Are PSRR plots all they're cracked up to be?
Posted 15th April 2015 at 04:19 AM by abraxalito
My experiments last autumn with putting a transformer between a chipamp and a drive unit left me with one nagging question. Why was the sound so much clearer with the trafo than without, given that the trafo's voltage ratio was relatively low (1.5:1)? Even Frank was surprised at so much change and suggested some other effect was in play (amp instability). I've not ruled out amp instability but I have made some progress on understanding chipamp PSRR.
The first trafo experiment was done with the TDA7265 - I've since tried trafos on two other chipamps - TDA8947 and LM4766 with similar subjective results - a much clearer sound, more depth to the soundstage and so many tiny details on recordings becoming clearer that the experience has become akin to headphone listening. The trafo ratio has increased - 2.5:1 for the TDA8947 (21V rails, bridged) and 5:1 in the case of the LM4766 (again bridged), running on 62V total supply.
My current listening system is LM4766 with trafos on both input and output. The input trafos I've designed in perform two functions - they reduce the output swing from the TL084 AXO (which is is 8VRMS, balanced) and re-reference the signal to the chipamp's local VEE so as to avoid the poor -ve rail PSRR inherent in that chip (my thanks to KSTR for this suggestion).
The PSRR plot for the LM4766 is, for the positive supply, one of the most impressive in the business. Above 100dB for the whole of the audio band. I have the amp set up for a gain of 32dB on each side of the bridge. My tweeters are roughly 4ohms and the trafo's giving a 25X increase in the effective impedance - so 100ohms or 50ohms each side. Subjectively even when driving such an easy load the tweeters didn't sound clean when I first fired up the amp. The reason for that was insufficient local decoupling. I've ended up with 6 * Rubycon ZLH 390uF/80V caps right up against the chip with lead lengths under 2.5cm. Only with this seemingly draconian decoupling is the HF subjectively clean.
Here's some rough and ready calculation for what PSRR I should expect based on the shown plot. Since the amp gain's 32dB that's directly subtracted from the 100dB so 68dB would be the practical worst-case PSRR. At the very top end the supply impedance is determined by the ESRs of these Rubycons - I've chosen the lowest ESR types I could find - the DS shows they're 51mohms. With 6 in parallel we're at 8mohms, say 10mohm total supply impedance including the wires.
Since the load is 100ohms, the ratio of supply impedance to load is 10,000:1 or 80dB. If the PSRR is to be believed then I'm hearing HF hash which is at least 148dB down. The PSRR is actually always above 100dB and my ears don't go up to 20kHz.
A quick sanity check - 390uF is 136mohms at 3kHz (my crossover frequency) so doing the supply impedance this way gives about a 3X increase - but the PSRR at 3kHz is 112dB, more than 3X better than at 20kHz. So either way I cut it, I shouldn't be hearing any hash if that PSRR plot is to be believed.
At this point, if I was an objectivist I'd disregard my ears and put it down to 'imagining things' because datasheets don't lie and the engineers at semiconductor manufacturers are paragons of competence. However this highly popular narrative doesn't convince me so stay tuned for some LTSpice simulation plots...
The first trafo experiment was done with the TDA7265 - I've since tried trafos on two other chipamps - TDA8947 and LM4766 with similar subjective results - a much clearer sound, more depth to the soundstage and so many tiny details on recordings becoming clearer that the experience has become akin to headphone listening. The trafo ratio has increased - 2.5:1 for the TDA8947 (21V rails, bridged) and 5:1 in the case of the LM4766 (again bridged), running on 62V total supply.
My current listening system is LM4766 with trafos on both input and output. The input trafos I've designed in perform two functions - they reduce the output swing from the TL084 AXO (which is is 8VRMS, balanced) and re-reference the signal to the chipamp's local VEE so as to avoid the poor -ve rail PSRR inherent in that chip (my thanks to KSTR for this suggestion).
The PSRR plot for the LM4766 is, for the positive supply, one of the most impressive in the business. Above 100dB for the whole of the audio band. I have the amp set up for a gain of 32dB on each side of the bridge. My tweeters are roughly 4ohms and the trafo's giving a 25X increase in the effective impedance - so 100ohms or 50ohms each side. Subjectively even when driving such an easy load the tweeters didn't sound clean when I first fired up the amp. The reason for that was insufficient local decoupling. I've ended up with 6 * Rubycon ZLH 390uF/80V caps right up against the chip with lead lengths under 2.5cm. Only with this seemingly draconian decoupling is the HF subjectively clean.
Here's some rough and ready calculation for what PSRR I should expect based on the shown plot. Since the amp gain's 32dB that's directly subtracted from the 100dB so 68dB would be the practical worst-case PSRR. At the very top end the supply impedance is determined by the ESRs of these Rubycons - I've chosen the lowest ESR types I could find - the DS shows they're 51mohms. With 6 in parallel we're at 8mohms, say 10mohm total supply impedance including the wires.
Since the load is 100ohms, the ratio of supply impedance to load is 10,000:1 or 80dB. If the PSRR is to be believed then I'm hearing HF hash which is at least 148dB down. The PSRR is actually always above 100dB and my ears don't go up to 20kHz.
