LM3886 x 2 in parallel

[jmar]

I bought one of these modules off Ebay that is a prebuilt stereo amp with (2) lm3886 on board.

Looks like I can run these in parallel pretty easily with the addition of the output resistors. 1/10th ohm at 3w on each output and then into a 4 ohm speaker.

I've seen in some schematics as well, the components regarding the noninverting input being shown. Looks like a 20k feedback resistor and then a 1K and a 100uf electrolytic to ground.

Will I need to adjust those components for parallel operation?

I don't have the schematic on this amp and like I say, it's setup for isolated stereo operation. (I'll trace the PCB when I have time -- but thought I'd ask).

Also, it's calling for a dual 28vac transformer for stereo -- what would be the right voltage for parallel mode?

Thanks.

 

[Minion]

A 2x28V AC tansformer is too much voltage , that will give you +/-42v DC which is the absolute maximum the chip can take but if the Voltage in your area happens to fluctuate 10% it will bring your voltage over the limit , you will also dissapate more heat so you will need a Bigger heatsink ....

I would suggest a 2x24v AC transformer for 8 ohm loads and a 2X18v AC for 4ohm loads (each chip) ....

For going paralell , I would try to match the Gain set resistors on each chip to .01% , I would also match the Output resistors to 1% and maybe use a 5w instead of 3w as a hotter resistor will tend to drift in value ....


Good luck

 

[pacificblue]

A 2x28V AC tansformer is too much voltage , that will give you +/-42v DC which is the absolute maximum the chip can take but if the Voltage in your area happens to fluctuate 10% it will bring your voltage over the limit , you will also dissapate more heat so you will need a Bigger heatsink ....

It is good to maintain a bit of headroom. On the other hand the IC withstands ±47 V at idle and the rails will sag under load. In AN-1192 the people from National chose 2x30 V transformers for 8 Ohm speakers. Of course they were big transformers, i. e. had only low regulation and heatsinking became an issue. 2x28 V AC should not pose a problem, as long as you keep the load at 8 or more Ohm for each IC, use big heatsinks and the non-isolated LM3886T package.

The isolated LM3886TF would indeed mean..

a 2x24v AC transformer for 8 ohm loads and a 2X18v AC for 4ohm loads (each chip) ....

In parallel application each IC sees only half the load. A 4 Ohm speaker will appear as an 8 Ohm load for each IC. An 8 Ohm speaker will appear as a 16 Ohm load for each IC. So you can take advantage of the IC's voltage rating as long as you don't connect 2 Ohm speakers.

For going paralell , I would try to match the Gain set resistors on each chip to .01% , I would also match the Output resistors to 1%

assuming that .01% should mean 0,1 % or better.

 

[jmar]

Ok --

So really running the LM3886 in parallel is really a no brainer.

4 or 8 ohm speakers are fine with a 24vac transformer and really I don't see any need to change the feedback resistors that they have on the board for normal stereo operation -- I see they are 1% -- I imagine they're matched well.

Just put the .1 ohm 5 watts in series with each output and that's it.

And parallel the inputs of course.

 

[pacificblue]

I don't see any need to change the feedback resistors that they have on the board for normal stereo operation -- I see they are 1% -- I imagine they're matched well.

The datasheet specifies 0,1 %. If you think, the National engineers did not specify that correctly, explain why. Don't recommend to ignore that specification based on your imagination only.

 

[abraxalito]

To get reasonable current sharing between two paralleled devices, the gain setting resistors matter a whole lot more than the current sharing ones. Those on the outputs can be 5% without problems.

As I see it you're likely to run into poor sharing at low frequencies with those component values (100uF and 1k). This is because of the relatively poor tolerances of aluminium electrolytics (20% typically). A 20% change in the 100uF cap gives of the order of a 0.5% change in the gain at 20Hz, so your relatively expensive 0.1% resistors are rather wasted. The solution is simple - up the 100uF to 1000uF.

