The real story about closed vs open loop class D amps
Very interesting paper which I'd not come across before which raises about as many questions as it answers. Thanks to Bob Cordell mentioning it in his book, here's the link :
https://www.eetasia.com/STATIC/PDF/20...URCES=DOWNLOAD
Compare the FFT in fig3 (closed loop) with fig4 (open loop) in the text - which do you think will sound better?
Here's my take on what the paper is claiming, FWIW. That PSRR measurements when the amp has a BTL (balanced) output aren't appropriate because there is output stage cancelling of them. My first head-scratch moment came because this is true of other amp topologies too, not just classD as BTL outputs are very much alive and well in NXP's range of chipamps for automotive use. So why restrict this to just classD?
Secondly what's the argument that PSRR in BTL amps isn't appropriate? They seem to be saying (though they don't express this very clearly) that whilst output stage cancelling is fine for purely linear systems, the IMD component of the PSU-related noise at the output won't be subject to cancellation. So they are in effect saying that purely linear models of PSRR are misleading at best. So why has it taken a BTL output stage to expose this PSRR fraud? If purely linear models aren't cutting the mustard then that too is independent of whether the OPS is BTL or SE. What's good for the goose is equally good for the gander, so all linear PSRR modelling isn't suitable for audio.
Which is a conclusion I'd myself come to anyway. Interesting to see it confirmed in this way though
Let's look at the argument for no OPS cancellation a little more closely. When under music stimulus noise is being added from the PSU eventually finding its way into the amp's output there are going to be sum and difference products of multiple orders resulting in IMD with a noise-like spectrum. Noise modulation in other words. It seems to me that running balanced is going to give cancellation of some (even order) of those IMD products in the same way as 2nd harmonic distortion gets cancelled when running balanced. The question is - how much cancellation can we expect in practice? I suspect the answer is 'not a lot' - but in the worst case I'd expect 3dB, a halving of the total noise power where the noise was totally uncorrelated between the two balanced channels. Still 3dB isn't anything to sniff at and the noise will never be totally uncorrelated. More correlation would lead to more cancellation. So I can see some benefit to running balanced still..
The second question arising in my mind is whether feedback really is helping us in this new, non-linear PSRR scenario. PSRR plots in DSs have always been shown for single frequency (sine) stimulus. Wrapping feedback around an amp, for low to moderate levels of feedback tends to multiply up the higher order components (spectral growth distortion highlighted by Bob Cordell in his book, page 502 in the first edition). Hence Bruno's dictum that feedback should be used to the max or not at all. However in any feedback system there's a unity gain crossover point and there will be the dreaded low-to-moderate levels of feedback immediately below that frequency.
Inspect the fig3 (TAS5706) plot at the HF end.To my failing eyes there's an apparent rise in the (odd-order) harmonics as the frequency goes from 7kHz and up. Their level is of concern - 7H is -70dB (0.03%) wrt the stimulus and the others get higher. This distortion won't be masked by the ear. But the distortion isn't the only concern - the noise 'floor' is also rising, a lot faster than the individual harmonics. The FFT bin bandwidth is easy to work out - about 5Hz from eyeballing the bottom left where the step sizes are obvious. So its quite misleading to show as they do a 'dynamic range' indication of 100dB between the noise 'floor' and the tone. When the noise is 100dB down in a 5Hz bandwidth that corresponds to a 36dB FFT gain so the integrated noise is actually -64dB and we've seen that its by no means flat at the HF end.
Take-away - it seems obvious to me that the TAS5706 is going to sound lousy with its apparent 64dB dynamic range on a sinewave test. What am I missing?
Update - one of the authors of the paper I linked has written an AES paper describing a way to characterize amps for PSRR - https://www.aes.org/e-lib/browse.cfm?elib=15214
Anyone have a copy?
Update2 - now that I've digested some of what's at Frank's link I'm more in a position to comment further on the original paper.
Firstly it seems output modulation method (AD or BD) plays a part. I've been trying to determine whether the TDA8932 chip uses AD or BD and its not clear from the DS so I'll have to get the scope out and have a look. The story goes that for AD (anti-phase carrier output switching) the PSRR shows up as 80dB at LF. The figure depends on matching between the two halves of the OPS - each needs to contribute exactly 6.02dB PSRR for cancellation to occur. A pretty impressive figure but it degrades at higher frequencies.
