Lavaradin Amp and "Memory distortion"...

Mark Finnis discovered that one a while back and posted it in a thread called (I think) Son of OptiMOS. I read it, and found it fascinating, but to my recollection, no one had ever heard the amp, nor had they even heard of it. One of these days when I have nothing better to do (which means I won't get to it for another century or two at the rate I'm going...), I'm going to take a whack at a circuit incorporating those ideas, as it seems to make sense.
If anyone has fiddled with it since then, by all means 'fess up.

Grey
 
Memory

Discussed this with several people. Does make for a nice technical sounding piece to read.

Have nobody I know of, or even know of as a reliable source, who has compared anything built this way, so don't even have what I consider a "respected opinion" that this is making a difference. Is not unusual for person modifying to have very good opinion of changes. As feedback is MUCH, MUCH quicker than any thermal drift could be, should be corrected out with no problem at all. Would like to see and hear such an amp, but at this point don't have anything to judge by. Sounds great, but so does clamp on gizzmo that gives you 100 miles per gallon in your car.

Regards,
Greg
 
I am sure that memory distortion is an important factor in determining the sound of an amplifier but in all fairness cascoding a circuit has other benefits such as lowering Miller capacitance and increasing the bandwidth of the circuit which in turn means lower distortion.
We should be careful not to attribute all the claimed benefits solely to the reduction of memory distortion and that it is the sole criteria for a good design.

Jam
 
I have not played with this theory eather but i know if you blow air around an Amp's input the DC off set will change rather quick and this is surly alott slower than any Audio so i give the Idea credit and will play with it wen done my other stuff but not a high priorty right now for me so have to content myself by keeping a thermaly stable envioment.
 
review of peufeu site

I stumbled on this very old thread and followed the link. Fascinating what this Frech guy by the name of Pierre has put together. I have seen worse monographs, and some of those even passed for master's theses in EE!

- He maintains that THD and intermodulation distortion measurements are meaningless because of the use continuous signals whereas the power envelope of music changes continuously.
- Changes in the power envelope cause nonconstant heating in the input and gain stage transistors- As V_BE changes, so do the bias points of all transistors and hence their distortion spectrum. The human ear appears to be pretty immune to constant harmonic distortion, but will perceive a changing distortion spectrum as unnatural.
So far, so good. In my opinion this
might already explain why some almost traditional circuit concepts sound particularly good, even if the same THD levels can be achieved without these goodies:
- cascoding input and VAS transistors helps a great deal in minimizing both power dissipation and its changes with the input signal
- sacrificing gain and hence global feedback for more local feedback will stabilize bias points; in particular, and this is a point overlooked by Pierre, emitter degeneration resistors will alleviate the effect of V_BE changes, the more the higher the voltage drop across the resistor is
- in symmetrical designs, i.e. with an NPN and an PNP input pair and an NPN and PNP VAS, the effects should compensate in first approximation

Pierre analyzed a conventional amp design with degenerated input pair, current mirror on the input pair, single transistor VAS (without emitter degeneration!) and complementary emitter follower output. He models the heat dissipation by a series of resistors and capacitors driven by a current source (this an an absolutely correct way to model heat conduction, I did a similar simulation for laser optics). The neat thing is that this current is proportional to V_CE x I_C. I don't know how that is done in Spice. Neither do I know how to convert the voltage that corresponds to the die temperature back into V_BE in the transistor model. But then Pierre uses Microcap which I am not familiar with.

He uses a 20 ms input transient that is constant at a high level slightly below the clipping level to offset the themal balance. The bias points need a couple of seconds to recover. He compares the voltage right after the end of the transient to the steady state voltage. This gives him an arbitrary measure for memory distortion. Switching the thermal model on and off for various transistors, he identifies the offenders (VAS, current mirror, input transistors, tail and VAS current sources in descending order).

This arbitrary measure is fine IMHO to optimize the circuit. But in order to really prove his point is valid in absolute figures, Pierre should show how the distortion spectrum is different right after the transient. First of all, he would need to get the absolute change in power dissipation right. Any musical signal, even a kick drum, will consist of a sum of damped sines and hence have no AC content. So the excursions to either side would probably be shorter than 50 ms (if the music is bandlimited to >20 Hz) and hence shorther than the thermal time constant of most transisor dies, resulting in some averaging.

