Sauermann Amplifier

[gsbrva]

Thanks for the responses.

I haven't had time to follow up lately as I wished (too much to do)!

Greg, I wasn't even aware of that this indeed is the whole point inyour view

No, not at all.  But, it is a complex topic and I got off on a side issue.

My original post was regarding proper use of gain structure on the analog side and the correct location (in my opinion) for attenuation and also using minimum gains and so on to control noise.  We really agree on this stuff so much Peter, that I feel bad arguing.  But, as a little aside, I'm finding that the very best amplifiers at my home are also the best for both horns and electrostats.  My favorites are Class A, unconditional stabile, ultra low distortion, etc.  Sometimes this creates a gain problem with certain sources.  So, while I agree that most people with high sensitivity speakers are using lower gain amplifiers (talking voltage here).  It doesn't always work out. 

I will summarize my complaints about advising use of digital volume as an absolute truth.

1)  Not enough evidence.  Yes, the NOS1 w/digital attenuation is heard to sound better than using an analog pad following.  But, that does not prove that the general idea is correct, only that it is correct with the NOS1.

2)  It's not hard to build an analog pad that far exceeds the bandwidth/distortion performance of any other audio equipment.  For instance, I have a pad on the output of a waveform synthesizer to feed the external sync on my scope.  It's flat to at least 50mhz when properly terminated.  This is not a special or rare device.  Resistors with low noise and excellent voltage coefficients are available.  I'm having a hard time believing that using a different range in the PCM1704 ladder and running the I/V at a different signal voltage has less sonic "character" than using a analog pad.  Also, if you don't like inserting a  voltage pad at the amplifier, why not have different fixed voltage outputs in the I/V.  You are already dropping the DAC current across a load resistance at some point to convert to voltage and using a different load results in a different output voltage (can be attenuation).  No need to discuss the  I/V design, that's not anyone's business by yours IMHO.  I just want to point out that there are spots to pad that introduce no extra components at all.     

3)  Also, I do have a problem with the loss of bit depth.  It seems like you are asking for two incompatible things to be true:  Either the "filtering" action of upsampling and arc prediction is needed or not.  If we use the bit depth for attenuation, then we are simultaneously changing the character of the sound by reducing the number of bits for upsampling.  Are we now saying that changing the number of bits for upsampling does not make an audible difference?  I was believing that one of the main advantages to XX/NOS1 was the method of filtering.


More later....happy listening!

Greg

[PeterSt]

Ok Greg ... This post contains a few words which was actually at the end of this post : http://www.phasure.com/index.php?topic=1976.msg20913#msg20913 but it got too long and I cut it out. But saved it though. Here you have it after all and it hopefully implies some answers or shows my thinking. Implies, because I made it before your last post obviously.

More food for thought.  Has anyone analyzed how often two adjacent words in redbook pcm happen to be within a few LSB (least significant bits) of each other.  If that doesn't happen often, it would help explain how arc prediction can work well with less than ideal bit depth (as when using 24db of attenuation).

Nah, that is not how I ever looked at things, although there may be some truth in it. Anyway, it will be tough to meet adjacent samples with about the same amplitude because it would imply a lower frequency wave without any higher frequencies (modulated on it). That's merely something for an MP3 approach I guess.
So no, there are no such secrets in there. Thus, when we were allowed to say that Arc Prediction is no real DSP within itself, then at least no DSP is added to make it work better. On that matter it's a "linear" algorithm (and if we'd compare it with normal filtering, then *that* for sure is pure DSP).

But think a bit different;
There won't be any way that a real 24 bit depth will be met. Not anywhere I think. In between lines : notice that I can see what's really happening because of the nature of the NOS1 (which does nothing itself to the sound).
Now, try to think from the other end. That end where I said that at -141dBFS there's still a nice little signal (and no other harmonics to be seen). How can it be ? The most officially it can't, and certainly not when I see that 24 bits can never be met.
Still it does ...

Of course, that 16 bit signal is made 24 (32) by Arc Prediction, but this looks quite contradictionary if we also take into account that each upsampling step takes out one bit to begin with. So, 16x takes 8 bits and 20 are left. Fine.
Then we attenuate 141dB and another 23 are taken. So we took 31 out from 24. Hmm ... I'm starting to think perpetuum mobile now.
This at least proves that when we attenuate e.g. 24dB it shouldn't be as worse as having lost 4 bits.

All I can think of (and this for sure is no law) is that the inherent noise of the system is a perfect dither means. But a couple of things are in order here for good (?) interpretation.
So, first off, higher noise will carry more "patterns" than lower noise. This is just because noise, generally, never is just "white noise" and it is creaeted by something. That something implies a pattern. Could be a voltage regulator, your USB transfer, anything. The more sources of this, the more complex the pattern will look and the more random it will be again, but there still will be a pattern. The more sources (at different frequencies) the longer it takes for the whole pattern to repeat. Nice.
At the same time, the more noise, the less we can get the signal to stick out (of that noise). So, although to some extend we can say that many noise sources will "randomize" more hence create a better dither means, it is useless because of the noise itself (which gets profound).

