Bob Cordell said:
Since that time I have also found that simpler is not always better. In our latest amp (the MX-R), the added transistors used to stabilize the operating point of the input FET (which provides all of the gain) improve both the sound and the measurements.
It's kind of interesting, really. The FET operates as a voltage-to-current converter, and then it has a load resistor that acts as a current-to-voltage converter. The whole system is so darned linear that it's kind of scary. If we made a low-powered class-A amp with this circuit (and avoided any problems with crossover distortion in the output stage), I think we could get the open-loop distortion down to around 0.001% or 0.002% range.
Bob Cordell said:
Well, the two problems with vertical MOSFET's are a zero tempco point at one or two orders of magnitude higher than you would like, plus the non-linear Cgd. In comparison, the only problem with the lateral is the lack of transconductance. But if you parallel a bunch of lateral devices, you can get as much transconductance as you want. The resultant composite device will be better than an equivalent vertical part in four ways:
a) The non-linearity in Cgd will be about an order of magnitude better.
b) The Cgd will be less than half.
c) The zero tempco point remains at a nice useful value, radically simplifying the bias circuitry.
d) The reliability of the design goes way up, as the thermal resistance from the composite (paralleled) transistors to the heatsink is dramatically reduced.
The only disadvantage is cost. But since we are talking about high-performance audio instead of mass-market audio, I see no reason to use vertical parts if you are going to use MOSFET's.
The one nice thing about MOSFET's versus bipolars is their relative immunity to instability when faced with a capacitive load. But in the end, I have come to prefer BJT's overall.
As an aside, you should know that there is an inherent defect in the IR P-channel parts. There is something wrong with the way they are made that results in the transconductance varying about 50% over time. The time scale in question is on the order of a few milliseconds, which puts the problem smack-dab in the audio band. So never, never use IR P-channel MOSFET's. If you search the archives, you will find some other posts I have written on this topic.
Charles Hansen said:
Hi Charles,
I'm a bit relieved to hear that the main concerns you have about vertical MOSFETs appear to be Cgd nonlinearity and positive tempco below several amperes.
For the moment, let me stick to comparisons with BJTs, rather than Lateral MOSFETs. The BJTs also have a large and very nonlinear Ccb at low Vce that is analogous to the Cgd of the vertical MOSFET. If you drive them both with a similarly low impedance, as in a Triple, the MOSFET is not likely to be worse than the BJT in this regard (and the problem should not be a big issue for either of them as long as the driver circuitry can support adequate current in both directions). Let me know if you think I am missing something here.
With respect to the tempco, the Laterals are indeed nicer than the verticals, but BJTs are much worse than the Vertical MOSFETs. Indeed, the development of the OnSemi ThermalTrak devices is a very helpful mitigating factor for the BJTs in this regard. If you haven't seen it, take a look at my JAES MOSFET Power Amplifier with Error Correction paper on my web site at www.cordellaudio.com under Published papers. In that paper I provide theory and actual bias stability measurements that show that Vertical MOSFETs are inherently more temperature-stable than conventional bipolars (even though the verticals do indeed have a positive tempco at normal bias current levels). One of the things I show there for both BJT and MOSFET designs is how the bias point changes and "recovers" after a prolonged amount of heat-generating program is removed. The BJTs move much much more than the Vertical MOSFETs, leading to very real concerns about transient under-bias or overbias of the output stage during program dynamics
Thanks to John and Nelson bringing it to my attention (and their mentioning your experience) I am very well-aware of the transconductance anomoly in the IRF P-channel devices. Being the show-me curious kind of guy that I am, I measured the cr@p out of it myself, and it is surely there. I had several communications with engineers at IR about it to no avail - no-one really understood its origin. The relatively long time constant of the phenomenon that places its effect in the audio band is truly curious. I suspect it has something to do with trapped charge and an extremely high resistance discharge path that influences gate inversion. I also brought it to Ed Oxner's attention and he was also unfamiliar with the phenomenon. It appears to me that the higher-frequency (reduced) transconductance is the "true" transconductance, and that the phenomenon acts to inadvertantly enhance the apparent transconductance at lower frequencies IIRC. Anyway, I agree that for best performance the IR P channel devices should be avoided, and I have not heard of the phenomenon occurring in other vendors' devices.