Now for BD modulation (carrier in phase at the output) the condition at zero input does not depend on switching - rather its like the OPS is putting out a common-mode signal but no differential mode one. Here it seems to me is the real reason for wanting to change the measurement method - with zero input the OPS behaves like its gone into a Hi-Z state with respect to differential mode output. Hence its totally unrepresentative of real music.
This could be the reason that the TPA3116's PSRR at HF looks so good from the DS - its around 60dB, compared to the TDA8932's 20dB. Which in itself suggests the TDA's using AD modulation. I'll test this hypothesis and then return to update the update.....
Update 3 - the result from scoping the outputs is that its using BD, the switching component is in phase between the outputs. So bang goes that hypothesis! Perhaps NXP has another way to measure PSRR....
https://www.eetasia.com/STATIC/PDF/20...URCES=DOWNLOAD
Compare the FFT in fig3 (closed loop) with fig4 (open loop) in the text - which do you think will sound better?
Here's my take on what the paper is claiming, FWIW. That PSRR measurements when the amp has a BTL (balanced) output aren't appropriate because there is output stage cancelling of them. My first head-scratch moment came because this is true of other amp topologies too, not just classD as BTL outputs are very much alive and well in NXP's range of chipamps for automotive use. So why restrict this to just classD?
Secondly what's the argument that PSRR in BTL amps isn't appropriate? They seem to be saying (though they don't express this very clearly) that whilst output stage cancelling is fine for purely linear systems, the IMD component of the PSU-related noise at the output won't be subject to cancellation. So they are in effect saying that purely linear models of PSRR are misleading at best. So why has it taken a BTL output stage to expose this PSRR fraud? If purely linear models aren't cutting the mustard then that too is independent of whether the OPS is BTL or SE. What's good for the goose is equally good for the gander, so all linear PSRR modelling isn't suitable for audio.
Which is a conclusion I'd myself come to anyway. Interesting to see it confirmed in this way though
Let's look at the argument for no OPS cancellation a little more closely. When under music stimulus noise is being added from the PSU eventually finding its way into the amp's output there are going to be sum and difference products of multiple orders resulting in IMD with a noise-like spectrum. Noise modulation in other words. It seems to me that running balanced is going to give cancellation of some (even order) of those IMD products in the same way as 2nd harmonic distortion gets cancelled when running balanced. The question is - how much cancellation can we expect in practice? I suspect the answer is 'not a lot' - but in the worst case I'd expect 3dB, a halving of the total noise power where the noise was totally uncorrelated between the two balanced channels. Still 3dB isn't anything to sniff at and the noise will never be totally uncorrelated. More correlation would lead to more cancellation. So I can see some benefit to running balanced still..
The second question arising in my mind is whether feedback really is helping us in this new, non-linear PSRR scenario. PSRR plots in DSs have always been shown for single frequency (sine) stimulus. Wrapping feedback around an amp, for low to moderate levels of feedback tends to multiply up the higher order components (spectral growth distortion highlighted by Bob Cordell in his book, page 502 in the first edition). Hence Bruno's dictum that feedback should be used to the max or not at all. However in any feedback system there's a unity gain crossover point and there will be the dreaded low-to-moderate levels of feedback immediately below that frequency.
Inspect the fig3 (TAS5706) plot at the HF end.To my failing eyes there's an apparent rise in the (odd-order) harmonics as the frequency goes from 7kHz and up. Their level is of concern - 7H is -70dB (0.03%) wrt the stimulus and the others get higher. This distortion won't be masked by the ear. But the distortion isn't the only concern - the noise 'floor' is also rising, a lot faster than the individual harmonics. The FFT bin bandwidth is easy to work out - about 5Hz from eyeballing the bottom left where the step sizes are obvious. So its quite misleading to show as they do a 'dynamic range' indication of 100dB between the noise 'floor' and the tone. When the noise is 100dB down in a 5Hz bandwidth that corresponds to a 36dB FFT gain so the integrated noise is actually -64dB and we've seen that its by no means flat at the HF end.