A quick sanity check - 390uF is 136mohms at 3kHz (my crossover frequency) so doing the supply impedance this way gives about a 3X increase - but the PSRR at 3kHz is 112dB, more than 3X better than at 20kHz. So either way I cut it, I shouldn't be hearing any hash if that PSRR plot is to be believed.
At this point, if I was an objectivist I'd disregard my ears and put it down to 'imagining things' because datasheets don't lie and the engineers at semiconductor manufacturers are paragons of competence. However this highly popular narrative doesn't convince me so stay tuned for some LTSpice simulation plots...
Total Comments 9
Comments
-
... [I] if[/I] that PSRR plot is to be believed ... :)
The point being, how does one extrapolate those numbers, tested in a certain fashion, to [I]all [/I]situations in a working setup? Does the PSRR behaviour alter when used at different gain settings - I did some sim's quite some ago where I used a fully modelled, through rather basic opamp circuit, and the PSRR behaviour at normally used closed loop gains was [I]not [/I]as expected! Also, does the PSRR behaviour alter when used at different loadings?
So much is assumed - only a final, working circuit in the configuration that will be used, that is rigorously tested in a fully debugged jig should be trusted, IMO, if one wants numbers, ;).
Curious how you going to check this out in LTspice, if a fully valid model for the opamp is not available ...Posted 15th April 2015 at 06:11 AM by fas42
-
Good questions Frank.
I'm working on part 2 right now and went back to first principles - that is the PSRR of a single transistor operating as an emitter follower. At least that's a common feature of all the opamp and chipamp output stages which aren't rail-to-rail types.
Yes to your question - PSRR in a closed loop (feedback) circuit is going to vary according to gain - more gain means poorer PSRR. Tom's ultra-high loop gain amp has proven that - but his poor results in THD for the 'bare' LM3886 on a non-bench PSU is one of the things which prompted me to look into this rather deeper. A quick finger in the air shows Tom's THD+N (0.003%) with naked LM3886 is more than 10dB worse than would be expected from the PSRR plot of the chip.Posted 15th April 2015 at 06:18 AM by abraxalito
-
Forgive me if this is an obvious question, but isn't the -PSSR just as important as +PSSR in a split supply device? Particularly because it's only 50% of what +PSSR is at 20K?Posted 15th April 2015 at 01:16 PM by MrSlim
-
Its important if your circuit configuration is the usual balanced one around ground yep. I've bypassed this issue though by creating a local 0V from a couple of TL431s in series (because the max per device is 30V and I want 31-37V depending on my power supply) and referencing that local 0V directly to pin4 (VEE) of the LM4766. So no signal-related current is flowing between local 0V and VEE meaning there can be no ripple. Therefore only the positive PSRR matters in my configuration.Posted 15th April 2015 at 02:21 PM by abraxalito
-
A couple of quick comments.
1. What makes you think that the roughness you hear (HF distortion?) should be related to the power supply ripple rejection figures? Sorry I don't follow the logic.
2. Your last reply "...no signal-related current is flowing between local 0V and VEE meaning there can be no ripple." sets off alarm bells. I'm not saying you are wrong, just that your explanation does not convince.Posted 15th April 2015 at 11:39 PM by rjm
-
Thanks for the question. The reason that the roughness on HF is related to the PSRR is because it reduces with more capacitance on the rails. More caps means lower supply ripple. No other changes were made than adding more low ESR caps close to the chip. I have no other explanation but that its related to power supply noise and the chipamp's ability to reject that. I got a similar effect with my TDA8566 - HF roughness went down from adding a lot (30 or so) of 10uF ceramic caps across its supply pins.
On your second point perhaps you could turn up KSTR's work in a similar vein - he convinced me that this is a reasonable method to eliminate the negative rail supply issues. He also reported improved SQ when he implemented his idea. So what about my explanation is lacking for you?Posted 16th April 2015 at 12:13 AM by abraxalito
-
A few things that may influence the PSRR:
By reducing the low frequencies there is an increase of the output switching between the two rails and therefore an increase of HF rail current.
Add to that the low impedance (2ohm) seen by each of the bridged IC. The bypass caps and IC are working harder than with a normal 8 ohm load and full audio signal.Posted 16th April 2015 at 09:08 AM by Mark Whitney
-
Mark do you mean because this is an active system the HF isn't riding on the LF in the same amp, so the OPS is changing halves more frequently? If so good point that I'd not considered before, I need to reflect on it.
Yep - bridged amps see half the impedance. When I was driving the tweeter directly that wasn't with a bridged amp though - just a normal TDA7265. Thanks to some dodgy engineering on some other speakers I tried trafos - one model is using TDA8947 (which is bridged) and fed into 4ohm units. If not for that I'd not have discovered trafos!Posted 16th April 2015 at 09:17 AM by abraxalito
Updated 16th April 2015 at 09:19 AM by abraxalito -
Compare the supply currents with a 10kHz signal to a 10kHz plus 100Hz signal.
Or the same test comparing a filtered and a unfiltered audio signal.
You could also add DC to each of the outputs and force class A operation.Posted 16th April 2015 at 10:15 AM by Mark Whitney
Updated 16th April 2015 at 10:20 AM by Mark Whitney
