Temperature compensation for output stages


I wondering about the temperature compensation circuits for power amps. Does anyone know if there are some smart way (besides using a temp chip and a microcontroller) of creating suitable mV/degree and suitable DC-level sort of speaking?

With BJT it's quite easy, simple use a diode or a transistor as the temperature sensing element. But when we use mosfets we require 8-10 mV/degree C in temperature coefficient. One way to creating this is to use 4 or 5 PN-junctions.

Better ideas out there? The microcontroller thing is not too bad these days, don't you think? A small PIC12 or AVR would do the job and they are also cheap.
I'm aware of the Vbe multiplier technic but this is not an advance way to control temp coeff and DC-level at the same time.

When we are using mosfets the temperature coeff variates7-10 mV/degree C is my experience. My idea was some smart circuit to adjust the two parameters rather independantly. One idea is to use a low voltage rail-to-rail opamp together with a temperature sensor of some kind.
mosfet bias

Us simple folks use a source resistor. But I guess that is not enough intellectual stimulation for everyone. How about Nelson Pass's scheme for the Aleph current source with a mod to remove the AC component from the base of the bipolar transistor. (Hint: two resistors and a capacitor.)




  • windows.jpg
    48.6 KB · Views: 1,679
I seem to remember (oh, oh, already suspect) a Siliconix FET or MOSFET design book that had the scheme of two Vbe multipliers in series: one to control bias voltage and another to control temperature coefficient.

I'll try to look it up tonight and post later. What I'm saying now might not be making much sense. :)

The Alexander amp has a circuit with a LM431 to fix the DC-level and a transistor to fix the temperature coeff. The disadvantage of this is that the temperature coeff is somewhat limited. You can't choose 10mV/ degree and have 5 volts across.

To use a microcontroller is nowadays no problem. We have 8-pin devices (PIC, AVR) and good possibilities to read temperature.
I'm a boring guy and I want to discuss the topic only.

"How about Nelson Pass's scheme for the Aleph current source with a mod to remove the AC component from the base of the bipolar transistor. (Hint: two resistors and a capacitor.)"

I believe I did discuss the topic. Perhaps if you explain why you are trying to control the output stage bias current so precisely it would help.

Do a websearch on the terms "offset" and "slope". A couple of opamps and a silicon diode should do the trick. The diode forward voltage reduces about 2.2mV per deg C I think.

The offset sets the bias at startup when the devices are at basically room temp. The slope refers to the gain of the circuit which translates into how many mV reduction in bias per deg C increase.

Actually, here's a page with a schematic that illustrates the principle. 5k pot is offset, 10k pot is slope (gain).
Set the offset with a cold amp, then set the slope with a hot amp.

Re: I'm a boring guy and I want to discuss the topic only.

HarryHaller said:
Perhaps if you explain why you are trying to control the output stage bias current so precisely it would help.


Picture this:

Voltage across the gates of the output stage = 5 V (at room temp)

Temperature coeff = 9-11 mV/ degree C

With a simple transistor solution I will get around 6-8 volts and if the voltage is trimmed to 5-5.5 volts the temperature coeff will be too low.

My problem concerns only MOSFET's, no problems with BJT's.
I read this thread at work and as I had been agonising over how to do the bias on my Circlotron doomsday amplifier, I put my brain into high gear to try and think of a solution. I think I have something workable.

First off, the temp sensing fet MUST have the same threshold as BOTH the MATCHED output fets. I measure threshold voltage by joining gate to drain and then run a 100 ohm resistor to +12v, source to neg, and measure drain source voltage. Usually you can find a pair that will match within 10mV. Rude and crude maybe, but it's simple and works well.

For the test I got a matched pair of N-channel IXYS IXFH32N50Q 500v 32 amp 375 watt fets and wired one of them in threshold measurement config as described above. The second fet had it's source and gate to the first one's source and gate. It's drain went to a 40 vdc supply via a 10 ohm current sense resistor. The threshold voltage was 4.06 v.

With the heastink at 20 deg C the second fet pulled a constant 108.6 mA. At 70deg C heatsink temp when heated via a soldering iron the second fet pulled 112.0 mA. ;) Is that reasonable enough?

For a complementary symmetry setup the matching still applies but the temp sense fet MUST have equal resistors on the gate to source and gate to drain so it makes EXACTLY double the drop it would otherwise. It is probably the best idea if you use the same kind of fet for sensing as you do for the main output.

I have been think since I started this thread. I haven't got much input from you (advanced subject!) but I got a crazy and extreme idea:

PIC or AVR processor with ADC and EEPROM, placed between the upper and lower cascode transistors. My circuit replaces the conventional transistor+2 resistor thing. Working voltage 2.5 - 8V and current consumption max 5-8 mA

Processor in sleep mode most of the time and internal RC-oscillator or 32kHz crystal.