Take-away - it seems obvious to me that the TAS5706 is going to sound lousy with its apparent 64dB dynamic range on a sinewave test. What am I missing?
Update - one of the authors of the paper I linked has written an AES paper describing a way to characterize amps for PSRR - https://www.aes.org/e-lib/browse.cfm?elib=15214
Anyone have a copy?
Update2 - now that I've digested some of what's at Frank's link I'm more in a position to comment further on the original paper.
Firstly it seems output modulation method (AD or BD) plays a part. I've been trying to determine whether the TDA8932 chip uses AD or BD and its not clear from the DS so I'll have to get the scope out and have a look. The story goes that for AD (anti-phase carrier output switching) the PSRR shows up as 80dB at LF. The figure depends on matching between the two halves of the OPS - each needs to contribute exactly 6.02dB PSRR for cancellation to occur. A pretty impressive figure but it degrades at higher frequencies.
Now for BD modulation (carrier in phase at the output) the condition at zero input does not depend on switching - rather its like the OPS is putting out a common-mode signal but no differential mode one. Here it seems to me is the real reason for wanting to change the measurement method - with zero input the OPS behaves like its gone into a Hi-Z state with respect to differential mode output. Hence its totally unrepresentative of real music.
This could be the reason that the TPA3116's PSRR at HF looks so good from the DS - its around 60dB, compared to the TDA8932's 20dB. Which in itself suggests the TDA's using AD modulation. I'll test this hypothesis and then return to update the update.....
Update 3 - the result from scoping the outputs is that its using BD, the switching component is in phase between the outputs. So bang goes that hypothesis! Perhaps NXP has another way to measure PSRR....
Total Comments 17
Comments
-
That feedback is digital, is it not? That's a whole different ball game, how does the chip do its tricks internally is the question - I note generally positive comments about SQ, in spite of very uninspiring distortion figures ...
And, I also just discovered the TAS5756, which I note you have also come across. This might do quite spectacular things if handled well - is this done in a different way from the TAS5706? Plenty of scope here for fine detail tweaking - everything done in one tiny box, erh, chip might just work nicely if fed with very clean power and signals.Posted 11th July 2014 at 01:07 PM by fas42
-
You are asking if the feedback for the 5706 is digital? Meaning there is no internal ADC ? The DS doesn't give enough detail to work out how they do it. Where have you learned about the SQ? On DIYA?
Yes I found the 5756 and mentioned it on the TPA3116 thread. No-one else has commented on it so far but I agree it could be great (even ignoring the DAC). Yes its different from the 5706 as there's an analog classD amp, not a digital-to-PWM converter. This tells me that most likely even they don't like what they've done in the 'direct-to-digital' chippery and are reverting to good old analog classD.....Posted 11th July 2014 at 01:39 PM by abraxalito
-
Yes, in a closed loop, digital input unit how is the feedback actually done? Probably there is some explanation out there, but I'm being lazy at the moment ... :)
Yes, the 5706 is mentioned several times on diyAudio, and generally given a thumbs up for doing impressively well, subjectively, for what it is.
[I]If[/I] the 5756 innards are really intelligently thought out then it may do remarkably well, there being no room for interference to creep in via sloppy, conventional implementation. Personally, as I've mentioned many times, the S-D DAC is not a problem, after switch-on conditioning always sorts that out, IME - if aiming for optimum sound in, say, 5 minutes from power up then just add extra, peripheral circuitry to keep the chip nicely simmering in the "off" times.
Interestingly, in the my PC setup which uses a cheap Realtek onboard DAC, you can easily hear this process happening. It's configured now to run at maximum analogue volume always, the digital volume control [I]is [/I]the gain setting. And from a cold start you can [I]hear[/I] the DAC glitching and stumbling, the digital artifacts are obvious, as if it had only about effectively 6 bits working. Steadily, as the chip runs in, warms to its work, :), these artifacts fade away and finally completely disappear, even with one's ear to the speaker.Posted 11th July 2014 at 11:44 PM by fas42
-
Ah I'm not interested in 'doing impressively well' with such a qualification. If it does impressively well unconditionally then my ears'll prick up.
