm sure I am not alone in misusing the device.
-Chris
I'll take your word for that.
I've only completed the first chapter of his book. I also would give these parts the benefit of the doubt on that point until I have read that section and experimented more. Who knows what I could be doing wrong here? I'm sure I am not alone in misusing the device.
-Chris
anatech said:
Oh, I doubt if you're doing anything wrong at all. I know you have years of experience. All I'm saying is that Self's gloom and doom about MOSFETs is based to a large extent on the old style SPICE MOSFET models that don't reflect reality in the region of operation where they are actually used in this circuit configuration.
Although MOSFETs are somewhat less linear than BJTs, it's easier to implement error correction with them. With error correction and BJTs, you run into the problem of very low VCE in the error correction devices themselves. To get around that, you'll need to do something like what Glen is doing. With the FETs though, the large Vgs compared to BJT Vbe makes error correction simpler. And an error-corrected FET output stage is much more linear than a non-error-corrected BJT output stage.
andy_c said:
Hi Andy,
Thanks for elaborating on this topic. Your graph (post#2458) reveals exactly what I mean. BTW, my home-brew models are not only based on specs of the data sheets, but also on real measurements of the weak inversion.
Mr C. Hanson, I do hope this will change your opinion against "super-fancy models"
Cheers, Edmond.
Re: Re: Bipolar - MOS crossover issues
This is not correct. The bulk of these inventions seek to bend
the load line of a push-pull amp to avoid cutoff, and assuming
that they start with a healthy bias to begin with, they smooth
the load line as compared with Class AB, lowering both the
measured distortion and the proportion of higher harmonics.
Your memory has failed you. In my circuit both sides remain in
conduction. The circuit to which you refer is probably Barry
Thornton's as used by SAE.
I had another design which I did not release which also fits that
description but where the switched diodes appear at the input
of the driver stage. Later I saw this in a Technics design instead
of the expected "Super-A" circuit of Sano, Hirosha, and Hirota.
Just for the record Charles, the dynamic bias circuit was
superceded at Threshold for strictly marketing reasons, not for
technical or subjective reasons. When the herd arrives to feed
on a good marketing story, it's time for me to move on.
As a concept, dynamic biasing is still viable, and I continue to
insist on it's similarity to a number of the EC circuits I have seen.
The notion of designating a point in the bias circuit where the
action of the bias circuit works to hold the output voltage to the
value of that point is a pretty general idea.
Charles Hansen said:
This is not correct. The bulk of these inventions seek to bend
the load line of a push-pull amp to avoid cutoff, and assuming
that they start with a healthy bias to begin with, they smooth
the load line as compared with Class AB, lowering both the
measured distortion and the proportion of higher harmonics.
Tim__x said:
Your memory has failed you. In my circuit both sides remain in
conduction. The circuit to which you refer is probably Barry
Thornton's as used by SAE.
I had another design which I did not release which also fits that
description but where the switched diodes appear at the input
of the driver stage. Later I saw this in a Technics design instead
of the expected "Super-A" circuit of Sano, Hirosha, and Hirota.
Just for the record Charles, the dynamic bias circuit was
superceded at Threshold for strictly marketing reasons, not for
technical or subjective reasons. When the herd arrives to feed
on a good marketing story, it's time for me to move on.
As a concept, dynamic biasing is still viable, and I continue to
insist on it's similarity to a number of the EC circuits I have seen.
The notion of designating a point in the bias circuit where the
action of the bias circuit works to hold the output voltage to the
value of that point is a pretty general idea.
anatech said:
Hi Chris,
You're right, MOSFETs, need more corrective feedback. My early designs were also equipped with BJTs, but because of my fears of the "hole storage effect" and second break down issues, I switched to MOSFETs. Happily, my ears are too old to hear the difference.
Cheers, Edmond.
PS, enjoy reading D. Self's book, it's really good, but skip chapter 11.
Hi Edmond,
I have also gone back to some simpler transistor circuits and trying variations on those. It's hard to argue when you have a three transistor circuit with +20 dB gain and less than 0.002% THD at 4 Vrms (1 KHz). I haven't gotten too much further than that so far. I think we are too quick to use op amps these days.
-Chris
Hmmm, too much good stuff in between to skip right to it. I will be doing a lot of reading though.
I have also gone back to some simpler transistor circuits and trying variations on those. It's hard to argue when you have a three transistor circuit with +20 dB gain and less than 0.002% THD at 4 Vrms (1 KHz). I haven't gotten too much further than that so far. I think we are too quick to use op amps these days.
-Chris
Originally posted by anatech Hi Edmond,
Hmmm, too much good stuff in between to skip right to it. I will be doing a lot of reading though.
