Bob Cordell Interview: BJT vs. MOSFET

[PB2]

Hi Pete,

I haven't looked at it real hard, but I'll take a chance here and speculate that, with the right interpretation, it may also be governed by my equation, as long as the right numbers are plugged in.

It looks like the net transconductance of the each output transistor, with respect to changes in bias spread, is, in this case, about equal to the transconductance of the power transistor in combination with its associated RE, just like in the emitter follower cases. I say this because the "gain" in this mode provided by the driver transistor (and ignoring its gm in comparison to 100 ohms), is about unity, since we have 100 ohms in the emitter to ground and 100 ohms in the collector load. This also ignores the input impedance of the output transistor (Beta is assumed to be high).

Cheers,
Bob

Thanks Bob, yes I think that your equation applies, since if we assume that the thermal sensor is slow and as a worst case not changing, then the output transistor Vbe does not change. The .1 ohm emitter resistor then acts in the same way as in the case for EF outputs.

Pete B.

 

[PB2]

Hi Bob,

Thinking more about your equation it seems to me that it would also apply with regard to hot spots on the die and secondary breakdown. Not much we can do about it, and it should be characterized in the SOA curves, I would think. However, doesn't your analysis suggest that higher Hfe output devices would tend to have worse secondary breakdown peformance? I'm thinking of the Sankens that are available in different grades. It's interesting that this is not reflected in the SOA curves - they're probably worst case.

Also, it seems to me that worst case behavior will only be seen starting cold, where a step function in power is applied, holding the case temp constant so that it does not help even out the die temp.

Does this make sense, or am I missing something?

It was on my mind here:
http://www.diyaudio.com/forums/showthread.php?postid=1218339#post1218339

Pete B.

 

[Bob Cordell]

Hi Bob,

Thinking more about your equation it seems to me that it would also apply with regard to hot spots on the die and secondary breakdown. Not much we can do about it, and it should be characterized in the SOA curves, I would think. However, doesn't your analysis suggest that higher Hfe output devices would tend to have worse secondary breakdown peformance? I'm thinking of the Sankens that are available in different grades. It's interesting that this is not reflected in the SOA curves - they're probably worst case.

Also, it seems to me that worst case behavior will only be seen starting cold, where a step function in power is applied, holding the case temp constant so that it does not help even out the die temp.

Does this make sense, or am I missing something?

Pete B.

Hi Pete,

These are good questions, but I'm not sure I have good answers without spending some time thinking about it in a quiet room, especially the one regarding SOA as a function of Beta. I suppose that, to the extent that high Beta increases effective device gm by lessening the effect of base spreading resistance, maybe so.

It's hard to say when worst case will rear its ugly head in regard to thermal stability in an amplifier, if one is looking at the amount of thermal positive feedback. To first order, it will be whatever set of conditions increases the product of all of those terms in the formula, gm being an important one, but not the only one. Anything that increases the idle bias and output stage power dissipation might get involved, but if the slower bias sensing loop has time to get involved, and is a bit overcompensated, it could actually make things go the other way in some cases.

In some cases, the scenario of starting cold could indeed be a bad case, however. I had a case of that actually happen to me, my own fault, due to poor physical design. I had a situation where the circuit board was vertical, and one of the driver transistors was below the bias spreader transistor that was not mounted on the heat sink (there were two Vbe multiplier transistors - the other one was mounted on the heat sink). At thermal equilibrium, when I set the bias, I did not realize that I was "depending" on the fact that the bias spreader Vbe multiplier transistor had been warmed by the heat given off by the driver transistor. Subsequently, on a truly cold start, I discovered that the amplifier started out way over-biased when first cold-started.

Cheers,
Bob

 

[PB2]

Hi Pete,

These are good questions, but I'm not sure I have good answers without spending some time thinking about it in a quiet room, especially the one regarding SOA as a function of Beta. I suppose that, to the extent that high Beta increases effective device gm by lessening the effect of base spreading resistance, maybe so.

It's hard to say when worst case will rear its ugly head in regard to thermal stability in an amplifier, if one is looking at the amount of thermal positive feedback. To first order, it will be whatever set of conditions increases the product of all of those terms in the formula, gm being an important one, but not the only one. Anything that increases the idle bias and output stage power dissipation might get involved, but if the slower bias sensing loop has time to get involved, and is a bit overcompensated, it could actually make things go the other way in some cases.

