Since i have no webspace available at the moment i try to
describe the circuit:

The concept is to build the A47 circuit but wiring the output
of the "first" opamp to the input (unity gain configuration).
Let the non-inverting input of the "first" opamp sit at half the
battery voltage by using a resistor voltage divider.
The A47 output now is the virtual ground.

Question: I have still to use the output resistors, right ?
Shall i keep the value ?

The only difference is that you're paralleling two op-amps on the virtual ground instead of using just one. Yes, you'd need to add output sharing resistors.

What you do is use the opamp in a current doubling configuration like 47 amp for more current drive to the circuit

i will see if i can dig up the link to the original

I thought about the exactly the same thing while I was trying to increase the output current for the virtual ground.

I dropped the idea without even testing it. As Tangent said, you still need the output resistors. Based on my experience most opamp's ouput impedance is already too high, no need to mention those open loop buffers. Adding any output resistor is simply non-acceptable.

[Edited by Fixup on 04-08-2004 at 12:50 AM.]

Thanks rickcr42 for assuring that this is not new.
So if it´s not new there is a chance it may work.

Fixup,
thanks for the comment.
If output resistors are not acceptable: What other concept
shall i use to have a simple, cheap rail splitter.
A TLE chip is too weak for my purpose and I do not want to use buffers after a TLE because buffers are expensive.

Shall i follow sijosaes concept he showed a while ago ?

Just use the same circuit in the picture posted by Tangent, but do these:

Use 10K instead of 22K, if your power is 5V or lower.
remove the 10 ohm resistor.
You may need a 1K-10K resistor in the feedback for some opamps.

I don't know much about 741 and I don't know if it's good for this case. I have used LMH6642 for quite a while and it does the job very well. Now I use something even newer and better. I cannot say which one is the best, but I can say the followings should be avoided:

AD8531
BUF634
TLV4110

These seem ideal for their very high load capbility, there is even a sample splitter circuit in BUF634 datasheet, but they are craps if used for splitters in headphone amps. TLE2426 should not be used neither, because it's too weak and its output impedance is too high above 10kHz.

The plan posted by Tangent has output impedance below 0.1Ω from DC to about 400KHz (depending what type caps used and if you add small ceramic bypass caps). More like 0.01Ω over much of the audio band. With or without the 10Ω resistor (though I'm sure it is not necessary except perhaps to make a tutorial point).

Wiring resistance is much more significant than impedance of this level.

If you think you need lower, and do not listen to DC, increase the cap size.

I would argue that a plan like Tangent shows, while it often works, is conceptually confused. But I won't argue that tonight; I need to get to bed.

PRR, maybe you can find the time and explain why tangents plan
is "conceptually confused" ?

[Edited by sijosae on 04-13-2004 at 07:38 PM.]

---

quote:

conceptually confused

---

I agree that it's better to use a high output current device here, but at the point in the article where that schematic comes from, I'm taking on the case where the person can't get a TLE2426. If they can't get a TLE, they probably can't find {insert exotic op-amp here}, either.

My purpose definitely wasn't to show the ideal rail splitter.

[Edited by tangent on 04-13-2004 at 10:12 PM.]

Don't have a cow. I'm not picking on you. Or that 741. I kinda like 741s, in their place.

The question is: when you have a big cap on the output of an amplifier, who is in charge? The cap or the amplifier?

For the medium power devices we are likely to use, with anything like the suggested 440µFd hanging on its output, over most of the audio band, the cap is in charge and the amplifier isn't important to the sound.

But also, as shown, the amplifier is not going to be happy. The 440µFd load is nearly a dead-short over most of the audio band. In this case, and for most reasonably similar cases, the op-amp's gain essentially vanishes above about 40Hz due to the heavy capacitive loading.

In this specific case, it would be mildly happier if the feedback were taken from the other end of the 10Ω resistor. That increases the DC impedance. And the op-amp is still very heavily loaded. It would start to work like an op-amp if the resistor were changed to say 100Ω, but that would reduce the DC balance.

And here's where the nub of the problem lays. Why are elegant designs like Cmoy's so prone to showing large DC error? Why are we talking about ultra-beefy rail-splitters that pit caps against chips in a fight that I am not sure we can win?

Look at the Cmoy from the DC point of view:

I've used new part-numbers, so refer to this drawing instead of Cmoy's. Ignore the 1G resistor, that was just to keep SPICE from puking. The 1mV "battery" is a realistic offset-voltage for a chip-amp. In Cmoy's defense, he did not have a 32Ω load in mind, and the problem is far less with higher loads. But people who use 32Ω loads do have big DC offset. Why?