I have also gone back to some simpler transistor circuits and trying variations on those. It's hard to argue when you have a three transistor circuit with +20 dB gain and less than 0.002% THD at 4 Vrms (1 KHz). I haven't gotten too much further than that so far. I think we are too quick to use op amps these days.
-Chris
Hi Chris,
Huh.. only three trannies? So you have beaten Darkfenriz (Adam) with his exiting ten trannies design.
Cheers, Edmond.
Hi Edmond,
I can't take any credit for that design. I'm running part of a tone control circuit from an old Marantz 1152DC integrated amplifier. It was destroyed and the board contained a similar circuit to what I wanted to play with.
Why waste time designing PC boards when one is finished and debugged for you already?
-Chris
I can't take any credit for that design. I'm running part of a tone control circuit from an old Marantz 1152DC integrated amplifier. It was destroyed and the board contained a similar circuit to what I wanted to play with.
Why waste time designing PC boards when one is finished and debugged for you already?
-Chris
You spelled my name right, no apologies necessary
BTW, there is a pair of diodes - I also used a trick that I got from
Barney Oliver of using power diodes to shunt across emitter
resistance so as to limit the maximum voltage at high current.
It made the job easier for my (admittedly crude) bias circuit.
BTW, there is a pair of diodes - I also used a trick that I got from
Barney Oliver of using power diodes to shunt across emitter
resistance so as to limit the maximum voltage at high current.
It made the job easier for my (admittedly crude) bias circuit.
spreading A/B transitions
May I ask the experts what you think of speading the individual A/B transitions when using multiple paralleled output devices? This might smooth out the overall distortion signature when each output transitions in slightly different regions -- this, as I see it, partly happens anyway with paralleled outputs with non-perfectly matched devices and non-uniform device temperatures throughout, so why not use it deliberately?
I've actually done something similar with paralleled op-amps biased individually to have them operate in class A for low levels and have them invert their output current one by one, not all at the same point, with increasing output current:
http://www.diyaudio.com/forums/showthread.php?postid=1258047#post1258047
Regards, Klaus
May I ask the experts what you think of speading the individual A/B transitions when using multiple paralleled output devices? This might smooth out the overall distortion signature when each output transitions in slightly different regions -- this, as I see it, partly happens anyway with paralleled outputs with non-perfectly matched devices and non-uniform device temperatures throughout, so why not use it deliberately?
I've actually done something similar with paralleled op-amps biased individually to have them operate in class A for low levels and have them invert their output current one by one, not all at the same point, with increasing output current:
http://www.diyaudio.com/forums/showthread.php?postid=1258047#post1258047
Regards, Klaus
Hi Klaus,
In my experience, this increases distortion. I've been matching outputs (and other parts) for years. Some designs are more sensitive than others. Of course, this may just mean that some designs commit larger evils that swamp out mismatch effects.
On Semi figures that your output stage THD could be reduced as much as to 10% it's previous level (before feedback) simply by matching the output transistors. They have "characterized" their new series of audio output transistors for distortion performance. Their figures are for the no feedback case to remove circuit variables.
So finally, after years of defending my position, On Semi came out and verified my practices. It is a pain though, although the new parts are so much better in device spread than they ever were before.
-Chris
In my experience, this increases distortion. I've been matching outputs (and other parts) for years. Some designs are more sensitive than others. Of course, this may just mean that some designs commit larger evils that swamp out mismatch effects.
On Semi figures that your output stage THD could be reduced as much as to 10% it's previous level (before feedback) simply by matching the output transistors. They have "characterized" their new series of audio output transistors for distortion performance. Their figures are for the no feedback case to remove circuit variables.
So finally, after years of defending my position, On Semi came out and verified my practices.
It is a pain though, although the new parts are so much better in device spread than they ever were before.-Chris
andy_c said:
Well, heck, these are *not* trivial differences!
There are two things that I'm confused about. Edmond or Andy, please feel free to edify me in these regards:
a) My understanding was that if we ignore the very low currents (that is presumably not of interest in the output stage of an audio power amp) that the main deviation from the square law for power MOSFET's was due to heating effects increasing the channel resistance.
In other words, a BJT will follow the exponential law for something like 10 decades (!), while a MOSFET will follow the square law for only 2 or 3 decades. But it seems to me that an audio power amp will largely be operating in those 2 or 3 decades where the MOSFET follows the square law very closely.
But you guys keep talking about "weak inversion", which I've never heard of. Please help me out and explain what "weak inversion" is, what it does, and if the channel heating effects are modeled in *any* level of the SPICE models.
b) If you look at the on-resistance curves for FET switches (which comprise a paralleled complementary pair of MOSFET's), the graphs look almost exactly like Self's graphs. These curves are presumably made by actual measurement. So is it just coincidence that the models Self is using (admittedly in a slightly different circuit configuration) reflect the real world of the switches?