One of the reasons I like the 5756 better than the TPA3116 is they don't seem to have included the on-chip regulator for the analog signal processing. To my ears that LDO reg craps out the bass performance, it must be too noisy at LF (no surprise really, its a CMOS/DMOS process). Seems not to be a reg on the 5756 as the power pin is brought out directly.
Your 'correct' sound in the way you've talked about it seems to have only one dimension. My sound though has dimensions even once the speakers disappear. So I still hear that for example the classD amp's bass sounds better than the classB - does this mean both create 'correct' sound? Or is classAB bass 'incorrect' and just needs a few more farads of capacitance?
Do you have any idea what changes on a digital chip as it runs in? Any hypothesis?Posted 11th July 2014 at 11:57 PM by abraxalito
-
Doing impressively well unconditionally is one of the hardest things to get happening, IME. I've felt extreme frustration over my years of fiddling, because I could only get the quality of sound I was after if enough factors were in place - and many of those required "warm up" times, hours often. This is what has made it so difficult to just give a recipe of what needs to be done, because there are so many variables, far too many. Hence, "conditional" is not a problem for me - it [I]always [/I]has been conditional ...
The ultimate goal is to have a complete system, which you can take anywhere, plug in anywhere, at any time - and within, say, 5 minutes it's running at near 100% of its potential. That's a pipe dream, I've never heard anything come anywhere near that - I can live with compromise!
How the 5756 is implemented, in every area, will determine whether it's a winner or not. If it behaves in the fashion of sounding like absolute crap on turn-on, and takes 2 hours to come to life, when it becomes exceptional, I can live with that - the power used to to keep the chip trickling along when not listening to it is trivial, if talking about generating convincing sound.
'Correct' sound will mean that the recording always sounds the same, the word "better" is the problem here, each person will have their interpretation of that. And, agreed, bass is probably the area where there may be the most "disagreement", people have different standards for such. I like intense, "tight" bass - no bass unless it's meant to be there; I've mentioned Boney M. a few times, the tracks as recorded have a bass quality which I find very satisfying, and systems nominally quite superior to what mine was at the time completely miss the boat in terms of reproducing that. Pete, the chap up the road who has a DEQX unit, with monster subwoofers, way beyond what I've ever had, has bass which sounds exactly right to my ears - it matches in tonal quality, in tracks that I'm very familiar with, in the the bass area with what I would expect.
The running in, to my mind, is largely about capacitance behaviour stabilising - the storing and releasing of charge. The S-D is all about about switching on and off of a lone capacitor, the single bit of its design. That capacitor has to be as "perfect" in behaviour as possible, otherwise there will be some type of distortion - and since capacitance is ultimately a materials behaviour exercising it causes the characteristics to change. Also, how good are the switches, how perfect are they? Are waveforms of the transitions always identical, or do they settle down in behaviour over time?Posted 12th July 2014 at 01:10 AM by fas42
-
I've only been fiddling for maybe four years or so but I have ascertained a considerable number of ingredients in the recipe that works for me. I've not exhausted the details by any means but so far it has been
a) Multibit DAC kitted out with enough caps on reference and supply to keep LF noise in check. One DAC per speaker driver.
b) Passive I/V
c) Steep filtering with passive filter into the band required by each driver of the speaker (2-way so far, 3-way still to try)
d) As little active circuitry for voltage gain as humanly possible. AD815 with CCS loading and lots of caps on the rails comes closest so far to being a 'passive voltage amp'. With enough current from paralleled DACs I think perhaps this voltage amplification could be disposed of. This is an experiment to try.
e) One amp per driver kitted out like DAC in a) with the addition of extreme HF noise decoupling from running in classB
f) Transformers to couple between DACs and amps if they're not in the same box running from the same supply.
Overall its a constant iteration of the basic principle of keeping the dynamic noise (noise modulation) as low as possible through attention to power supply related noise and IMD-generated noise. Frustration has been entirely optional
Putting this another way I'm still working hard trying to break the original hypothesis that satisfying sound is characterized by having the lowest noise modulation. So far the hypothesis has stood up well in that changes to the system which have been intended to lower noise modulation have indeed resulted in improved listening satisfaction.