In typical use with typical speakers, the music will be around 90 dB(A), so 1 W would be ok. This corresponds to 2 V_rms into 4 Ohms. So the distortion spectrum of a 1 or 10 kHz signal with 2 V at the output should be calculated. Then a loud drum or whatever could be simulated by a 500 ms 20 Hz sine burst at +20 dB relative to the 1 W signal. The distortion spectrum immeadiately after the sine burst should then be computed and compared to the original spectrum. If the 2nd harmonic changes from -105 to -107dB and the third goes up from -115 to -112 dB , then it will probably no be a relevant effect. But if there are 5 dB changes at -90 which might be possible in an average design, the whole idea of memory distortion would stand a high chance of being valid.

In terms of optimization, Pierre proposes to cascode all current sources and use pretty strong emitter degeneration on the current mirror (1k) - all fine with me.
For the input pair he compares a classic degenerated pair to cascoded, CFP and "new". For the cascode, he uses JFets to cascode bipolar input transistors - unusual enough! Also, he takes the input voltage for the Fets from the bases of the input transistors. I have never seen an input cascode like this and will have to think about it. I guess the gate-drain capacitance will result in a nice miller capacitance, completely defeating the tradidional reasons for using a cascode on the input, i.e. to get rid of nonlinear input capacitances. This nonwithstanding, the cascode helps to reduce memory distotion considerably by keeping the input transistors at constant voltage. This aim seems to be the reason for the weird cascode connection. In my eyes, it might be better to reference the cascode voltage to the emitters à la Pass and Borbely.
Secondly, he tries to keep the input transistors at constant current by employing complementary feedback pairs as input devices. Again, I have never seen thelikes and will have to contemplate this. Reduction of memory effect is on par with the cascode.
Finally he combines CFP and cascode in the "new" circuit to keep voltage and power constant which results in yet higher reduction of memory distortion.
The simulations of input stage linearity (diff current as a function of diff input voltage) of both CFP and "new" look very good so I will probably use this in one of my designs once I have mulled over dynamic effects.

I don't find his improvement of the VAS too convincing. He takes the undegenerated (!) VAS and cascodes it - fine. Then he adds the "magic" resistor between VAS collector and cascode emitter. The basic idea is to compensate the parabolic power dissipation in the gain transistor by taking some of this power away to be dissipated in the resistor the value of which is chosen so that the power in the resistor becomes a tangent to the parabula. That is fine, only I think the resistor should be between emitter and rail which would help to linearize the VAS stage considerably.

Another point that one might derive from the fact that the time constants for memory distortion (if it is relevant) are on the order of 100 ms to 10 s is that it might be worthwhile to rethink the feedback and gain tailoring.

The GBW of a circuit is basically fixed by the output transistors and their drive circuit. One could opt for high open loop gain an begin to roll it off at very low frequencies. As this results in nonlinear phase of the amplifier with frequency, current wisdom has it that the open loop gain should be lowered so that the dominant pole can be at >20 kHz. If memory distortion is indeed relevant, the added open loop gain at LF might, however, be desirable to get rid of memory distortion through global feedback. I will have to think whether local emitter degeneration might even me more, i.e. doubly effective...

Hope this will start a lively discussion!


Eric
 
Hi Eric, This was a lot to read through!!

My thoughts about the memory distortion is :

The idear is good. When you change the temperature on a silicon device we all know that it will change it parameters.
If you take a look on BJT datasheet you will se a increase in currentgain, a lower vbe drop as a result of a temperature rise.

So far so good.

Pierre has written on his page that a high openloop gain will compensate for this phenomena.
The reason is that it will only need a small currentchange to get a high voltage switch. So the current through each stage is more constant compared with an amp with a lower openloop gain.

That is not enough to get a constants temperature through the transistors/fet.

So he tries to get a constant temperature in the devices who senses voltage changes.
It is impossible to get a constant power/temperature through every device in the circuit so he stick to the important parts.

This is done by him by adding resistors inside a cascodestage. IT will lower the voltage across the transistor (who have been cascoded) when the current through it rises. => constants power ... This works only when the value of the resistor is selected with care.

By moving the resistor to the emitter instead of the collector you will generate large voltage shift in the stage driving the VAS which again turn into a change in power/temperature in the stage before the VAS.. Raise in memory distortion.

Will it be a drawback? What will have the biggest impact on the sound quality? The linearization of the VAS or the "magic resistor"?

For the modified diff amp:

I agree with you regarding the miller capacity. To cancel this out you need to add another cascode stage to cascode the cascode FET's. ;) hehe
But it should not be attached to +in and -in but to AC-gnd.