Cheers,
Bob
lumanauw said:
Yes, of course. Mr. Cordell has done this very successfully with his designs. When I tried it in our original design, I did not like the sonic result. Since then I have not used vertical MOSFET's.
Here is another interesting aside and another reason why I don't like vertical devices. Over the years, we have had a small number of unexplained failures in that original amp that was all-MOSFET. The problem turned out to be that at high line voltage, the bias in the output stage would run away.
After studying this for a while, we finally tracked it down. The problem was the N-channel IR MOSFET with the sensing device built-in (IRCP054, I think it was). When you put it on the curve tracer, it looked great up to about 30 Vds. But above that, the lines would curve upwards even though the part was rated at 60 volts. So when the line voltage would increase, the supply voltage would increase and then Ids would increase. This would heat the device excessively and then the bias would increase until it melted.
But here is the kicker. We looked at over two dozen possible replacements from every manufacturer we could find (none of which had the sensing terminal, but that wasn't a big deal because we could always measure the bias by putting an ammeter in place of the rail fuse -- there was no source resistor in this design). EVERY SINGLE ONE had the same problem except for a Samsung part that had been taken over by Fairchild (I think that's who it was).
Of course, THAT part was also being end-of-lifed, so we had to buy a couple of thousand so that we had a lifetime supply. But the point is this. Vertical MOSFET's are designed for SWITCHING NOT AUDIO. They often have all kinds of quirks and non-linearities that can cause trouble when being used in a linear application. So one must be very careful when trying to use them for audio.
The IR P-channel parts are perfect example of this. When I first found the problem with time-varying transconductance in 1992 and brought it to the attention of the IR applications engineers, they said they didn't care because the parts were designed for SWITCHING AND NOT AUDIO. I have a friend who is an applications engineer for Infineon (then Siemens) in their power MOSFET division. He knew exactly what was wrong with the IR parts and explained the mechanism to me. (Unfortunately, I have since forgotten what he told me.) But I never could use the Infineon parts, as they don't make many (any?) P-channel devices and I like complementary circuits.
I just think that your life will be easier if you use parts designed for audio instead of trying to adapt parts designed for switching. (Unless of course, you are designing a switching amplifier!)
Finally, on a completely different topic, the Hi-Fi+ review of our amplifier using the ThermalTrak devices is now posted on our website:
http://ayre.com/PDF/HiFi_Plus_MX-R_Review.pdf
janneman said:
Yes, it starts off as a slab weighing 75 pounds and when it is finished it weighs 22 pounds. So over 2/3 of the metal ends up as chips (recycled, of course!).
It is a very nice way to build things. Very rigid, great thermal dissipation (the entire chassis is the heatsink), you can put mounting points anywhere you want, and there are separate shielded chambers for each PCB and each wire. But the downside is that it is VERY expensive -- about 5x the price of doing it with folded sheet metal.
---To all 3-level EF Output Stage Experts---
For a bridged Class-A amp with a regulated driver power supply and a capacitor filtered lower voltage output power supply, what stages of a 3-level emitter follower output stage using ThermalTrak bipolar output transistors+diodes can be connected to the driver power supply?
Are there distortion or thermal or bias tracking advantages to attaching all 3-levels of the EF output to the lower voltage capacitor filtered output power supply?
Every diode drop wasted in the output swing burns measurable power with Class-A. With modern bipolars, a 2 level EF output can have a combined beta of 20,000. Does the 2,000,000 combined beta in a 3 stage EF output really make a big difference? Lower distortion? Better sonics? Better load tolerance?
For a bridged Class-A amp with a regulated driver power supply and a capacitor filtered lower voltage output power supply, what stages of a 3-level emitter follower output stage using ThermalTrak bipolar output transistors+diodes can be connected to the driver power supply?
Are there distortion or thermal or bias tracking advantages to attaching all 3-levels of the EF output to the lower voltage capacitor filtered output power supply?
