amplifier's power supply topology

[abraxalito]

Well if the DMM is an average cheap one, it won't give you true RMS, just average but calibrated to RMS on a sinewave. Are these measurements with a sine wave or using music?

[sangram]

Square, actually, but what I do is I set the level to what the maximum comfortable listening level is, then retain the same volume setting and run a signal gen into the amp input. The source is a PC, so it's pretty easy to do this without changing anything.

Maybe building a small attenuator and make a PC oscilloscope is a good idea now that you mention it.

[abraxalito]

Originally Posted by sangram

Maybe building a small attenuator and make a PC oscilloscope is a good idea now that you mention it.

Sounds like a good idea. I don't have a PC handy near my amps so I recently got a Sony PCM-10 flash-card recorder and use this (with a 200:1 input attenuator) if I need to check distortion by looking at the recorded file in Audacity.

[twest820]

Originally Posted by abraxalito

I've never seen such high ripple on the power supply in any amp I've designed.

I'd assumed a 4 ohm load and 50Hz mains with the usual small signal ripple model for a full wave rectifier. From there applying Ohm's law and a little math gives you

Vripple / Vsignal = 1 / (2 f C Zspeaker)

where f is the mains frequency, C the reservoir capacitance supplying the output signal, and Zspeaker the impedance presented by the speaker+and passive crossover+whatever in a power spectral density weighted sense. With this model the ripple on 8 ohms at 60Hz should be around 20dB below signal, though your case is probably coming up against the limits of the small signal model's validity range.

The above assumes a constant current draw over the discharge and using peak to peak ripple so it's a bit conservative; adjust as needed for sinusoidal current draws, different sorts of transients, or different definitions of ripple. I'm guessing you're using sinusiodal draws which are symmetric about 0A. In which case the measured ripple'd be more like 23dB down. Might be another +/-3dB in there depending on how exactly our capacitance accounting lines up.

Originally Posted by abraxalito

How does PSRR translate into IMD?

I think this gets into the mental model one uses to think about what's going on. We both agree the supply caps produce ripple that's a down modulation of the signal and the amp ends up with an output error that's the ripple attenuated by PSRR. Seems to me it's equally valid to think of the output error as additive noise or as a modulation of the input signal.

I should go back and check the PSRR measurement conditions on the 3886; it's possible the supply induced output error's attenuated only by PSRR and not excess loop gain, in which case I've overestimated the part's performance.

Originally Posted by abraxalito

my experience of that is that its caused by poor grounding and decoupling, not by interchannel crosstalk

Do you have data on what the noise levels introduced by the grounding problems were? The 3886s I've listened to and measured haven't been particularly great either subjectively or objectively---roughly 0.5% THD at typical home audio tweeter levels in biamp, input stage limited. It's certainly possible the ground problems were above that level and interchannel crosstalk through the supplies is below that level, in which case performance is not supply limited and it'd make more sense to invest effort in changing to a higher fidelity part.

Originally Posted by sangram

I rarely see anything above 14V before the sound starts to crack up real bad. Mostly 11-12V, though most of real listening its 7-8V. I assume these are RMS figures as they come off a DMM. I'm just not sure it can actually manage 20VRMS output cleanly.

How much bypass capacitance, what mains frequency, which chipamp (in my limited experience the 3886's somewhat unsual in its ability to drive close to the rails), and is the test being done within the DMM's passband? Posting the schematic wouldn't be a bad idea.

[sangram]

twest: I use the reference schematic, with 22K/1K feedback resistors, a 470uF Ci cap, DC-coupled input and no zobel (tried, makes no difference).

PSU has 10,000uF per rail per chip, separate rectifiers and separate cap banks but a single 48VCT transformer.

I use a 1KHz square wave, so I guess it should be within the DMM's range.

Mains here is 230V/50Hz, give or take about 20% tolerance depending on the power company.

[abraxalito]

Originally Posted by twest820

I think this gets into the mental model one uses to think about what's going on. We both agree the supply caps produce ripple that's a down modulation of the signal and the amp ends up with an output error that's the ripple attenuated by PSRR. Seems to me it's equally valid to think of the output error as additive noise or as a modulation of the input signal.

To me, addition of noise isn't really the same as modulation of the signal. But I'm not sure that the difference is important here, we both agree the noise is unwanted. The result of modulation though is unwanted signals at different frequencies.