I think all these extras will add some extra delay to the settling time of the amp when it have to settle after an impulse!?

As you write about GBW. All this extra will maybe require a pole at a low frequency to get stable. And if this is right!? the effect of these "memory cancelling" changes will be degenerated because of the falling openloop gain at high frequency?

One of the good things about keeping the first pole outside the audio band is that you can get a low phase change at 10KHz - 20KHz.

Drawback : is hard to get stable in an amp with global negative feedback but it is very easy to optain in an amp without global negative feedback.

But it will shurely have a great benefit on a speaker where the drivers are alignet to zero phase between them. A delay/phase change in the sound is very easy to hear.

mmmhhh?? i can not remember the rest right know!

I will get back later.

Sonny
 
degenerated VAS

Hi Sonny,

you are right that a degenerated VAS will result in slightly higher voltage swing in the current mirror. But even if the VAS is degenerated to a voltage gain of 100 (I realize it should be regarded primarily as a current in stage), the swing will be be +/- 0.5 V for a +/- 50 V output signal. Any cascode can handle that without flinching....


Greetings,

Eric
 
Thermal Tails

Yes, yes, and yes:

This concept, in the guise of whichever name one prefers, is IMHO a big AHA when grokked.

It is very easy to ignore dynamic mechanisms and pretend that static or steady state signals are sufficient to design, simulate, build and test audio equipment with good sonics.

Just about any device, active or passive, will be at least somewhat temperature dependent. Even bulk foil resistors [g].
Those with a resistive/dissipative characteristic [transistors, tubes, resistors] have a bonus tendency to thermally self modulate under dynamic conditions. Delta power, delta temp, delta parameters. As an extra added bonus, devices in/on a common substrate [integrated circuits for those who fell asleep in that part of the class] will thermally cross modulate. Oh, fun. A signal event causes each device to share a thermal shock wave with it's neighbors. Interestingly enough, typical time constants associated with these thermal waves correspond with frequencies in the low kiloHertz range, right in the area of the ear/brain's max sensitivity. Or, at least some of us claim to hear glare, grit, fuzz, mush or general blase-ness associated with most monolithic opamps in the signal path. Even CFB ones...

I casually claim that thermal parametric modulation/shift is an important and usually overlooked factor in sonics, probably as significant in the strictly analog domain as jitter is to digital/analog conversion.

Here is the general scenario:
The input diff pair, through loop feedback, attempts to correct errors elsewhere in the signal path encompassed by the loop. But, any errors, including thermally generated parametric shifts, from/in the input diff pair will not be corrected by the feedback loop, and will pass right through, amplified by the closed loop gain. Not just thermally induced DC input offset drift. The input diff pair is in a state of "constant perturbation", attempting to correct errors while being modulated by the last error. Yes, under steady state "thermal equilibrium" conditions the loop may be unconditionally stable with sufficient phase margin, but still be chasing it's tail so to speak, and almost none of this will show up with standard sine/square tests. Some intermod test signals may show a slight increase, but that is really fumbling around in the dark. What one ends up with is a "tracking" fuzz signal, not too many tens of dB below peak levels, and ususally above the static noise floor, that again, simply does not show under conditions of thermal equilibrium, ie steady state testing.

So, enough complaining on my part of what is broken, here a a few constructive ideas:

Ignore it. Can't measure it, must not exist. All sounds the same ennyhow. Just listen to mp3.

Less or no loop feedback. Does not solve it, but perturbs is less.

Run hot. If the idle power is significantly larger than the delta power, then the effects are of course smaller. Pass, the one and only, is onto something here.

Use tubes and iron. No magic, just a nice big thermal bath, with constant power mode almost by default.

Design for constant power in critical devices, like Pierre in the original link.

Save up a boatload of money and buy Halcro electronics.

For less money, hire live acoustic musicians to serenade you.

Share and enjoy...
 
Thermal effects

Mr. Wildmonkeysects is actually on to something here but it is not quite as hopeless as one might think. It is interesting that many of the classic "audiophile" design techniques tend to minimize this
effect:

1. Oversized resistors ie. 2Watt in place of 1/2 watt.

2. Low tempco resistors. Maybe "naked" Vishays sound better due to better convection to the resistor foil element.

3. Dual transistors in same package share a more common thermal environment.

4. Multiple capitors in parallel.

5. Class A and cascode operation.

6. Oversized heatsinks.

7. The effect of thermal coupling in op amps is a well known effect and extreme attention is paid to this designing a precision op amp!