Edit - I've just had a look at the AN-1192 and it seems they didn't get this one proof-read by Bob Pease. 0.1% resistors are most certainly not required on the + inputs to the amp chips. 1% output resistors are also overkill.

 

[pacificblue]

To get reasonable current sharing between two paralleled devices, the gain setting resistors matter a whole lot more than the current sharing ones. Those on the outputs can be 5% without problems.

A 20% change in the 100uF cap gives of the order of a 0.5% change in the gain at 20Hz, so your relatively expensive 0.1% resistors are rather wasted. The solution is simple - up the 100uF to 1000uF.

Edit - I've just had a look at the AN-1192 and it seems they didn't get this one proof-read by Bob Pease. 0.1% resistors are most certainly not required on the + inputs to the amp chips. 1% output resistors are also overkill.

More like 2 % at 20 Hz, but the resistors are still not wasted. There is only little musical content around 20 Hz. It is much more important to get the rest of the audible spectrum equally shared.

The DC offset should also be similar on both ICs. That is why the blocking resistors are also 0,1 % types. Looks like overkill, but maybe Bob Pease did proof-read and he knows something about the ICs we don't.

1000 µF in the gain loop makes little sense, if the DC blocking cap on the non-inverting input is not adjusted accordingly. And even then it is not really a practical value.

The 1 % output resistors must be judged with the power dissipation in mind. Paralleling chipamps only makes sense, if you want to take advantage of the amplifier's power limits. The nearer you get to the limits, the smaller an imbalance can be tolerated. If the rails stay well clear off the LM3886's rating, you can use bigger tolerance. With a 2x28 V transformer 1 % seems more recommendable, while 5 % is certainly sufficient for a 2x18 V transformer. But then again, with such a low voltage there is no need for paralleling.

 

[abraxalito]

More like 2 % at 20 Hz

Do show your calculations. I'm guessing they went something like this :

A 100uF cap has an impedance at 20Hz of about -j80 ohms.

At the tolerance extreme, this goes up to -j100 ohms.

This difference of j20 ohms is indeed j2% of the 1k resistor. So is that how you reached your answer?

but the resistors are still not wasted. There is only little musical content around 20 Hz.

It matters not one iota that the content at 20Hz be 'musical'. If this is being fed from a turntable, the rumble content won't be insignificant and will certainly extend well below 20Hz. If the OP listens to well recorded organ music on CD, there will be significant musical content. As an engineer I prefer not to legislate for the OP's musical tastes or choices of source. Besides, there will still be a diminishing gain error at frequencies higher than 20Hz. With your (miscalculated) 2% figure, there's still an error of 0.2% at 200Hz - are you going to say there's no musical content there too?

It is much more important to get the rest of the audible spectrum equally shared.

Do please show your reasoning.

The DC offset should also be similar on both ICs.

Have you looked at the LM3886 datasheet to understand what causes an offset at the outputs?

That is why the blocking resistors are also 0,1 % types.

Do please show your reasoning, bearing in mind that the input bias current for this chip shows up at 1uA maximum, whereas the offset voltage can go up to 10mV worst case.

Looks like overkill, but maybe Bob Pease did proof-read and he knows something about the ICs we don't.

An interesting hypothesis - what do you base it on? If he did indeed proof-read this AN, he's let various other howlers through along with these, some unrelated to the device characteristics.

1000 µF in the gain loop makes little sense, if the DC blocking cap on the non-inverting input is not adjusted accordingly.

I'm looking at fig 6 (page 8) and I see only one DC blocking cap, a 1uF mylar. Since that's in circuit for both paralleled amps, please explain how its going to cause a sharing problem.

And even then it is not really a practical value.

Do please explain why you think 1000uF is impractical.

The 1 % output resistors must be judged with the power dissipation in mind.

Here, I have no grasp of what you're trying to communicate. The tolerance of a resistor and its dissipation are orthogonal quantities.

