AUDIO

S everal factors may contribute tothe lack of accuracy:

1) Mismatch between relativetemperature coefficients of mosfetsand bjts as consequence of thevariability of the gate thresholdvoltage as well as the temperaturecoefficient of VGS

2) Thermal delay and attenuationof the coupling between outputdevice and sensing element.

3) Drivers - if included - thatoperate at a different temperature.

4) Long-term drift of thresholdvoltages as a result of aging1.

5) Errors in adjusting the bias level

for each individual amplifier.So it seems natural to replace the

VBE multiplier - which in factprovides a kind of error feed forward2

- by a control loop based on feedbackof the bias current itself. In the past,several attempts has been undertakenin this direction, but none of themseem to me suitable for high-endapplications, as they are intrusive alsoon other parts of the amplifier. Thiscould raise distortion3,4, complicateHF compensation5,6 or beincompatible3,4,5 with acomplementary source followerarrangement (which I prefer), or

could be too complex7. Nevertheless,reference 5, ingenious in its ownright, inspired me to a re-design thatovercomes these shortcomings.

The new design comprises threesections: a bias current sensor, anisolator and an integrator. Each ofthem will be discussed below indetail.

Bias sensorThe bias sensor’s purpose is to detectany deviation from the nominal biaslevel without being influenced by thecurrent distribution over the outputtransistors, i.e. independently of the

December 2003 ELECTRONICS WORLD 17

It’s said that biasing vertical d-mosfets in a class-AB output stage isless critical than biasing their bipolar counterparts. So manydesigners are content with the classical VBE multiplier as biasgenerator. However, the accuracy needed for utmost performance (interms of cross-over distortion and quiescent dissipation), cannot beprovided by such a circuit, even when it is thermally coupled to oneof the output devices. Edmond Stuart explains

Autobias for mosfetaudio output stages

Fig. 1. Circuitdiagram of theautobiasgeneratorconnected to atypical mosfetoutput stage

current delivered to the load (RL). Thisis accomplished by sensing thevoltages across the source resistors ofthe output stage and passing them to anon-linear network, comprising twocurrent mirrors (transistor pairs T3,4and T5,6 respectively). Two equalreference currents are supplied to theinputs of the mirrors via R6 and R8.Since the voltage across each sourceresistors adds an off-set to VBE of T3and T6, the currents reflected by themwill be lower than the reference.

Under quiescent conditions, thisoffset is such that each mirror reflectsonly 50% of the reference current. Asdictated by the laws of physics*, thiscondition is met if the offset voltage is18mV (at room temperature).Together with R1 and R2, this 18mVdefines the quiescent current of theoutput stage (e.g. 100mA if

R1=R2=0.18Ω). To maintain thiscondition the reflected currents aresummed together and subtracted froma third reference current supplied byR9. The resulting difference, IERR, tellsus whether the output stage is under,correct or over biased, thus IERR canbe used as a feedback signal, Fig. 2.

Next, what happens in the presenceof small or medium signals? This isillustrated by the middle curve of Fig.3. It shows the relationship betweenerror signal and output current at thenominal bias level. As one can see, thecrux is that the error signal stays veryclose to zero. Apparently, the non-linear behaviour of the mosfets, inparticular in their weak and moderateinversion region, matches thecharacteristics of the bias sensor quitewell. The other curves show the errorsignal if the output stage is forced to

an under - or over biased state.Beyond output currents of ca. 0.6Athe error signal is pinched off. Thislooks like a disadvantage, but itturned out to be beneficial, asexplained below.

Finally, what happens in thepresence of large signals? Suppose T1carries a large current and T2 is turnedoff. In this case the reflected currentin T3 is zero, while in T6 it is 100%.Summed together and subtracted fromthe third reference, the resultant errorsignal is again zero. This is exactlywhat we want, as an output stageoperating in class B provides noinformation about the quiescentcurrent, thus the error signal has bepinched off in order to preserve thecharge of the integrator’s capacitor(C3).

Since music has a high peak/averageratio - some 20dB - the average signallevel - even at maximum volume - iswell within the capture range of thebias sensor. Sine waves at full powergive no trouble either, as the relativetime traversing the capture region islong enough to let the control loop doits work. However, a large squarewave pushes the output stagecontinuously in class B and no errorsignal is produced at all, leaving theintegrator in an undefined state.