Time constants for semiconductors are smaller than you would think and numbers in the millisecond and smaller range sound similar to those I encountered in telecom protection circuit design. Maybe some experiments with coupling semicondutors to thermal masses is in order.

H.H.
 
What a inlightening Discussion I find some truth to all or most of the info presented. I would like to add the Following........

(1) true IC's share a comon thermal mass and thus all components are thermaly stressed by the same amount. Well as was pointed out most opamp designs deal with this by the Placement of critical input stage devices in sutch a mannor as to reduce the sensitivity to electrical imbalance caused by thermal effects. While it is true that requiring an opamp to dissipate large amounts of power relitive to it's Bias level is going to Reduce Performance regardless of design, IC makers have a wounderfull device to deal with this problem called a Buffer or simply a compimentry pair of Emmitter followers on a chip. This allows the Opamp to control precision and the Buffer to handel the Thermal management caused by the power into the load.

(2) stable thermal operation most certently be obtained by having a Large Bias level relitive to the changing signal dependent levels. Class A is just one way to obtain this, another way is to afix the semiconductors on a large thermal mass or a large heatsink. this will stabilize the tempeture of the semiconductors regardless of the Relitive bias levels. this the rationall for using 2 watt resistors to dissipate less than 1/8th of a watt.

(3) having the openloop gain of an Amplifier to extend beyond the 20KHz. Audio range can be unstable if at the same time the open-loop gain is large. to get stability and a dominent pole above 20K can be done by reducing the Openloop gain. this will allow a stable single pole rolloff.
 
input pair:
Nonlinearities in the gain of the input pair will be corrected by global feedback, just like any other gain in the circuit! Just think of how nonlinear a long tailed pair without global feedback is and how linear a discrete or integrated op amp is.

large thermal masses attached to transistors:
the relevant time constants are for processes within the die and between die and its immediate surroundings (lead frame and first mm of plastic housing). External masses will be too slow.

resitors:
are way more linear than semiconductors, so resistors inside the loop won't matter, however for gain setting feedback resistors I also try to use low TC, oversizing and parallel devices

capacitors:
I don't get the point. Paralleling will lower ESR but there shouldn't be thermal issues.


Greetings,

Eric
 
I really don't understand most of the technical stuff you guys are talking about but I had achance to hear the Lavardin amps at ashow and then audition them at home. I heard the IT integrated at the show and the IS reference at home. I thought that the amps were not particularly impressive compared to my Jadis Orchestra Reference and Naim 72/hi-cap/180. They have a softer sound than the usual transistor amps but I found it also robs them of detail and separation to a great degree. I really didn't see what the fuss is allabout since they were highly reviewed worldwide. In fact I found them much too expensive, that's around 4000-5000usd I think for a 2x35w integrated and the build quality was not what you would expect at the price. The case was folded metal and the parts quality was nothing spectacular. The sound quality was the same at the show and I had two amps at home (IS+ISref) so it is pretty safe to say that it was not system/room/sample dependent.
 
Lavardin patents and schematics

in case anybody is interested, the patents behind Lavardin, and also the basic schematic can be seen here:
http://www.delphion.com/details?pn10=US05512858

Mr Gerard Perrot has been known as "Hephaistos" in the 80ies, when he launched a series of articles about amplifier design in the French DIY magazin "L'Audiophile". He developped the notion of thermic distortion and showed by experiments, that certain transistors show more of this, others less, and tubes nothing of that. He then made another series on amplifier topologies, and compared them both by measurements and by ear. I found those articles very interesting. The schematic presented in the patent is the logical conclusion from all these articles.

Mr Perrot from Hephaistos laboratories also held some lectures on AES conventions, please use the AES search engine for that:
http://www.aes.org/publications/preprints/preprints_search.cfm

regards,
Hartmut
 
I think i posted this on some other thread, but whattahell :) I ended up building a stereo unit of the "Lavardin-clone" headphone amp on protoboard, originally intended as an experiment as a preamp.
Well, the thing is that as a preamp it worked quite nice, but no so nice that i'd preffer it over a Kaneda, Pass, Elliot or some other simpler preamp (even a quality op-amp based one). Yet, as a headphone amp it was VERY VERY good, and i could see what the guy reffered as the "solid state harshness" being gone. Really smooth sounding. I was considering it for the headphone out on my future preamp, but i first want to try the design Rod Elliot proposed using the high speed TI op-amps.