I'm sure Tangent and FA will follow this explanation. The typical 1mV offset, when fed-back, forces 1mV across R7. The ratio R7/R6 forces 10mV across R7. We have 11mV across R6+R7.

That also means 11mV across Rload. If Rload is 32Ω, then 0.011V/32Ω= 0.34mA DC flows through Rload. The only place this can come from is Rsplitter, which is in effect two 10K resistors in parallel (via the power battery) or 5K. 0.34mA times 5KΩ is 1.7V DC.

The point that should be "half-way" between battery terminals is 1.7V away from half-way!

Somehow a 1mV error has been multiplied 1,732:1 !

The audio (and DC) gain of the op-amp and R6-R7 is 11:1.

But Rload and Rsplitter create a second gain-set loop. If Rload is 32Ω and Rsplitter is 5KΩ, then this ratio is 32/5000= 156:1. This multiplies the basic 11:1 gain. 156*11= 1,719.

Any DC error in the op-amp becomes 1,700 time larger at the "halfway" point.

Note that if Rload is 300Ω, as Cmoy intended, the multiplication is only 194:1. A 1mV op-amp error becomes a fairly mild 200mV error at the halfway point.

In either case, the DC across the headphones (for a single channel) is 11*Vo where Vo is offset voltage. Typically 11mV. (However it does get messy when we have two channels on one jack and one splitter.) The headphones won't fry. Actually it works well, except headroom is reduced as "halfway" error rises. 1.7V error here is 1.7V less peak output level on one side of the wave. Since music is statistically symmetrical, this is like 3.4V less peak-to-peak level.

The answer is a lower splitter impedance. Using a passive splitter, we could go to a pair of 1K resistors. Rsplitter is then 500Ω, halfway error tends to be more like 0.17V, which is no big deal on 9V supply. However two 1K resistors is a 4.5mA drain, which is more than some chips, and liable to impact battery life.

I think in many cases, a pair of 2.2K resistors bypassed with BIG caps is a very happy solution.

But how big? What does it matter what the output impedance is?

If the splitter impedance is "dirty", such as an amplifier with crossover distortion, that will get into the audio. And for impedances like ours, we need much-much more than 1mA transistor current to eliminate crossover (feedback masks it but IMHO not very well). Also any amplifier doubles your power consumption because it lacks energy-averaging/storage over the audio cycle. A biased electrolytic cap is actually a pretty clean low-Z device and very efficient.

If the impedance is clean, how low does it have to be? For less than 1dB power loss it should be under 1/10th of load impedance, say 3 ohms. But with common-ground stereo any impedance causes crosstalk. A high-spec 40dB at 20Hz suggests 25,000µFd! However perception and recording practice tends to eliminate stereo in the bass. 26dB separation at 100Hz gives 1,000µFd, a very reasonable value.

With two 2.2K resistors and two 470µFd caps, there is no doubt "who is in charge". There is no active device straining to fight a big storage capacitor. It is conceptually elegant.

[Edited by PRR on 04-15-2004 at 02:31 PM.]

---

quote:

I'm sure Tangent and FA will follow this explanation

---

Damn straight, cos I'm not sure I do - see, now I wish I had paid attention in class...but I am trying to understand all this, so bear with me

I tried that Tonmeister link Sanaka posted, and then re-read your post. And then read it again.

From what you are saying, it sounds as if including what Cmoy refers to as R5 (say a 10k) would lower the voltage across Rsplitter to 5mV, since Rload would effectively be 10,032R.

I have a horrible feeling I am wrong, but nothing ventured...

EDIT: ...nothing gained (little pun there ). I am wrong. I thought R5 was 50k, but it is, of course, 50R. Using that would give a DC offset of 0.67V.

Cheers,

Rexel

[Edited by Rexel on 04-16-2004 at 02:22 AM.]

And now i am going to use an opamp to work on big electrolytics.
I could kick my behind because i overlooked this.

But what to do ?
Could it be an advantage to stay with
the RC railsplitter ?

You say 2.2kOhm and 470µF.
This gives an extra load of 2mA to a 9VDC battery which is acceptable.
Can we use two low-current LED with two series resistors instead
of two simple resistors ?
This could be nice because we hit two points:
Rail splitting plus power-on signal.

What do you think ?

One question, however: What would be the effect of a suitably sized cap in series with the load resistor? It seems to me that this would let you shut down the DC load current which is the real source of your problems, while allowing audio signals to pass through. Are the side effects of a cap in series with your load worse than the original offset problem?

