Power Supply Resevoir Size

Status
This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.
How about posting the .asc file(s)?

What diodes are used in the bridge? Maybe right-click on them and put some "real" power diodes in?

Need to model at least the main power-rail conductors, between the smoothing caps and the load, as series inductance and resistance (25 nH and 1 mOhm per inch).

Lower priority: Maybe try two of my transformer model, or similar, since the one being used doesn't have leakage inductance. (Might need a snubber, then, however.)
Can do, Tom. The "new" .asc file is a combination of material already posted on the forum, being AC to DC Power Supply - diyAudio and http://www.diyaudio.com/forums/atta...ells-power-amplifier-book-cordell-tmc-tpc.zip :

View attachment PowerSupply+CordellOutputStage.asc

The active device models are not included -- rather than pass them around, download them via the Cordell posting noted above.

Agreed, this is a first, rough cut, using the simulations as first posted. It's now wide open to make the scenario as realistic as possible, to gain greater insights. As an example, Bob's amp has good +ve PSRR, but the -ve rail is nowhere as well rejected. Hence noise on the -ve rail is likely to do major damage ...

Frank
 
Now this is telling ...!

As a quick experiment, I added the amp output stage to Tom's completely different PS setup, as per that attached to his recent post here. And tried with a gain of 75; the +ve voltage rail looked like:

Compare_PSU01.gif

This should be compared with http://www.diyaudio.com/forums/atta...pply-resevoir-size-basicpowersupply04-bc5.gif, and on first glance looks a lot more impressive! But, appearances can be deceptive: Tom's PS uses less overall capacitance so the peak to peak ripple is greater, check the vertical axis ticks. And, what turns out to be the key difference, his overall smoothing cap ESR is lower, by a factor of about 4. So what I did was change the ESR of the smoothing caps from 0.03R to 0.007R in the original PS circuit, nothing else, and had a closer look at a small part of the waveform:

mightydub PS:

Compare_PSU02.gif

gootee PS:

Compare_PSU03.gif

In spite of the fact that the supply circuits are radically different, the noise waveforms match very closely, and that is linked precisely to the effective ESR of the caps ...

Frank
 
Interesting. You might want to change the input line voltage amplitude on my PSU model to 330 or so, to get the same DC output voltage from the rectifiers.

I've been playing with something similar, since last night, but using a 40-second WAV of AC-DC's "Highway to Hell", since it has some punishing bursts.

The first thing I've done is tried to turn up the gain until just below clipping, which seemed to happen at about 61 Volts, into 4 Ohms. I thought that a gain of 750 was good until I got about 10 seconds into the song. Now it looks like a gain of 119 is about right, for the first 20 seconds of this source material at least.

Second, I noticed that since this is essentially a unity-gain power amp, we can directly compare Vin and Vload, and subtract (e.g. Vin-Vload) to plot the output error, and maybe then we can see what effects the reservoir caps have on the output accuracy.

Actually, there were small intrinsic offset and gain errors that I had to remove, first. So I added really-large ideal caps (100k uF both after the rectifiers and at the load) and then adjusted the plotted expressions for Vin and Vout to match as well as possible, which meant adding an offset of slightly less than 2 V to Vin and mutliplying Vload by a little less than 1.1. Right now the error expression looks like this: abs(V(vin)+1.966717-(1.082*V(load))) .

Then I was able to run sims for only certain time-segments of the song, by setting the "time to start saving data" and the "end time", and get the RMS and Average of the error for exactly the same time segment, for different simulation runs.

For example, I did that for 0.3 to 1.0 seconds, which is in the first burst of sound, in the song. I tried it with various reservoir capacitor configurations. That section is not extremely high in amplitude, so there was not a lot of difference. But, surprisingly, 30000 uF and 470 uF were both slightly better than in between (although 470 uF had some significant problems later in the song).

I also used current sources across the outputs of the power supply, with the load side disconnected, and plotted the output impedance of the supply versus frequency, with and without the power rail conductor model included.

I will post the plots and .asc file later this evening.

Cheers,

Tom
 
Last edited:
The first thing I've done is tried to turn up the gain until just below clipping, which seemed to happen at about 61 Volts, into 4 Ohms. I thought that a gain of 750 was good until I got about 10 seconds into the song. Now it looks like a gain of 119 is about right, for the first 20 seconds of this source material at least.

Second, I noticed that since this is essentially a unity-gain power amp, we can directly compare Vin and Vload, and subtract (e.g. Vin-Vload) to plot the output error, and maybe then we can see what effects the reservoir caps have on the output accuracy.
Sounding good,Tom (pun intended :D)! Only thing, to be fair to Bob's design, this is only the backend of his amp and lacks the distortion reducing feedback intrinsic to the circuit -- the point so far is purely to drive the PS hard to see how the voltage rails behave. So, to really assess the interaction of a realistic PS with a good performing amplifier design the whole rather than just a part of the latter needs to be incorporated.