Paralleling chipamps only makes sense, if you want to take advantage of the amplifier's power limits. The nearer you get to the limits, the smaller an imbalance can be tolerated. If the rails stay well clear off the LM3886's rating, you can use bigger tolerance.

If we already know the load then I might agree with you. But we don't - loudspeakers do vary. But you're missing the point here - the effects of the gain errors swamp the effects of having the sharing resistors closely matched. Do the math.

With a 2x28 V transformer 1 % seems more recommendable, while 5 % is certainly sufficient for a 2x18 V transformer. But then again, with such a low voltage there is no need for paralleling.

Depends what the load impedance presented by the loudspeaker is. Also it depends how low distortion the OP wants - looking at the datasheet, its clear that lower impedance loads make the distortion higher. So paralleling amps has benefits beyond higher powers.

 

[pacificblue]

Wow, what a lecture!

Do show your calculations. I'm guessing they went something like this :

100 µF ±20 % leads to worst case values of
80 µF at 20 Hz = j99,472 Ohm
120 µF at 20 Hz = j66,314 Ohm

Worst case for 0,1 % resistors is
Ci1 = j66,314 Ohm, Ri1 = 999 Ohm, -> Xi1 = 1001,199 Ohm, Rf1 = 20020 Ohm -> a = 20,996
Ci2 = j99,472 Ohm, Ri2 = 1001 Ohm, -> Xi2 = 1005,930 Ohm, Rf2 = 19980 Ohm -> a = 20,862
a average = 20,929, a error = ±0,32 %

Worst case for 1 % resistors is
Ci1 = j66,314 Ohm, Ri1 = 990 Ohm, -> Xi1 = 992,219 Ohm, Rf1 = 20020 Ohm -> a = 21,358
Ci2 = j99,472 Ohm, Ri2 = 1010 Ohm -> Xi2 = 1014,887 Ohm, Rf2 = 19800 Ohm -> a = 20,510
a average = 20,934, a error = ±2,03 %

there's still an error of 0.2% at 200Hz - are you going to say there's no musical content there too?

I am not getting you there. You think 0,2 % is a severe error, but the resistors need not be closely matched. Isn't that a contradiction?

Do please show your reasoning.

Your previous answer shows that you already followed my reasoning. I won't answer to nit-picking.

Have you looked at the LM3886 datasheet to understand what causes an offset at the outputs?

Should I? Are the mechanisms different from the ones in other amplifiers?

Do please show your reasoning, bearing in mind that the input bias current for this chip shows up at 1uA maximum, whereas the offset voltage can go up to 10mV worst case.

You should probably ask the designer or the datasheet author. I can only guess that they tried to match Ri to Rb in an attempt to improve CMRR and thus decrease DC offset.

An interesting hypothesis - what do you base it on?

What do you base yours upon that he didn't? Infallibility, because his name is well-known?

I'm looking at fig 6 (page 8) and I see only one DC blocking cap, a 1uF mylar. Since that's in circuit for both paralleled amps, please explain how its going to cause a sharing problem.

I don't know, how it is going to cause a sharing problem, and I didn't write that it does.

It makes little sense to filter the non-inverting input at 3,38 Hz and the inverting input at 0,16 Hz. Those two filters should have similar corner frequencies and add up to a reasonable quality factor. More than half an octave of separation between them should be an exception.

Do please explain why you think 1000uF is impractical.

It has no advantages that justify its size and cost. You will find few amplifiers with a 1 mF capacitor in the inverting leg, if any. Are you prepared to accept the accumulated experience of the majority of amplifier designer's as proof?

The 1 % output resistors must be judged with the power dissipation in mind.

Here, I have no grasp of what you're trying to communicate. The tolerance of a resistor and its dissipation are orthogonal quantities.

Yes, sorry, that sentence can be misleading. I was refering to the power dissipation in the chipamp, not the resistors. Based on my assumption that paralleling is used to make the most of the ICs' power output, they (the ICs) would be used next to their limits.

the effects of the gain errors swamp the effects of having the sharing resistors closely matched. Do the math.