A few things are on the 'still to be determined' list :
i) the optimum place to put the volume control
ii) How to incorporate more bits (hi-res recordings) than 16
iii) Related to ii above - to what extent is the set-up limited by my choice of the TDA1387 as DAC?
iv) Is there another solution to LF noise without requiring oodles of caps?
v) Is there anything to be gained from digital filtering prior to the DACs?
vi) Any advantage of applying the 'limit the bandwidth' principle to the LF cut off of the bass/mid driver?Posted 12th July 2014 at 01:43 AM by abraxalito
Updated 12th July 2014 at 02:10 AM by abraxalito -
Okay, first an update on 'correct' sound - my take on that is that it's in place if I can put on a recording, [I]any[/I] recording - and nothing about the sound bugs me, I can let go, and go with the sound emotionally - no artifact is irritating enough to distract me, making me aware that it's there - hence the "no bad recording" mantra.
Personally, I've found that rough and ready pop and rock recordings are almost ideal for debugging a system - everything is in them which will stimulate discernable distortion artifacts, and the pattern of sounds repeats constantly through a song, making it easy to pick the 'problem' occurring. If a tough pop recording passes muster, then nearly every classical recording sails through with ease ...
Volume controls are bad news, get rid of them! Use digital, or electronic chips, or similar - the impact of ordinary pot type mechanisms is far too audible.
16 bits are fine, I'm happy with downsampling from hi-res if necessary, offline - I have not heard a single higher resolution recording that has made me say, "Gee, look at what I'm missing!!"
I still don't get what you mean by LF noise - noise is low level distortion in my book, and I never separate it into high or low. That said, I'm happy with using regulation to getting LF ripple under control, at the moment.
Any filtering I would do offline - prepare a doctored version of the track to play, to suit all the following equipment.Posted 12th July 2014 at 02:37 AM by fas42
-
Further on the DAC playing up when not fully conditioned, I'm listening to Perahia doing Mozart right now, at very low volume from switch on only a short time ago - the basic sound is very clean, good treble - but talk about noise modulation! On the right speaker a strong whistling or bzzz'ng, like someone blowing over a bottle, warbles all over the place depending upon the music; on the left, much better, like a white noise, gate switching on and off constantly - this channel has settled down much more already. The level of these artifacts modulates ferociously with the volume and content of the music - but, from a distance the music content is fine, comes through cleanly, has a distinct identity and presence separate from the these obvious distortion elements.Posted 12th July 2014 at 04:32 AM by fas42
Updated 12th July 2014 at 04:49 AM by fas42 -
I'm fairly cool with 16bits too, but customers do demand more on occasion (for whatever reason) so I'd like to have a solution for this.
LF noise is a fairly new discovery for me. LF ambience is on many recordings, its what gives a venue its acoustic signature. On the TDA1387 without phenomenal amounts of decoupling the low level, low freq ambience information gets masked (presumably by noise modulation). Dan (Max Headroom) alerted me to the importance of low levels of LF noise - the difference is perceived as an enveloping LF 'bloom' when its low enough.Posted 12th July 2014 at 04:49 AM by abraxalito
-
Okay, that LF ambience I would term general ambience cues, not especially in the deep bass range - and, yes, it depends on the DAC working properly, much digital playback fails because that low level information is not retrieved well enough for the ears to decode what it's meant to represent, which, as you say, is the acoustic signature of the recording space.
That same low level information is critical in unscrambling very 'messy' or 'difficult' pop recordings - the original Led Zeppelin I mastering has this component very strongly in its sound makeup, and typically fails to impress on playback because sounds are reproduced that don't make sense - the verdict is that it's a "bad recording" ...Posted 12th July 2014 at 05:54 AM by fas42
-
Richard, that paper you mention in your update has a powerpoint style summary of the paper here, [url]https://wenku.baidu.com/view/cf49390b763231126edb118f[/url], 2/3 of the way down. Might be enough to give you the gist of it ...Posted 13th July 2014 at 03:23 AM by fas42
-
Thanks for that Frank, looks like there's more to the story than indicated in the initial paper so another update is in order. Before I write that though I'll need to study more in-depth the different OPS modulation styles as these seem to impact the PSRR....Posted 13th July 2014 at 05:09 AM by abraxalito
-
Maybe you two have way to sensitive ears (and I thought I have to) and very high demand on your electronics.