I will post the plots and .asc file later this evening.

Cheers,

Tom
Before commenting further I'll wait to see what you have there. Thanks for getting on board!

Cheers,
Frank
 
Continuing in the vein of testing the "whole" amp behaviour, I was curious how much a variation of PS voltage rails would affect Bob's design, from the point of view of fairness, using it within design constraints. In other words, forgetting about maximum ratings for the moment of the individual components, how does the behaviour change between perfect 35V rails and 70V rails? To do this I created a model with duplicate, complete, amp circuits, one running off the lower voltage and the other off the higher. Fed both with the jazz track, and added a trace of the difference of the two outputs:

DuelingAmps01.gif

So, looks pretty good, the green and blue traces of the two amps overlay each other nicely, and the difference is just millivolts in value, just a DC offset really. But, there is a bit of AC noise, which worsens at points, which look interesting, so we'll zoom in:

DuelingAmps02.gif

Hmmm, some high frequency instability here, very low in amplitude, but it may be telling us something. Try zooming in again:

DuelingAmps03.gif

Now this is very high frequency, about 840kHz. This might be happening because the Spice models aren't accurate enough, or there is potential for instability in the circuit. Note again that the PS rails are ideal voltage sources in both cases. I haven't tracked down the origins of this yet, but it certainly shows the scope LTspice has for digging into the dark corners ...

Frank
 
Last edited:
I used current sources across the outputs of the power supply, with the load side disconnected, and plotted the output impedance of the supply versus frequency, with and without the power and ground rail conductor models included.

PSU_Cordell-output-only_tg_ZOUT.jpg

PSU_ZOUT_without-rail.jpg

PSU_ZOUT_w-rail2.jpg

PSU_ZOUT_both.jpg

It looks like we would want power supply decoupling and high-frequency bypass capacitors after the power rail, at the load, in order to present a low-enough impedance.

I also disconnected the reservoir capacitors and plotted the output impedance of just the rectifiers and transformer assembly.

PSU_ZOUT_no_caps.jpg

Cheers,

Tom
 
I got some temperature and frequency-dependent capacitor spice models, from Cornell Dubilier's JAVA Applet at CDE Capacitor Impedance Calculator and Spice Model Applet , and used them in the following simualtion runs.

I did simulations with 3x 10000 uF, 1x 10000 uF, and 1x 1600 uF, playing 22 seconds of AC/DC's "Highway to Hell". No, it's not my favorite song.

Plots follow.

Zip file should contain ALL files needed to run the simulations, in LT-Spice, except for WAV file for input.

Here's the song: http://www.youtube.com/watch?v=Xv24N8H1KyI

The plots correspond to the first 22 seconds or so of the song. (That song can also be "the highway to hell" for a simulation of a DC servo circuit, which is what I originally used it to test, with LT-Spice. One interesting aspect of using WAV files for input AND output is that you can then LISTEN to the output of your simulated circuit!)

Cheers,

Tom
 

Attachments

  • PSU_with_freq-dependent_caps.jpg
    PSU_with_freq-dependent_caps.jpg
    70.2 KB · Views: 319
  • PSU_30kuF_freq-dep_45C_6inch-rails.jpg
    PSU_30kuF_freq-dep_45C_6inch-rails.jpg
    142.9 KB · Views: 310
  • PSU_10kuF_freq-dep_45C_6inch-rails.jpg
    PSU_10kuF_freq-dep_45C_6inch-rails.jpg
    170.8 KB · Views: 310
  • PSU_1600uF_freq-dep_45C_6inch-rails.jpg
    PSU_1600uF_freq-dep_45C_6inch-rails.jpg
    171.9 KB · Views: 306
  • CDE_10000uF.zip
    11 KB · Views: 50
Last edited:
I did simulations with 3x 10000 uF, 1x 10000 uF, and 1x 1600 uF, playing 22 seconds of AC/DC's "Highway to Hell". No, it's not my favorite song.
Just a quick comment for now, Tom ... what that highlights, apart from nasty looking voltage rails(!!), is the apparent tradeoff between maintaining voltage headroom and the load drain modulating the voltage: with the 30,000uF the spikeness is much reduced but the caps are steadily being drained, from 71V to about 66V, the rectifier pulse of current is insufficient to top up the charge, at some point a balance will be reached, but what voltage will that be? Then, at the other end of behaviour, the minimal 1600uF produces terrible glitches of voltage drops, to 45-50V, but then recharges very quickly because of the lower capacity, so the headroom voltage overall remains close to 71V.

So which way is better, and what are the other "solutions" ...?

Frank
 
Is that an amplified inverted copy of the drum kick on positive rail ?

See the post after yours (i.e. post 372).

It shows that it's exactly as I've been saying all along:

The rail CURRENTS _ARE_ the signal, non-inverted, and they come from the capacitors.

Each rail does one polarity. So the currents look like a push-pull version of the signal.