In some cases, the scenario of starting cold could indeed be a bad case, however. I had a case of that actually happen to me, my own fault, due to poor physical design. I had a situation where the circuit board was vertical, and one of the driver transistors was below the bias spreader transistor that was not mounted on the heat sink (there were two Vbe multiplier transistors - the other one was mounted on the heat sink). At thermal equilibrium, when I set the bias, I did not realize that I was "depending" on the fact that the bias spreader Vbe multiplier transistor had been warmed by the heat given off by the driver transistor. Subsequently, on a truly cold start, I discovered that the amplifier started out way over-biased when first cold-started.

Cheers,
Bob

Thanks Bob, I'll continue to ponder this myself.

I have a feeling a step function is the worst case, also for testing the heat sink thermal loop in a complete amp.

Pete B.

 

[PB2]

Hi John,

If you had used a simulator (as I did), you had only wasted one hour.

Cheers, Edmond.

You're joking? Or?

Did you actually try idle current compensation based on sensing the current in Re? Seems the heat sink is a measure of average power, so if we sensed V and I with the right time constants we ought to be able to duplicate a thermal feedback loop.

Pete B.

 

[Edmond Stuart]

You're joking? Or?

Did you actually try idle current compensation based on sensing the current in Re? Seems the heat sink is a measure of average power, so if we sensed V and I with the right time constants we ought to be able to duplicate a thermal feedback loop.

Pete B.

Hi Pete,

Am I Joking? Absolutely not, I was at most a little bit of sarcastic. As a matter of fact, I spiced (years ago) the LT1166 in conjunction with a pair of vertical MOSFETs and the results were disappointing in terms of THD. Spicing it saved me a lot of time!
Did I sensed the current in Re? Yes I did, but I kept the unity gain frequency of the bias feedback loop below AF, somewhere between 1 and 10Hz, to avoid unwanted interactions that increase the distortion. See my article "Autobias for Mosfet Audio Output Stages", EW December 2003, pp. 17...20.

Cheers, Edmond.

 

[Charles Hansen]

For convenience I'll be using a system with the following constants Tr=0.5 Tcvbe=-0.0022 Gm=10 Vce=45.4545, this gives us a beta of 0.5. When the ambient temperature goes up one degree the Pd goes up by 1 watt and the temperature rises by an additional 0.5 degree, this is important, the temperature does not rise 1.5 degrees, it rises 0.5 degrees relative to ambient which happens to have risen by 1 degree already. Now that it has risen an additional 0.5 degree we can plug the numbers and see what further increase in temperature we get, in this case another 0.25 degree.

Tim and Bob,

Thanks for hanging in there with me on this. I think I've finally gotten my drug-addled brain around this. Learn something new every day! (Well maybe not *every* day...)

I'm still not completely convinced that the equation isn't simplified to the point where it's not really reflecting the real world so well, but I can see the value in a "back-of-the-envelope" calculation to get a rough boundary on the problem.

The bottom line is that I think it is a good thing to have a stable bias over time and temperature. The graphs in Bob's AES article are where the rubber meets the road, and he shows a much more stable bias with MOSFET's compared to BJT's. It's not clear of all of the reasons for this, and what can be done to overcome them. But at the present time, I am still sticking with my ThermalTrak BJT's. We are getting excellent bias stability and excellent sound.

If heat and energy efficiency and cost were not concerns, my second choice would be the lateral parts from Exicon (Semelab). But I still have a bad taste in my mouth from the vertical parts based on a multitude of problems with them. Plus everything else I've made sounds significantly better than that design that used vertical parts. YMMV.

 

[hitsware]

>If heat and energy efficiency and cost were not concerns, my >second choice would be the lateral parts from Exicon (Semelab).