I also refer to those resistors as load-sharing resistors out of habit. The truth is probably more that those resistors act as buffers between the IC outputs, so that any imbalances can be converted into heat in them, instead of the ICs' output stages. Close matching them is necessary, when you take the ICs to their voltage limits as I explained before.

Depends what the load impedance presented by the loudspeaker is.

a 4 ohm speaker

 

[abraxalito]

Wow, what a lecture!

Not really, since a lecture is generally unidirectional. Here I'm interacting with your remarks so its more of a tutorial. But I agree its rather long - that's because you've made so many mistakes. So here's what I'll do - I'll just deal with one group of your mistakes at a time to keep the post to a manageable length. So for now, I'll just deal with the calculations for the gain matching of the two amps.

100 µF ±20 % leads to worst case values of
80 µF at 20 Hz = j99,472 Ohm
120 µF at 20 Hz = j66,314 Ohm

No disagreement there.

Worst case for 0,1 % resistors is
Ci1 = j66,314 Ohm, Ri1 = 999 Ohm, -> Xi1 = 1001,199 Ohm, Rf1 = 20020 Ohm -> a = 20,996
Ci2 = j99,472 Ohm, Ri2 = 1001 Ohm, -> Xi2 = 1005,930 Ohm, Rf2 = 19980 Ohm -> a = 20,862
a average = 20,929, a error = ±0,32 %

By which you show that the capacitor tolerance on the 100uF has moved the gain matching error well beyond that for the resistor tolerance. Which was my point. If you do this again but use 1000uF you'll find the error reduces considerably. To save you the trouble, I'll do it myself :

C1a = j6.6 ohm, R1 = 999 ohm; X1 = 999.02 ohm, Rf1 = 20020 ohm; a = 21.04
C1a = j9.9 ohm, R1 = 1001 ohm; X1 = 1001.05 ohm, Rf1 = 19980 ohm; a = 20.96
average a = 21, error in a ±0.04 giving ±0.2%

This is the same (to the number of significant figures in the calculation) as having no capacitor at all, so the value of the relatively expensive 0.1% resistors hasn't been thrown away.

Worst case for 1 % resistors is
Ci1 = j66,314 Ohm, Ri1 = 990 Ohm, -> Xi1 = 992,219 Ohm, Rf1 = 20020 Ohm -> a = 21,358
Ci2 = j99,472 Ohm, Ri2 = 1010 Ohm -> Xi2 = 1014,887 Ohm, Rf2 = 19800 Ohm -> a = 20,510
a average = 20,934, a error = ±2,03 %

You offered your 2% figure to correct my initial figure of 0.2%. Now I see your comment of 'more like 2%' is based on gain setting resistors with 1% tolerance. However, I've never suggested the gain setting resistors be 1% types - in my judgement they should indeed be 0.1%. I agree that with 1% resistors, the match is unaffected by capacitor tolerance. So it seems you misunderstood my original comment and took it out of context - that context was already accepting the need for 0.1% resistor match in the gain setting resistors. Actually my 0.2% was based on conservative hand waving, I accept your more accurate figure above of 0.32% - which only makes my thesis stronger and yours weaker.

I am not getting you there. You think 0,2 % is a severe error, but the resistors need not be closely matched. Isn't that a contradiction?

Please point out where I've said that the gain setting resistors need not be closely matched? As I've set out above, they should indeed be matched to 0.1% but that degree of match is rather wasted with a too small value of capacitor.

 

[abraxalito]

Second episode of the long-running tutorial.....

Should I? Are the mechanisms different from the ones in other amplifiers?