It is much easier for me, once i have a pair of relative good speakers ,the sound quality is depending on the recording and a media carrying it. For example I don't understand how a simple, or a set of resistive dividers can degrade the sound so much at this would bother me. Not sure of the LF ambiance either, you mean like a bass pipe decaying in a church? The "dynamic noise (noise modulation) as low as possible" How low, is there a value, a number here? How would one measure if a DAC can retrieve the low level information, or not. Is playing back a -70dBFS tone and analyzing it would be good enough, or would one need a digital storage scope to see the decreased S/N on a dynamic signal? It's a surprise to me that the Led Zeppelin I would be any good for testing audio, other than maybe amplifier power output, no acoustic instruments in it, other than the drums.Posted 13th July 2014 at 11:03 AM by miklos
-
Perhaps you are assuming that because Frank shares something here that I find the same as he does? No, I've yet to experience the 'degradation' of a resistive divider and I use them to set the right levels between my DAC and amp when I'm too lazy to wind a new trafo with the right turns ratio.
No, I don't so much mean the decays of bass (organ) notes rather the whole LF 'bloom' or ambience which sets apart holographic soundstaging from the merely decent.
As for how low, no there's no number as I've yet to decide a measurement. I'm sure its measurable though, just requires some ingenuity. Playing back a single tone - no, that wouldn't be a worthwhile exercise as its so unlike music. The test stimulus really does need to be as music-like as possible, if you've not come across such a stimulus I suggest you search on 'MTPR'.Posted 13th July 2014 at 01:00 PM by abraxalito
-
If a normal volume control was close to an ideal resistive divider then I wouldn't have issues with them - but they're not, they typically also contain two poor quality sliding contacts - and that's where the damage is done.
Simple way to test it. Set a pot to give the right resistive division, measure that setting out of circuit - then swap in and out either the pot, or pure resistors that match the pot setting. I did the equivalent decades ago, and the quality compromise was too great ...
miklos, what I'm after is that the heard quality doesn't depend on the recording 'quality' and media - the goal is that 100% of everything I put on hits the spot. Led Zeppelin I is an amazing recording, it has a massive acoustic which is immensely deep, that makes normal, "audiophile" recordings sound pretty pathetic - it [I]would[/I] be a huge shock to many people to hear it reproduced properly, at realistic volumes ...Posted 14th July 2014 at 12:22 AM by fas42
-
Richard,
I'm sorry to take you and Frank ,so to speak under one hat.
Actually I tried multi tones already. A year ago I purchased a QA400 analyzer to replace my old HP one. This new one is able to dig deeper ca. -130dB, instead of -90 with the HP. In the Cool Edit I made a test CDR with a set of test tones, among three Phi6, Phi12 and Phi3x4. the frequency's I took from Jon Rich's web page. The "Phi" frequency's are chosen so that the generated harmonics would not fall under the fundamentals and not be masked by them. I measured a bunch of CD players (ca. 35-40), Bernard was doing interesting measurements on DAC's 10 years ago and presented his result on Diyaudio. Looking at CD players and playing those tones to, going from Phi 6 to 12, I could see the noise floor (distortion) increase. Also tried a noise signal with a gap in it, but I'm not sure how good it is. I imagine the best would be a gated multi tone signal and see how the transition region would look like. Any case I would like to see your results on this interesting subject.Posted 16th July 2014 at 06:39 PM by miklos
-
Hi miklos, no worries - your multitone results look promising. Is Phi6 a six-tone signal and Phi12, twelve tone? If so I suggest you build up more tones than 12, a minimum of 100. With 100 tones the crest factor could be up to 40dB.
I only put up results with a six-tone signal on my blog when I was developing my hi-end chipamp but I think the noise floor was already limited by problems with the DAC in that case. We are looking for noise floor changes - 'noise modulation'. To do decent measurements I'll need an ADC which doesn't suffer from noise modulation itself, i.e. an SAR one, not S-D as 99.9% are nowadays. I do plan to develop one eventually.Posted 16th July 2014 at 10:56 PM by abraxalito