Why not the Renesas equivalents ?
They are better 'bang for buck' ....
Here anyways .... (Sacto Ca)

 

[PB2]

Hi Pete,

Am I Joking? Absolutely not, I was at most a little bit of sarcastic. As a matter of fact, I spiced (years ago) the LT1166 in conjunction with a pair of vertical MOSFETs and the results were disappointing in terms of THD. Spicing it saved me a lot of time!
Did I sensed the current in Re? Yes I did, but I kept the unity gain frequency of the bias feedback loop below AF, somewhere between 1 and 10Hz, to avoid unwanted interactions that increase the distortion. See my article "Autobias for Mosfet Audio Output Stages", EW December 2003, pp. 17...20.

Cheers, Edmond.

I'd like to read your article, is it available online, or can you email it to me?

Pete B.

 

[Charles Hansen]

Why not the Renesas equivalents ?

The Semelabs are clones of the Renesas (nee Hitachi) parts. But the Semelabs have been improved somewhat. The P-channel Hitachi parts are "triode-like" until you get above 40 Vds (not good). They don't particularly match the transconductance of the N-channel parts very well. Plus the tempcos of the two parts don't match real well either. The Semelabs solve all of these problems.

The only bad thing about the Semelabs is that they don't make the smaller driver MOSFET's.

 

[PB2]

Re: OT Paper

Anyone have this paper scanned? I have it and will scan it if not, if anyone has is scanned already I'd appreciate an email of it.
"Distortion in low-noise amplifiers" by Eric F. Taylor, Wireless World August, 1977

and also the September part 2 issue?

Pete B.

A copy has been sent to me, so you can all stop worrying now ;-)
Thank you!

Pete B.

 

[Edmond Stuart]

I'd like to read your article, is it available online, or can you email it to me?

Pete B.

Hi Pete,

You have mail.

 

[GK]

Hi Pete,

You have mail.

Hey Edmond, me too please

Cheers,
Glen

 

[Bob Cordell]

Tim and Bob,

Thanks for hanging in there with me on this. I think I've finally gotten my drug-addled brain around this. Learn something new every day! (Well maybe not *every* day...)

I'm still not completely convinced that the equation isn't simplified to the point where it's not really reflecting the real world so well, but I can see the value in a "back-of-the-envelope" calculation to get a rough boundary on the problem.

The bottom line is that I think it is a good thing to have a stable bias over time and temperature. The graphs in Bob's AES article are where the rubber meets the road, and he shows a much more stable bias with MOSFET's compared to BJT's. It's not clear of all of the reasons for this, and what can be done to overcome them. But at the present time, I am still sticking with my ThermalTrak BJT's. We are getting excellent bias stability and excellent sound.

If heat and energy efficiency and cost were not concerns, my second choice would be the lateral parts from Exicon (Semelab). But I still have a bad taste in my mouth from the vertical parts based on a multitude of problems with them. Plus everything else I've made sounds significantly better than that design that used vertical parts. YMMV.

Thanks for hanging in there, Charles. We are definitely on the same page with respect to the desire for a stable bias over time and temperature. The ThermalTraks are wonderful, and I think it would have been difficult for you to achieve what you have in your amplifier with ordinary BJTs.

With respect to the greater thermal bias stability of the vertical MOSFETs shown in my paper, it is certainly fair to say that it is in part due to the lower gm of the MOSFET devices. I think that in the realm of thermal stability and the nature of the device, the verticals lie in between the lateral MOSFETs and BJT's in the spectrum. They all have their individual advantages and disadvantages.

Quite frankly, if I was tasked to do a Class-AB no-NFB amplifier (and no EC), I would probably also opt for the BJT ThermalTrak route, as you have.

If, however, I was tasked to do a Class-A amplifier of the same sort (and add significantly to Global Warming), I would then be tempted to do it with vertical MOSFETs, since Class-A operation drastically reduces the problem with transconductance droop, on the one hand, and the large amount of total idle bias and number of devices in Class-A reduces the amount of the transconductance droop in the first place.

Cheers,
Bob

 

[Bob Cordell]

Hi Pete,

You have mail.

Hi Edmond,

I'd appreciate a copy as well.

Bob

 

[Edmond Stuart]

Hi Edmond,
I'd appreciate a copy as well.
Bob

Hi Bob,
you have mail too.
Cheers, Edmond.