IsolatorTo avoid any adverse interactionbetween common mode anddifferential signals at the gates (nodeA and B), an isolator has to beinserted somewhere inside the servoloop. Putting it between bias sensorand integrator greatly simplifies thecircuit, as the integrator can nowsimply use the bias voltage as supply.Given the bipolar nature of the errorsignal, two opto-couplers (U1, U2) areneeded, one for charging, the other fordischarging C3. They are specified foroperating at low currents (<1mA).Apart from their primary task(isolation), they serve one morepurpose: masking the tiny deviationsof the error signal, as can be seen atthe middle curve of fig. 3. Thereduced transfer ratio at very lowcurrents from which any opto-couplersuffers (see Fig. 4), meets thispurpose nicely. R11 delivers thesupply voltage to U1, while beinglimited by LED D3 reducing VCE ofT3 and T6 to approximately the samelevel of T4 and T5 and preventingsimultaneous conduction of U1 andU2.

IntegratorDepending on the mosfets actuallyused, bias voltage can vary from 2 to10V. To handle this range it leads

ELECTRONICS WORLD December 2003

AUDIO

18

Fig. 2. Simulatederror signal as

function of biascurrent

(IREF=1mA,R1,2=0.167ΩΩ, RL=

∞∞).

Fig. 3 Simulatederror signal as

function of outputcurrent at bias

levels (from top tobottom) of 60, 90,

120, 150 and180mA. Noticethat at 120mA

this signal staysclose to zero.

AUDIO

almost automatically to the kind oftopology, as shown in fig. 1. In spiteof its simplicity, this integrator, or tobe more precise - integrating shuntregulator, exhibits a dynamic outputimpedance that is low enough (<2Ω)to cope with AC currents from thedriver stage. In order to excludeinteraction at AF, integrator capacitorC3 is rated such that the unity gainfrequency of the servo loop fallsbelow the audio spectrum,somewhere between 1 and 10Hz,depending on transfer ratio of theopto-couplers. D4 discharges C3during switch off and protect the gateof T7. For reliable operation, lowleakage at the integrator input isessential, so D4 should be protectedfrom light. Zener diode D5 is rated atthe maximum expected bias voltageplus a small margin and should beadapted to fit a particular design. Ifsome component fails or if nasty testsignals have been applied, this diodeprotects the output stage againstexcessive common mode currents.

AccuracyThe bias level is very sensitive tomismatches of the transistors pairsand reference currents. Base-emittervoltages T3 and T4 respectively T5and T6 should match at least within0.5mV. A quad transistor like aMAT04 or CA3086, selected on lowVos, meets this requirement. For thesame reason, R6, R8 and R9 shouldmatch at least within 0.5% as well asthe equivalent emitter series resistors;hence R7 and R10 are rated slightlyhigher than R4 and R5. Since aMAT04 is equipped with smallreverse connected diodes betweenbase and emitter, Schottky diodes D1,D2 are added to protect them.

Now we come to a moot point: thevoltage of 18mV across R1 and R2

which all relies upon, varies linearwith the absolute ambienttemperature*. Of course, it variesmuch less than the junctiontemperature of the output devices andthermal runaway is precluded, butstill it is not constant. It is not yetclear to me whether this should beregarded as flaw or feature, as onecould argue that the decreasedtransconductance of mosfets atelevated temperatures just needs anincreased bias level. Interestingly, abias control IC from LinearTechnology6 shows the sametemperature dependence, which couldnot be explained by an inherentshortcoming of the basic circuit. So Iconcluded that this property has beenadded on purpose. Asking why,Linear Technology was unable togive a satisfactory explanation. Ifanybody could shed light on this

matter, please let me know.

Experimental resultsTo see if the circuit is generallyapplicable, I tested it on severalcombinations of mosfets, all capableof delivering 12 to 20A, but ofdifferent types and brands. Tominimise temperature effects,measurements were done at a reducedsupply voltage of 2x16V, the mosfetsmounted on a large heatsink withforced air cooling and at a frequencyof 1kHz. Static measurements weredone at an even lower voltage,2x7.5V. Since these were very timeconsuming, I have done this only incase 1 and 2.

In the first instance, dynamicbehaviour of bias current wasobserved by means of anoscilloscope, but changes at variousoutput levels were hardly visible anddifficult to quantify, except in case 1,which showed an increase of 5% atmaximum output power. Instead, I

December 2003 ELECTRONICS WORLD 19

Ripple amplitude of the error signals measured at VO = 8Vpp and 1kHzas well as estimated ripple on bias voltage (VRPL) at 20Hz.