For DC, yes. For Audio, it is the value of the capacitor(s). Since these are typically 470uFd, the node impedance is under 20Ω at 20Hz, around 1Ω at 400Hz. Note that even 1Ω impedance here is -30dB leakage between channels.

> including what Cmoy refers to as R5 (say a 10k)

A series resistor inside the feedback loop does nothing for offset. Or 32Ω power, except that it may let the chip run a bit cooler by taking some of the excess supply voltage as heat in the resistor.

Putting the series resistor outside the feedback loop will reduce offset gain, but the value required for significant improvement is much larger than low-Z headphone impedance so you lose all damping. It is not clear that all phones need damping, but we usually do drive them from a fairly low impedance, not an impedance higher than the phones.

And 10K in series with the load will just kill output. Fooling with 50-100Ω resistors may make a bad situation less-bad, but are not a fix, or even an elegant hack.

> What would be the effect of a suitably sized cap in series with the load resistor?

That is certainly my preference; I maintain that it solves all the problems without any adding great flaws. And certainly a lot of recordings were made on gear with cascades of interstage caps.

Anyway, the splitter-caps ARE in series with the load, unless your splitter buffer is stronger than the caps (which also means more than twice as strong as the audio stages).

But output caps have become very unfashionable.

[Edited by PRR on 04-19-2004 at 06:21 PM.]

sijosae,
I think PRR answered you very well.

PRR,
Very good reasoning. You answered some of the questions I had in my mind for a long time. As I said before, this virtual ground is much more complicated than it appears. There are at least three aspects (PRR mentioned them all):

1) Load impedance. For 300 ohm headphones, anything or just two resistors will be just fine. But, 32 ohm headphones cause all the troubles.

2) DC offset. Big caps does not help in this regards. An opamp buffer or TLE2426 solves this issue elegantly, as long as you don't use any cap at its output (yes, fa-schmidt). In my earlier amps which did not have opamp buffers, I had to use two 1K resistors as the spliter, to keep the DC offset low enough for 32 ohm headphones (see my old posts). I did not come up with the "1K" value via calculation like PRR just did; I chose it after lots of tests, trading off between low DC offset and low battery drain.

3) Channel crosstalk. Again, for 300 ohm loads, this is no issue. But for 32 ohm loads, The difference between 0.1 ohm and 0.01 ohm impedance makes a HUGE differnece on crosstalk, no need to mention 3 ohm or so. Although RMAA crosstalk numbers do not look so good, real-life listening is still okay as PRR said.

Like PRR said, output caps have become unfashionable. But, virtual ground needs lots of careful treatments. Although I found the related problems as soon as I built my first CMoy and I did figure out my own solutions, PRR is the first one who addressed these issues theoratically very well. Did you guys notice that the Airhead still used output caps (see the pictures on Headroom's website)? My latest solution on this issue takes the best parts of the buffered splitter and PPA.

See, how many details are behind such a simple virtual ground circuit! Thank PRR for providing such a good example on how to solve a practical problem using theorem and expertise. I usually can solve a problem with a good solution, but cannot explain it so well.

I'm not sure I understand your assertion that the caps on the splitter are in series with the load, unless you're talking about a non-DC signal. True, with the proper sizing mentioned the caps will be doing all the driving for signals in the audio band, essentially replacing Rsplitter in the diagram, but the caps shouldn't affect the DC offset voltage which was the original problem. Is there something I'm missing?

As far as the sizing for the 470µFd caps goes, are there any disadvantages to going larger aside from space/cost constraints? I'm assuming from the discussion that this is a suggested minimum, but if I can find some larger ones that fit in my enclosure would they cause problems?

[Edited by Whillowhim on 04-23-2004 at 03:02 AM.]

I tried again with 680uF caps, then 4 680uF caps, then 4 680 uF caps with 2 1uF caps (trying to mimic the Pimeta). The voltage readings did not waver.

Just for comparison, with 2 4K7 resistors and 2 470uF caps I was getting 1.5-1.7V offset.

I also tried the TLE2426 in place of the voltage dividing resistors, using all of the above cap variations and the same cans. Almost, but not quite as good. The voltages rails were rock solid at usual listening volumes (for me) but started to wobble slightly at higher volumes. I could not, however, actually hear any difference.

Cheers,

Rexel

EDIT: I got my readings mixed up and duplicated the readings going into the headphones. The rails differed by 100mV at pins 4 and 8, not 5mV

[Edited by Rexel on 04-23-2004 at 06:32 AM.]

Result:
Rails differ 300mV. DC offset below 10mV (multimeter precision doubted).