Tempsensor made of transistor + 2 resistors to create enough voltage for the ADC.

ADC for sensing voltage across the gates of the output FET's.

R-2R ladder to create a DAC, directly from the processor or via a shift register (depending on current consumption and emission).

My idea is to calibrate the amp and then store the values in EE-memory.

Calibration like this:

Enter "learn"-mode

1 Set desired bias with cold heat sinks

2 Store temp and voltage

3 Increase current in order to get the heat sinks really hot (60-80 deg C).

4 Store temp and voltage

5 The processor calculates the temp coeff and store it. Since I have a processor it's possible to let the temp curve be unlinear.

Sounds my idea like an overkill? I have to think a litte bit more about the DAC-thing. My goal is to make it easy and "quite".
OK. this is an all-stops out effort to measure the variation in mosfet quiescent current with change in temperature. That is to say, any variation using the compensation method I am suggesting. The test I did last night was pretty much a seat of the pants effort at home. Today at work I used several meters and a thermocouple temp sensor that all have traceable calibration, and besides, one of my job functions is to actually perform temp measurements on power electronic stuff, so no shortcuts were taken. These figures are accurate.

Actuall, afterI began testing I realised the test setup is actually a current mirror of sorts. But I can't stress enough "all the devices MUST have the same Vgs threshold voltage".

The test setup consisted of both TO-247 fets bolted opposite sides of a 20mm square aluminium bar 150mm long. A small hole was drilled between the two fets and filled with heatsink paste. The thermocouple was inserted into this hole so it would be shielded from air currents and get the best possible thermal coupling. The thermal compound used on the main fet was Thermstrate TC, (amazing stuff, that) and the bias/sense fet had a silpad under it. The aluminium bar also had a 25 watt metal clad resistor attached so I could adjust the aluminium bar to whatever temp I wanted. The thermocouple meter was a fluke 52.

The main fet had a constant 50 vdc across it, and every 5 degrees I recorded it's current, the bias voltage, the bias fet current and the temp. It's all here on a spreadsheet with graphs. This was a BIG effort. I'm certainly going to make use of what I found. I hope others like it too.

P.S. The normal way of setting bias with a transistor and a pot is all wrong I think because when you set it at a particular temp what you are doing is changing the slope, or tempco, of the bias so that it only works at THAT temp. At other temps it will either dry up or dig in too hard. The only way to do it is to have the temp sense fet gain fixed at x2 with equal value resistors and use a fet that has the same characteristics as the output devices. That way the nonlinearites in the tempco will track the whole way. But remember, make sure the sense fet has intimate thermal contact with the output fets, e.g. opposite side of the heatsink, and... *match the 3 device thresholds* ;) Comments invited please.



  • mosfet bias vs temp measurements.zip
    24.8 KB · Views: 263
This approach looks quite simple. Essentially 2 Vbe multiplier ccts in series between the output device gates. Only *one* is thermally coupled to the output fets. Both are adjustable, one for offset (the initial or cold bias voltage) and the other for slope (rate of bias voltage decrease per deg C increase).
See page 15.

I just can't seem to stay away from this thread...
LT have a chip specifically for this biasing bizzo. LT1166. In their words:
The LT®1166 is a bias generating system for controlling class AB output current in high powered amplifiers. When connected with external transistors, the circuit becomes a unity-gain voltage follower. The LT1166 is ideally suited for driving power MOSFET devices because it eliminates all quiescent current adjustments and critical transistor matching. Multiple output stages using the LT1166 can be paralleled to obtain higher output current. Thermal runaway of the quiescent point is eliminated because the bias system senses the current in each power transistor by using a small external sense resistor. A high speed regulator loop controls the amount of drive applied to each power device. The LT1166 can be biased from a pair of resistors or current sources and because it operates on the drive voltage to the output transistors, it operates on any supply voltage.


Paid Member
2002-03-29 5:19 am

There's no agonizing to do. Some reading and thinking, yes,
but no agonizing. And surely you don't need microprocessors
to do this job.

Go read Doug Self's book, or his articles in EW/WW. I don't
know if the bias stuff is covered on his web site so you just
may need to spend $35 and buy the book.

Self covers it all in painstaking detail.
The LT1166 seems to be able to do the job but I'm not very keen on using unusual chips with no second source. There seems to be a very little problem with this chip and this is speed limitation. Since the chip creates a control loop it seems that extremely low distortion isn't possible.

My microcontroller idea are interesting and also rather compact and not very expensive. My brain is at the moment puzzled by creating analogue signal in a cheap simple way.

Microcontrollers are very rare in hifi but it's nowadays a component as the transistors. Nothing wrong in using those parts but they need to be carefully designed into the whole amp because they can hang up and also interfere the audio signal with clock signals.

My experience with PIC and AVR is very good. They very seldom creates problems in our products (at work) nor in real life.