Thanks in advance.
Re: Re: Re: Re: Re: Re: Bipolar - MOS crossover issues
There are 4 paralleled devices in each leg of an "H" bridge, for a total of 16 outputs per channel. Each device is running at ~120 mA. Coincidentally, this was the same idle current per device that I was using in the design with the lateral MOSFET's. But there are only half as many output devices, so half the total idle current.
But the bipolar design is rated at 300 watts, while the MOSFET design was rated at 200 watts. I made a prototype of a 300 watt lateral MOSFET design and it used 12 paralleled devices in each leg. So for an apples-to-apples comparison, the BJT design has only 1/3 the idle current. But it measures better, sounds better, is cheaper....oh, you know the rest.
Wull, yahbut....why?
The distortion is already about 2 orders of magnitude lower than a transducer. And since both the amp and the transducer are zero-feedback mechanisms, the audibility of the distortion between the two devices is going to be roughly comparable.
On the other hand, error correction is a feedback scheme, and I find that the audible distortions created via a feedback scheme have a noticeable sonic signature. It's kind of like the distortions created by a digital recording system are very different from the distortions created by an analog recording system. Each system has a different sonic signature. And as any analog lover will tell you, the analog distortions are much more benign in nature. So I prefer to have a zero-feedback amp with 0.01% distortion than a feedback amp with 0.001% distortion. YMMV.
Bob Cordell said:
There are 4 paralleled devices in each leg of an "H" bridge, for a total of 16 outputs per channel. Each device is running at ~120 mA. Coincidentally, this was the same idle current per device that I was using in the design with the lateral MOSFET's. But there are only half as many output devices, so half the total idle current.
But the bipolar design is rated at 300 watts, while the MOSFET design was rated at 200 watts. I made a prototype of a 300 watt lateral MOSFET design and it used 12 paralleled devices in each leg. So for an apples-to-apples comparison, the BJT design has only 1/3 the idle current. But it measures better, sounds better, is cheaper....oh, you know the rest.
Bob Cordell said:You could always use error correction...
Wull, yahbut....why?
The distortion is already about 2 orders of magnitude lower than a transducer. And since both the amp and the transducer are zero-feedback mechanisms, the audibility of the distortion between the two devices is going to be roughly comparable.
On the other hand, error correction is a feedback scheme, and I find that the audible distortions created via a feedback scheme have a noticeable sonic signature. It's kind of like the distortions created by a digital recording system are very different from the distortions created by an analog recording system. Each system has a different sonic signature. And as any analog lover will tell you, the analog distortions are much more benign in nature. So I prefer to have a zero-feedback amp with 0.01% distortion than a feedback amp with 0.001% distortion. YMMV.
anatech said:
Yes, unlike (say) Nelson Pass, the Levinson designs were pretty much straightforward examples from textbooks or app notes, but usually implemented with high quality parts and good strong power supplies. The only unusual thing was the "Adaptive Bias", but that was copied straight from the AES article by the Sansui engineer.
Nelson, on the other hand, is *constantly* inventing new circuits. Ya gotta hand it to him for that -- no flies on that guy!
anatech said:
Yes, it is a CCS. One obvious variation would be to replace the CCS at "A" with a transistor complementary to "B" to make a fully complementary design that is driven at two points (separated by a bias voltage). That is what Levinson did.
Re: Re: Re: Bipolar - MOS crossover issues
I didn't say it wasn't viable. I just said that it causes more problems than it solves. (Of course this is my opinion and YMMV.)
And even if you moved on from a dynamic bias circuit for marketing reasons, I know you well enough to state with confidence that you wouldn't have moved on if you hadn't come up with something that was equal (if not better). In short, you wouldn't be satisfied going *backwards* from a performance standpoint just because the "herd" was copying you.
Nelson Pass said:
I didn't say it wasn't viable. I just said that it causes more problems than it solves. (Of course this is my opinion and YMMV.)
And even if you moved on from a dynamic bias circuit for marketing reasons, I know you well enough to state with confidence that you wouldn't have moved on if you hadn't come up with something that was equal (if not better). In short, you wouldn't be satisfied going *backwards* from a performance standpoint just because the "herd" was copying you.
Charles Hansen said:
Yes! That very thing got me also. It wasn't until studying Edmond's models using BSIM3 and how they differed from the traditional SPICE MOSFET models (levels 1-3) that I started to understand this better myself.
Well, since we're talking about crossover distortion here, that's a low current thing. So ignoring what's going on at low currents in the models turns out to be a trap. It's a trap that Self fell into, and I did too, until I later understood better what was going on. More in a moment.
Self-heating is something I haven't studied in any detail at all, so I can't provide any meaningful information about that. But the non-trivial differences in the graph are at low currents, in the transition region where one device is turning off as the other draws more current.