Every diode drop wasted in the output swing burns measurable power with Class-A. With modern bipolars, a 2 level EF output can have a combined beta of 20,000. Does the 2,000,000 combined beta in a 3 stage EF output really make a big difference? Lower distortion? Better sonics? Better load tolerance?
AndrewT said:
Hi AndrewT
The simple set of rules are at the conclusion:
Ro=R/2 + 1/gm at crossover
Ro= R far from crossover
the variation of Ro with i between these two values are given in the graph for different values of Rgm.
This gives the reason behind the design rule of critical biasing at gmR=1
The development shows also that what matters is not Io or Re but the product gmR. This clarifies also the comments of Self where he found experimentally that the voltage drop on Re is what matters. Rgm is equivalent to (Io R)/Vt. This means that for a given resistor Re there will be an Io that makes the biasing optimal ( Rgm=1) and i.e. minimal variation of Ro with only a slide bump in Ro for i= about 4Io. Equivalently, for a given Io there will be a Re that makes the biasing optimal.
It shows also that a slide underbias is not bad ( until gmR=0.5)
So a rule should be 0.5<gmR<1 providing thermal stability is well controled.
It raises the question of optimal T tracking of the bias. It seems that gm should stay constant with T and not Io as is usually done.
Then the condition of minimal distortion ( gmR=1) remains while T is changing.
The detailed calculation is given because Oliver's paper is difficult ( IMHO) to read and full of typo's errors again IMHO.
It clarifies also this Gm doubling stuff
JPV
LineSource said:
In my experience, the more stages that are regulated, the better the sound quality. Of course, regulating the output stage is very expensive and creates a lot of heat.
Please keep in mind what happens if you have different power supplies for each output level (pre-driver, driver, output) and a rail fuse blows. Then the preceding stage reverse biases that transistor and it acts like a diode to the rail capacitors. This can create excessive current that can damage the stage before the stage with the blown fuse. (Sorry if this isn't clear -- re-read it a few times until it makes sense.)
LineSource said:
No, but you will avoid the reliability problems noted above. Look at the characteristic curves of a bipolar transistor and you will see that not much changes when the Vce changes unless you get to very low levels.
I haven't looked at Locanthi's paper for a long time. In my opinion, the stepped power supply rails are NOT the important part of his design, and in fact may cause problems. The important part is using three emitter followers instead of two, and running the first two stages in class A by not tying them to the load.
LineSource said:
Yes, yes, and yes.
For example, say that the loudspeaker drops to 1 ohm at a certain frequency and the beta of a double is 10,000 (more realisitic than 20,000 I think, plus easier calculations). This means that the gain stage has a 10 kohm load reflected on it, which is enough to affect its operation substantiall. Adding a third follower reduces the loading on the gain stage by a factor of 100 so that now the loading is on the order of 1 megohm. This improvement is easily measurable in an open loop condition. Again, look at the Stereophile test report of the MX-R amplifiers with a triple "T-circuit". The distortion is only 0.015% at 100 watts into 8 ohms and that is without any feedback.
JPV said:
Sorry, I don't have time to read this paper, so I can't comment on it directly.
However, in Doug Self's books he discusses this subject extensively and I find nothing to disagree with him about. You can learn a lot by downloading one of the free simulators and playing around with different values of emitter resistors and bias currents to see what happens.
Which leads us to another advantage of BJT's over MOSFET's. With BJT's, there is one optimal bias current that is relatively low (usually with about 10 mV to 20 mV across the emitter resistor). But with MOSFET's the distortion goes down as the bias goes up, without limit. This means that a MOSFET amp must run relatively hot with relatively high power consumption to get distortion as low as a BJT design.
Oliver's paper in the early '70's set the criterion for optimum emitter resistance of a given complementary transistor output stage. We usually define it as the voltage drop across each resistor when the unit is idling. I tend to chose 15-22 mV, Chas prefers 10-20 mV. There is NO exact number that does everything right, but being in that range is optimum. Other factors weigh in such as beta nonlinearity and finite drive resistance that change the equation slightly.