Do you have data on what the noise levels introduced by the grounding problems were?

No, I don't have equipment designed to measure that stuff. But even if I had an AP, I'm not sure how I'd go about measuring it. I did try with just a PC soundcard and didn't see anything different in the two cases.

The 3886s I've listened to and measured haven't been particularly great either subjectively or objectively---roughly 0.5% THD at typical home audio tweeter levels in biamp, input stage limited.

You're saying the LM3886's input stage is causing 0.5% distortion? Sounds to me like you've got a serious circuit problem or a half-dead chip.

[twest820]

Originally Posted by abraxalito

You're saying the LM3886's input stage is causing 0.5% distortion? Sounds to me like you've got a serious circuit problem or a half-dead chip.

Naw, it's right in line with what you'd expect from the datasheet. Typical output levels on a biamp's tweeter channel for home audio are around a hundred microwatts---with 90dB efficient drivers the total power per speaker is a few milliwatts to deliver the slightly loud 65dBish average SPL many folks tend to listen at---and the tweeter power tends to be 10-15dB down from the total channel power. The THD+N curves in the 3886 datasheet are in good agreement with Douglas Self's rule of thumb crossover distortion triples for every order of magnitude reduction in power. If you project past the datasheet's 10mW limit the results are in good agreement with the measurements I mentioned.

Also, if you look at the equivalent circuit of the 3886 you'll see it's not actually an op amp and hence has a few limitations around the input. tomchr and I found if the part's operated class XD there's no change in distortion, which confirms the problem is a control loop limitation and not crossover distortion in the output devices.

[abraxalito]

Originally Posted by twest820

The THD+N curves in the 3886 datasheet

Those are THD +N curves - to me, they look to be dominated by noise, not distortion.

are in good agreement with Douglas Self's rule of thumb crossover distortion triples for every order of magnitude reduction in power.

I've read Doug Self's earlier book and this rule of thumb did not ring any bells. So I've had a peek in his latest edition, and found this:

reducing the output power from 25 W to 250 mW, which is pretty drastic, only increases THD percentage by six times

That's on pp178,9 of his 5th edition. So there's a two orders of magnitude reduction only giving six times. So we'd expect a bit under 2.5X, per order of magnitude, not quite as much as tripling. He also mentions a bit earlier in the chapter that at lower levels, the harmonic structure tends to be skewed away from the higher order ones so at lower levels the distortion becomes more and more benign.

The thing to remember here is that Self's measurements were done on a deliberately underbiassed amp to maximise the distortion and one with considerably lower noise than a LM3886 with its heavy input LTP degeneration.

<edit> I forgot to comment on your earlier remark that its the input stage which is causing the distortion. Clearly not, any crossover distortion will be caused in the output stage. Self says those measurements made above only became meaningful once other distortion mechanisms present had been sufficiently minimized.

[twest820]

Originally Posted by abraxalito

Those are THD +N curves - to me, they look to be dominated by noise, not distortion.

Originally Posted by abraxalito

any crossover distortion will be caused in the output stage

Neither of these is in agreement with the measurement data I have. Do you have data which shows otherwise?

[abraxalito]

Originally Posted by twest820

Neither of these is in agreement with the measurement data I have.

Post it up by all means. If you have measurement data which shows input stage crossover distortion that would be really something.

Do you have data which shows otherwise?

No, because I'm happy to rely on National's measurement data in their datasheet for the LM3886. Let's look at my earlier claim that at low level the THD +N figures are dominated by noise, given all that I've so far introduced.

I'll take the page 12 top right graph of 1kHz THD+N into 8R. Here we see the distortion plus noise figure touch 0.002% just before clipping, which is shown as being 39W. Let's assume initially that this figure is distortion dominated.

Applying Doug Self's six times more distortion for a hundred times less power rule once, we reach 0.012% at 390mW. Apply it again and we obtain 0.072% for 3.9mW. The graph reads 0.035% for 10mW. Extrapolating the graph to 3.9mW we won't get double the figure as the slope of the graph isn't steep enough at 10mW. At 40mW it shows 0.017% - multiply this by 2.5 and we get a figure short of 0.072%. So I conclude its noise dominated here even if it was distortion dominated just before clipping - which is debatable.

The story is rather different at 20kHz though, I'm not claiming that the graphs are noise limited there. Were you only referring to the 20kHz plots when you claimed the input stage was causing 0.5% distortion?