Yes, but Self's MOSFET graphs of the derivative of output voltage with respect to input voltage for a source follower show the really lousy stuff going on at low currents. It's important that the models accurately reflect reality at low currents so we don't get fooled.
Let's set aside this self-heating thing for another discussion later and concentrate on weak inversion. One simple way to look at this is by the model equations for the simplest level 1 SPICE model. This model gives the drain current as:
Id = 0 for Vgs < Vto, and
Id = K(Vgs - Vto)2 for Vgs >= Vto
So for the level 1 SPICE MOSFET model, the drain current for Vgs less than the threshold voltage is modeled as zero. But that's wrong! The drain current for a real n-channel MOSFET starts to climb slowly for Vgs > 0, and grows in an exponential way until Vgs becomes somewhat less than the threshold voltage Vto. This is the weak inversion region, also known as sub-threshold conduction. IOW, the actual device is drawing drain current at a voltage less than Vto when the level 1 model claims that current is dead-on zero. Now, when Vgs is somewhat larger than the threshold voltage, the drain current vs. Vgs looks like a square law. In between, there is a transition region.
Now, there is a problem in the simulation by representing the drain current vs. Vgs as two discrete expressions as above. For the first expression above (Vgs < Vto), the drain current is a constant, zero. So all subsequent derivatives must be zero. For the second expression (Vgs >= Vto), the first derivative with respect to Vgs is 2K * (Vgs - Vto). Taking the second derivative, we end up with 2K. So the second derivative is 0 for Vgs < Vto, and steps instantaneously to 2K for Vgs >= Vto.
This behavior of the level 1 model is bad. We would like a single expression for Id vs. Vgs at a fixed Vds that gives us the current for any Vgs we want. That's what BSIM3, BSIM4, EKV and other more complex MOSFET models give us. This "unified expression" property ensures continuity of the Id vs Vgs expression and all its derivatives, eliminating "kinks" when we start taking derivatives (as Self does). So what I'm claiming is that this discontinuity of the second derivative of Id vs. Vgs is what causes the "kinks" (sudden changes in slope) in Self's plots of the derivative of output voltage with respect to input voltage for the source follower.
Hmm, Your question has me confused. The on resistance is for the triode region, which corresponds to clipping. I'm not sure how to respond to that one.
Anyway, for more info on the device physics of weak inversion (AKA sub-threshold conduction) see the link I posted earlier. This is a PowerPoint document, so you'll need PowerPoint or the MS PowerPoint viewer installed.
andy_c said:
Thanks for your help with this, Andy -- you rock!
I'm still a bit skeptical, though. I know that at sub-threshold voltages that the device doesn't follow the square law. But let's look at two key slides in the PowerPoint presentation you linked. If I counted right, they are #23 and #24, that show Id versus Vgs on a linear scale first and then a log scale.
Based on the red X's, we can assume that these graphs are supposed to be of the same part. There is a mistake however because in the first graph Vgs = 3 V gives Id = 90 A, while the second graph Vgs = 3 V gives Id = 10^-4 A. Big discrepancy!
But let's ignore that for now. The main thing is the second graph. Although the numbers are hard to read, it appears that each division on the vertical scale corresponds to one order of magnitude. So let's normalize the top of the second graph to 10 A, which would be about right for a MOSFET used for an audio amp (give or take an order of magnitude!)
Then the cluster of the three red X's around 0.5 Vgs are at 100 uA or so. Clearly in this range the behavior of the device is exponential. And I would assert that in the range of 100 uA that the crossover distortion is not going to be affected all that much by this deviation from the square law.
So the question becomes *when* does the transition from exponential to square law *really* occur? And how does that affect the crossover distortion as the device turns on?
In our re-normalized log graph, the red X at Vgs = 1 V is at 100 mA. This is clearly in the area of interest for crossover distortion. But is that really where the device changes from square law to exponential? Just by eyeball, I wouldn't say so. By my eye, I would say it would be another decade down, say at Id = 10 mA or so.
Now it may be that this is enough to ameliorate the sharp notches that Self was showing. Or maybe not. I think we are on the fringes of the transition zone between exponential behavior and square law behavior, to the degree where the accuracy of the model is going to be *critical* in predicting what the output stage is going to do. And I guess this is where you would want to compare the model against the *real* part, either by comparing the actual characteristic curves or by comparing the transfer function of the output stage.
So I guess you are saying that the old models were *way* off base in this regard. How good do you think the new models are in this regard?
And I must admit that I am surprised that the change to exponential behavior is that strong so close to Vgs. I was under the impression that it didn't take effect for another couple of decades below that. Apparently I was mistaken. (For the first time, right, Andy?)