Perhaps I'm misunderstanding you here - you said the offsets should be similar on both ICs. So you could have been meaning one of two things - that its a design goal to make them similar, or that the implementation we're discussing (I take it from AN-1192) will result in similar offsets. I'm though wondering why offsets are a problem here, given that they'll be of the order of low tens of mV worst case whereas gain errors will result in greater imbalance when using 0.1% resistors and a 100uF cap to determine gain.

You should probably ask the designer or the datasheet author. I can only guess that they tried to match Ri to Rb in an attempt to improve CMRR and thus decrease DC offset.

I can't see how 'improving CMRR' makes any difference to DC offset - they're quite unrelated in this particular configuration. I wouldn't dream of asking the appnote authors, its obvious to me they're a bit out of their depth here.

What do you base yours upon that he didn't? Infallibility, because his name is well-known?

No, because he's a highly competent, experienced analogue engineer and would be very unlikely to make the really basic errors these writers have made. So, back to my original question - what's your hypothesis based on?

I don't know, how it is going to cause a sharing problem, and I didn't write that it does.

It makes little sense to filter the non-inverting input at 3,38 Hz and the inverting input at 0,16 Hz. Those two filters should have similar corner frequencies and add up to a reasonable quality factor. More than half an octave of separation between them should be an exception.

I'm totally lost as to the rationale behind these remarks - to me it looks like some dogma having no bearing on reasonable design. In the world of electronics, 'quality factor' (or 'Q') means something totally different. So what meaning have you ascribed to the term here?

It has no advantages that justify its size and cost.

We're speaking here of the 1000uF capacitor - I've already set out the advantages of it - improving the low frequency current sharing between paralleled amplifiers. Its physical size can be fairly small as the working voltage isn't great. Alternatively if you object to the size and cost, there's always the option of scaling up the values of the gain setting resistors. Increasing the 20k to 100k would mean a five fold reduction in this capacitor for example so a 220uF would do fine. Or you could go the way of the other designs in the appnote and use servos - bear in mind though they're considerably more complex and costly than the bigger cap.

You will find few amplifiers with a 1 mF capacitor in the inverting leg, if any.

That might be for varied reasons - I've given one above - the gain setting resistors might be chosen higher. Another reason might well be that the amp uses a servo rather than a big capacitor - electrolytic caps aren't by and large considered high-end components and plenty of designers prefer to avoid them all together. In any case there's no need for such a large value when amps aren't being paralleled. I think you'll find few commercial amplifiers out there based on paralleled chip amps.

Are you prepared to accept the accumulated experience of the majority of amplifier designer's as proof?

Not as proof, no, but certainly as strong evidence. However your statement above doesn't follow logically from your earlier claim since I've explained a few reasons why commercial amps may well not have 1mF caps. Nevertheless, since you've introduced evidence of amplifier designers into the conversation, have you heard of Doug Self ? He sometimes contributes here and he's written quite a lot on amplifier design. I happen to have a copy of his book 'Audio Power Amplifier Design Handbook' open in front of me. On page 220 of this tome, he presents a fully discrete design ('Trimodal Amplifier') which does indeed use a 1mF capacitor in this very position. Which amplifier designers are you going to cite who disagree with Mr Self ?

I also refer to those resistors as load-sharing resistors out of habit. The truth is probably more that those resistors act as buffers between the IC outputs, so that any imbalances can be converted into heat in them, instead of the ICs' output stages. Close matching them is necessary, when you take the ICs to their voltage limits as I explained before.

But you've not explained at all why close matching the sharing resistors is necessary. If you did the math, you'd find out why its not. In this case, I'm not going to do it for you.

 

[pacificblue]

Not really, since a lecture is generally unidirectional. Here I'm interacting with your remarks so its more of a tutorial.

Seems I was at least right about the nit-picking part.
I am not sure, whether I want to spend a lot of time with your 'tutorial'. I expect that in the end we will rather come to realize that we interpreted each other wrongly, than really learning something new.

An example for what I mean is your following contribution...

A 20% change in the 100uF cap gives of the order of a 0.5% change in the gain at 20Hz, so your relatively expensive 0.1% resistors are rather wasted.