 

[PigletsDad]

Yes, I believe that is pretty much correct. There are probably some other time constants involved, but they are likely much less significant. I'm not much good at estimating heat capacities to arrive at the corresponding thermal time constants, however. But they can be measured. I would also tend to guess very roughly that the thermal "mass" of an element is somewhat proportional to its physical mass.

If you look at the SOA vs pulse time for a MOSFET, for example, you can see that it implies that the fastest time constants are measured in milliseconds.

By the same token, its pretty obvious that the time constant of any decent amplifier heat sink is in the minutes.

Cheers,
Bob

For pretty much any solid, the thermal mass per unit volume is the same - density and specific heat go inversely. This is another way of saying that most atoms have much the same radius. To be more nitpicking, for metals, the specific heat per mole is about 25J/K.

For aluminum (usual heatsink material) it is close to 0.9J/K/g. The volumetric heat capacity is of Al is 2.4J/K/cc. The conductivity is of pure Al is 237 W/m/K.

We can put these together to get the thermal diffusivity (in square metres per second) which is about 10^-5 for Al. This means that at length scales of say 1mm, temperature differences relax on a time scale of 0.1s. Go up to 1cm, and relaxation times are now 10s; 10cm (roughly heatsink size) and you are looking at time constants round 1000s.

HTH

 

[Bob Cordell]

For pretty much any solid, the thermal mass per unit volume is the same - density and specific heat go inversely. This is another way of saying that most atoms have much the same radius. To be more nitpicking, for metals, the specific heat per mole is about 25J/K.

For aluminum (usual heatsink material) it is close to 0.9J/K/g. The volumetric heat capacity is of Al is 2.4J/K/cc. The conductivity is of pure Al is 237 W/m/K.

We can put these together to get the thermal diffusivity (in square metres per second) which is about 10^-5 for Al. This means that at length scales of say 1mm, temperature differences relax on a time scale of 0.1s. Go up to 1cm, and relaxation times are now 10s; 10cm (roughly heatsink size) and you are looking at time constants round 1000s.

HTH

Neat! Thanks!

So heatsinks are way not isothermal on even a relatively slow dynamic basis (e.g., 1 minute), right?

I'm also wondering if therefore the behavior from a bipolar transistor heat source to a heat sensor 10 cm away is more like that of a slow transmission line than that of a single slow time constant. Thermo was not my greatest moment in college .

Cheers,
Bob

 

[PB2]

Neat! Thanks!

So heatsinks are way not isothermal on even a relatively slow dynamic basis (e.g., 1 minute), right?

I'm also wondering if therefore the behavior from a bipolar transistor heat source to a heat sensor 10 cm away is more like that of a slow transmission line than that of a single slow time constant. Thermo was not my greatest moment in college .

Cheers,
Bob

Hi Bob,

I'd say that it is distributed, but that an RC should probably work just fine as a model. More R and C as you move further away.
Since it heats up gradually I don't see how it could be a simple delay function. We've been saying that there is a long delay, and by that I mean a long time constant for the RC.

Pete B.

 

[PB2]

For pretty much any solid, the thermal mass per unit volume is the same - density and specific heat go inversely. This is another way of saying that most atoms have much the same radius. To be more nitpicking, for metals, the specific heat per mole is about 25J/K.

For aluminum (usual heatsink material) it is close to 0.9J/K/g. The volumetric heat capacity is of Al is 2.4J/K/cc. The conductivity is of pure Al is 237 W/m/K.

We can put these together to get the thermal diffusivity (in square metres per second) which is about 10^-5 for Al. This means that at length scales of say 1mm, temperature differences relax on a time scale of 0.1s. Go up to 1cm, and relaxation times are now 10s; 10cm (roughly heatsink size) and you are looking at time constants round 1000s.

HTH

Thanks for this.

Edmond, let's say that the sensor is 1 cm away, and we want as a sanity check to model this with an Re feedback based system in an attempt to match the distortion performance, knowing and expecting the same compromised thermal behavior. Seems to me your 1 to 10 Hz bandwidth was too much. One cm being 10s would suggest .1 Hz and that's not taking into consideration die to case/heat sink thermal resistance which suggests an even longer time constant. Did you try .1 or even lower bandwidth?

Pete B.