The rail VOLTAGE just sags a little when the transistors open up, to let just the right amount of current out of the caps to create the load voltage.

Edit: It looks like the current for the negative rail cap (fC5) somehow got inverted, in the first image in post 370 (the 1600 uF version of the zoomed drum strike).

Cheers,

Tom
 
Last edited:
Here's a better zoomed view of the 3x 10000 uF caps' currents precisely tracking the load's output signal voltage.

They don't get to take a breath until the next charging pulse...

Cheers,

Tom

That's an ickey distortion right there in the middle! At the back slope of the second soundwave peak, when the caps at the 'zereocrossing' of the signal are in transit between exclusive supplying and simultanious supplying and charging. You see some overshoot when the caps 'drop' in that transition area, affecting the output (note the small vertical drop exactly at the point in time where the cap is supplying and starts being loaded while the negative rail cap starts providing the current.

This is actually a very intriguing aspect. Zerocrossing distorting effects seem to be introduced by caps 'alternating' between each other in supplying load. This could either be induced by the output stage crossover, or be an additional source of crossover inherent to symmetrical supply output.

All in all awesome work here in this thread!
 
Just a quick comment for now, Tom ... what that highlights, apart from nasty looking voltage rails(!!), is the apparent tradeoff between maintaining voltage headroom and the load drain modulating the voltage: with the 30,000uF the spikeness is much reduced but the caps are steadily being drained, from 71V to about 66V, the rectifier pulse of current is insufficient to top up the charge, at some point a balance will be reached, but what voltage will that be? Then, at the other end of behaviour, the minimal 1600uF produces terrible glitches of voltage drops, to 45-50V, but then recharges very quickly because of the lower capacity, so the headroom voltage overall remains close to 71V.

So which way is better, and what are the other "solutions" ...?

Frank

Frank,

Good insight. That is the trade-space that needs to be quantified.

The voltage goes back up, a few seconds after that sim ends, by the way. But that points out a possible criterion for choosing the reservoir capacitance: They have to be large-enough to have enough charge to supply a sustained "short term" demand, i.e. at least until the next charging pulse, but also need to be small-enough that a) they don't drag the voltage too low for a worst-case burst between charging pulses, and b) they aren't so large that they can blast out a LOT more current than a charging pulse can supply, because then, if they're unlucky with the size and timing of the demand versus the size and timing of the charging pulses, they could just continue to force the rail voltage lower and lower. That's part of the equation we need!

This looks like a probability problem, too. What fun! (groan).

ESR and ESL play a role there, too, both those of the capacitors and those of the interconnects on both sides of them. And the transformer and rectifiers will probably need to be able to supply extra-big charging pulses, sometimes.

Cheers,

Tom (going to bed now)
 
By the way, the plotted error (bottom plot pane in post 370's two plots, for example) appears to be meaningless, so far, since it varies only very slightly due to capacitor effects. It looks like I need to just devise an "ideal" amplifier, using spice's controlled sources, or something, so that the amplifier errors can be completely eliminated. That way, the error (difference between output and input) could give a valuable representation of any changes in the inaccuracies of the reproduction, due to changes made in the power supply or distribution, and should be able to be used to pinpoint their sources and might lead us to patterns or rules (or equations :) that we can actually use.

I will probably also repeat the plots I did today, but with extremes of capacitance, both large and small, to possibly shd more light on the trends and tradeoffs that were noted.

Now I really am going to bed...!

Cheers,

Tom
 
I have to agree with MagicBox - awesome thread.

Tom - you guys have done all the hard work, a mantra was the least I could do.

the droop is probably due to the transformer leakage inductance. it (assuming its the same as the one a few pages back) is 2%, but the crest factor is quite high so the ~ 6% droop will be this (only needs CF = 3, quite reasonable).

it wouldnt be hard to prove. extracting the leakage inductances from the transformer model would also make it easy to play with.

It would be interesting to see what the leakage inductance is for non-toroidal transformers - without interleaving I would expect it to be 5-10% so these would have far more droop.

PFC Boost converters would work nicely here - placed at the amplifier they would have very little output capacitance and draw nice sinusoidal input current. hell, it wouldnt be hard to make them compensate for the nonlinear transformer magnetising current, and draw unity power factor low-thd sinusoidal mains current. which must sound better, right? oops, crossed the fine line between the sublime and the ridiculous.

so turning it around, for a given transformer (ie leakage inductance) the allowable droop (before clipping sets in) sets an upper limit on the DC bus capacitance. somehow I dont think that will prove to be a useful design rule. ah well....

ISTR doing some maths a while ago to figure out conduction angle, crest factor and thd given load, cap & leakage inductance. I'll try and find it.....
 
Tom & Fred, I think between the two of you we are going to conclude that there is a definite rule of thumb and that just throwing money away to add as much capacitance as you can afford is not a solution. I think this is going to be an eye opener for most audio designers and possible a reference thread to any future designs.
 
Status
This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.