Case Type Manufacturer RL (Ω IERR (µA eff) IINT (µA eff) VRPL (mVeff)

1 IRFP240 Int. Rectifier 4 15.9 3.3 12.9IRFP9240 Idem ∞ 62.4 157.0 568

2 2SK1530 Toshiba 4 4.9 <0.1 <0.42SJ201 Idem ∞ 3.2 <0.1 <0.4

3 IRFP240 Intersil 4 14.2 1.3 4.7FQA12P20 Fairchild ∞ 5.0 <0.1 <0.4

4 IRFP240 Intersil 4 7.2 0.2 0.8SFH9154 Fairchild ∞ 5.6 <0.1 <0.4

5 FQA19N20 Fairchild 4 11.6 0.7 2.5FQA12P20 Idem ∞ 5.1 <0.1 <0.4

Fig. 4. Transferfunction of theisolator.

* VBE = VT Ln2, where VT = kT/q, thethermal voltage (25.86mV at 300K).

AUDIO

used a DVM to measure the rippleamplitude of the error signal beforeand after the isolator under load andno load conditions. Next, I estimatedthe ripple-on-bias voltage at 20Hz,according to VRPL = IINT / (2 π f C3 ),instead of a direct measurement at20Hz, because thermal modulationcould be disturbing. See table forresults.

The first trial was ratherdisappointing, not to say confusing.The error-function, Fig. 5, is heavilyskewed and ripple currents are high(see table, case 1). Transconductanceswere reasonably matched - within 20%, so something else spoiled it. Itappeared that the output conductance(GOS) of the IRFP9240 was theculprit. At a drain current of 125 mA,GOS is about 5mA/V, while theIRFP240 shows a GOS of only0.4mA/V (corresponding to an Earlyvoltage of 25V and 312V

respectively). So an increase of outputvoltage of 1V -without load- willincrease the quiescent current by4.6mA. No wonder that the errorfunction is skewed. At higher draincurrents the IRFP9240 behaves better:at 1A for instance, the Early voltageraises to ca. 75V. This explains whythe ripple on IERR becomes muchsmaller when the output stage isloaded with a low impedance.However, speaker impedances are notalways that low – at resonance up totenfold, so a no load condition has alsoto be taken into account. Withoutblaming the manufacturer, I cannotrecommend this mosfet pair. After all,these devices were designed forswitching, not for drivingloudspeakers.

In the next trial I used acomplementary pair from Toshiba,especially intended for linearapplications: 2SK1530 and 2SJ201.

Due to closely matchedtransconductances (within 5%) andhigh Early voltage (over 300V) forboth N- and P-channel parts, theresults were far better. The errorfunction, Fig. 6, is in accordance withthe simulation, although skew isslightly higher and in the oppositedirection. Ripple currents were hardlymeasurable.

In the last three cases I tested severalother samples (courtesy of Fairchild)which are less expensive, but, as incase 1, primarily targeted for fastswitching. Results were almost asgood as in case 2 and I see no reasonnot to use these mosfets, except thatthe higher gate threshold voltagereduces the maximum output powersomewhat, that is, without using aboosted power supply for the drivers.

I have also investigated a fewcombinations of two 20A N-channeland three 12A P-channel devices.Because no improvements were seen, Iwill not discuss them any further.Using bjts instead of mosfets willprobably not work at all, as Spicesimulations were very discouraging.For lack of Spice models, lateral d-mosfets have not been investigated.

ConclusionProvided that output devices areselected with some care, in particularwith regard to trans- and outputconductance, the proposed circuitcomes up to all expectations. Since thecircuit acts only on the bias voltageand is not intrusive on any other partof the amplifier, it should be easy toincorporate into new or existingdesigns like mentioned in ref. 8 and 9.

References1. Finnegan, T, ‘Going linear with

power mosfets, part 2’, EW, Sept. 1992, p.781.

2. Self, D, ‘Thermal dynamics in audio power’, EW, May 1996, pp. 410-415.

3. Roehr, W., ‘The autobias amplifier’, JAES, April 1982, pp. 208-216.

4. Siliconix, ‘Mospower Application Handbook’, Chapter 6, pp.105-110

5. Gevel, M. van de, ‘Audio power with a new loop’, EW, Feb. 1996, pp.140-143.

6. Datasheet LT1166, Automatic bias system, Linear Technology.

7. Brown, I., ‘Opto-bias basis for better power amps’, EW+WW, Feb. 1992, pp. 107-109

8. Hitachi Mosfet Handbook, Section 6.1, pp. 143-151.9. Stochino, G.,’300V/µs power’, EW, April 1997, pp.278-282

ELECTRONICS WORLD December 200320

Fig. 6. Same as fig.5,but measured on a

2SK1530/2SJ201 pair.

Fig. 5. Same as fig. 3,but measured on aIRFP240/IRFP9240

pair at bias levels of57, 85, 118, 142 and

171mA.