... which led me to the conclusion that you find the use of 0,1 % resistors a waste, while you wanted to convey that you find the use of a 100 µF capacitor inadequate.

This is the same (to the number of significant figures in the calculation) as having no capacitor at all, so the value of the relatively expensive 0.1% resistors hasn't been thrown away.

What makes you think a 0,32 % gain error means throwing away the value of the relatively expensive 0,1 % resistors? Compared to the ~ 2 % error with 1 % resistors the improvement is significant. We can only speculate about how and why the engineers at National came to the conclusion that they did not need a gain error as small as 0,2 %.

I accept your more accurate figure above of 0.32% - which only makes my thesis stronger and yours weaker.

I don't even remember setting up a thesis. And calling anything in your posts a thesis is strongly overstating its importance.

I wouldn't dream of asking the appnote authors, its obvious to me they're a bit out of their depth here.

Now that is a bold statement. If you think, you can do better than they, go ahead. You will find many eager readers here. Although you lost my attention completely by that. Save your time on completing your 'tutorial', because I won't read or answer it any further.

 

[abraxalito]

Seems I was at least right about the nit-picking part.
I am not sure, whether I want to spend a lot of time with your 'tutorial'.

You're not the only person here - if there are others building paralleled chip amps, its possible they might find my contributions helpful. As for the 'nit-picking' claim, you've no evidence that I'm nit-picking - this is attention to important design details.

I expect that in the end we will rather come to realize that we interpreted each other wrongly, than really learning something new.

What I've come to realize so far is you've interpreted me wrongly. Another thing I've come to realize is you're not interested in admitting to many of your mistakes. Never mind, there are others here who may learn from them.

An example for what I mean is your following contribution... ... which led me to the conclusion that you find the use of 0,1 % resistors a waste, while you wanted to convey that you find the use of a 100 µF capacitor inadequate.

Yep, its obvious you made an error of interpretation.

What makes you think a 0,32 % gain error means throwing away the value of the relatively expensive 0,1 % resistors? Compared to the ~ 2 % error with 1 % resistors the improvement is significant.

I see you're still confused, even though I've explained a couple of times so far. Well, maybe third time lucky! I've never considered 1% resistors adequate to the task here - the sharing will be way too poor if they're used. So I agree with you, but bringing 1% resistors in here is a red herring. What you don't seem to realise is that the 0.32% error is due to the capacitor primarily, by increasing its value we reduce that figure.

We can only speculate about how and why the engineers at National came to the conclusion that they did not need a gain error as small as 0,2 %.

You may continue to speculate, I have no need to because its obvious - they're not competent in this particular appnote. If speculation's your thing, here's one : perhaps they were a couple of students who National gave a summer job to. If you look at that appnote again, you'll see they didn't specify a 100uF capacitor as you're suggesting, their circuit has 68uF which makes the matching worse than 0.32% at 20Hz.

Now that is a bold statement. If you think, you can do better than they, go ahead. You will find many eager readers here.

I can do better, as I've been demonstrating these past couple of posts. I'll continue to explain to any interested readers here exactly how that appnote is flawed and how they can fix it.

Although you lost my attention completely by that. Save your time on completing your 'tutorial', because I won't read or answer it any further.

Again you seem to be under the bizarre illusion that I'm only doing this for your benefit. That's kinda quaint Your loss that you're not paying attention, not mine.

 

[pacificblue]

Do show your calculations. I'm guessing they went something like this :

Do please show your reasoning.

Have you looked at the LM3886 datasheet to understand what causes an offset at the outputs?

Do please show your reasoning, bearing in mind

If he did indeed proof-read this AN, he's let various other howlers through along with these,

Do please explain

But you're missing the point here

Do the math.

I've just had a look at the AN-1192 and it seems they didn't get this one proof-read

you've interpreted me wrongly.

Yep, its obvious you made an error of interpretation.

I see you're still confused,

You may continue to speculate, I have no need to because its obvious - they're not competent in this particular appnote.

I can do better, as I've been demonstrating these past couple of posts. I'll continue to explain to any interested readers here exactly how that appnote is flawed and how they can fix it.

Your loss that you're not paying attention, not mine.

I am probably also wrong in assuming that this is a forum? Seems to be abraxalito's ego-shooter instead.

 

[pacificblue]

perhaps they were a couple of students who National gave a summer job to.

When you look at the top of AN-1192, you can read that the authors are John DeCelles and Troy Huebner.

Joh deCelles was already the author of AN-898 in 1993, so it is pretty unlikely that he was still a student on a holiday job, when he wrote AN-1192 in 2004.

From: All About Class D Audio Amplifiers

Troy Huebner has been with National's Audio Group since 1996 originally working as a product engineer but switching to applications in 1998. His main area has been the high power class D and Overture amplifiers. Troy obtained a BSEE degree from Brigham Young University, Provo, Utah (1996).

Obviously the co-author was also not a student on a summer job anymore by 2004.

 

[AndrewT]

can you two bring your combined knowledge together and propose what must be done with a parallel chipamp design to make it work reliably?
In particular detail any differences from the application note, an1192.

 

[abraxalito]

When you look at the top of AN-1192, you can read that the authors are John DeCelles and Troy Huebner.

Yep, I already did that research. That's what hypotheses are for, to be shot down. Now, how about doing a similar exercise for your hypothesis that it was indeed proof read by Bob Pease ?

 

[abraxalito]

can you two bring your combined knowledge together and propose what must be done with a parallel chipamp design to make it work reliably?
In particular detail any differences from the application note, an1192.

Sure, any particular implementation shown in that appnote?

Some of my comments just apply to how the components have been over-specified in various places - the tolerances on the output current sharing resistors and on the various input resistors. Correcting those won't affect the reliability, just improve the performance per dollar spent.

However, the issue of the capacitor tolerances affecting the gain matching does affect the reliability and performance. I'll post up something more detailed on the simplest paralleled schematic in that appnote soon.

 

[pacificblue]

what must be done with a parallel chipamp design to make it work reliably?

From AN-1192:

The objective is to provide simple high power solutions that are conservatively designed, highly reliable and have low part count.

For a beginner or a one-time-builder, who does not have sufficient knowledge to draw his own conclusions, I keep up my recommendation to stick to the datasheet and rather take a too conservative approach.

Here is another quote from AN-1192:

1% gain setting resistors (Ri and Rf) will give good results but it is recommended 0.1% tolerance resistors be used for setting the gain of each op amp for closer matched gain and equal output current and power dissipation.

Once the gains are matched so that both ICs deliver virtually the same output voltage swing, the load sharing resistors come into play. The tighter the heatsink and the higher the rail voltage, the tighter should the tolerances of those resistors be.

In the end each designer has to decide for himself, whether his amplifier can tolerate a current difference of ±5 % or if he needs those hard to source ±1 % power resistors. If he cannot assess that, he should take the safer approach or be prepared to spend time and money for repairs, re-designs and/or re-building.

 

[AndrewT]

A stereo pair of parallel chipamps requires a minimum of 4 output resistors.
Buy 6 (two go into stock) and series connect them using a 317 CCS passing 500mA. Set your DMM to 200.0mVdc and measure the voltage drop (~50.0mV) across each resistor.
Pick a pair that have as near as possible the same voltage drop. Now pick a second pair.
That will give you two pairs of near +-1% tolerance 0r1 resistors.

Alternatively, buy 40off 1r0 1% 600mW metal film. Just parallel up 10 of each to give a 6W +-1% 0r1 resistor. The simple act of paralleling 10off will statistically give <0.5% variation.

That sort of useful information can be used by any competent builder.

If he/she can't do that then they should not be building a parallel chipamp.