When all noise sources are perceivedbly eliminated, it should be a complete random thing again (of which I don't believe it exists), then thinking of molecule noise (though in electronics) which will be fairly random. When this happens, we'd have and the perfect dither and the sufficiently low level to utilize it. The signal will still be there, as in my example (sorry I don't have a picture of it).

It is still more complicated when we see that nobody is going to hear that -141dBFS signal, but still we could say that -somehow- it must be so that 24 bits are resolved, while nothing in the chain really did it (also not the PCM1704).
And no, even in my system (or better, the NOS1) no way 24 bits can be resolved when we see that the FS output creates the noise at -120dB to begin with. Too bad.
Ah ... oh ... but when I attenuate 20dB the noise goes down as well with 20dB. Ok, doesn't help because SNR is and remains 120dB (actually 117 because the signal itself is at -3dBFS and now suddenly we recognize the specs of the 1704U-K ...)

And thus the dynamic range (which is that 117dB) depicts that 24 bits can not be resolved. Too bad again.
Ehm, yea, but at attenuating 141dB the signal is still visible.
Oh.

From another angle (but maybe I already told this), when I attenuate 21dB digitally, THD specs go down with 6dB only. Hmm .. Now what.
This indeed partly is about the 1704 not being optimal at Full Scale, which merely is in between the -10 / -22 DBFS range (there it's quite flat). So that matters too. But furthermore it is a bit tricky to look at THD specs, because they are always THD+N and I am actually not much interested in how it relates to noise, while I am only interested in creating a nice "shaped wave" (eliminate stepping distortion and such). So ...
When I attenuate 21dBFS and THD+N goes doen by 6dB only, then I must gave GAINED on THD. How ? well, the closer noise floor says 21dB thus when it's 6dB worse in practice, the wave shape (THD part) must have become 15dB better.
In the mean time I must have lost 4 bits (or 3) but I guess now it is important that we look at 16 bit sources. So there we didn't lose a thing.

Nice story eh ? well, again too bad, because the story is wrong. Remember, at attenuating 21dB the noise floor also drops with that (well, 20dB max). Now what ?
Now nothing, and all what happened is that the chips perform relatively better at -21dB compared to FS. So, somewhere deep down I *am* using 24 bits, which is just because I use them all (and not only 1 more for one upsampling stage), and thus -again deep down- it must be so that attenuating with 21dB must lose 3 bits. That's 18dB while the figures tell 6dB. Fine, then the chips behave 12dB better, relatively. In absolute sense though it is still 6dB worse compared to Full Scale, and my problem will be that I used all the 24 bits to begin with. Had I used 4 only (for 16x) then -if all is right- output at -21dBFS would have (12dB) been better for THD+N compared to FS. Too bad once more that it would be worse to begin with.

If I am not stuck myself in this little story then for sure you out there will be, and STILL the whole story is wrong (I seem to like wrong stories). Again ? yes;
What I suggest all the time is 4x upsampling only. How ? well, indirectly via the analyser's capacity which limits to 24/192 (for its ADC). So ... when I engage 4x and compare that to 16x no difference will be seen in THD+N because the analyser's resolution can't go beyond 4x. Sadly this means that my given specs of 0.0018% THD+N are also not right and in fact they are way better than that. Now think ...

We drop bits (by attenuating) but we don't drop sampling speed. This means that at some stage the analyser will meet a kind of balanced situation between the number of samples it can take (192000/s) and the resolution in the bit depth.
In order to understand this we must first see that a bit depth of 24 is way too much for the slow sampling speed of 192000. So, we take 192000 samples per second, and could theoretically register 192000 adjacent (1 step difference) volume levels. But 24 bits has 16,777,216 volume steps. So that's a way too high granularity compared to what the sampling speed can deal with. Until the number of bits drop low enough to have it even (and each sample could show an adjacent (1 step) level change).

With 768000 samples per second this works a kind of other wat around; Still the bit depth is way too high, but at attenuating we will meet the balance more early (at attenuating, thus at less attenuation). Without real math, 4 times more early and I am not sure whether I must say that this will thus be "24dB more early" with which I try to say that the attenuation of 21dBFs plus the test signal which is at -3dBFS is exactly that 24dB where I see that flipping point of relative better THD. So then it would shift to FS ...
(and the chips are the best at FS afterall)

If this were true (which it undoubtedly is not) then the chips (but measured at the output of the NOS1) should show 0.00011% THD+N which seems feasable. The official TI specs are 0.00136% and it only needs to be ~18dB better to meet this 0.00011% while the differential + parallelled setup (4 chips/ch) may go in that direction.
But hand me 60-90K for a good analyser and I can just look ...

FYI: The way I measure is 100% comparable with the way TI measures (I've been through some nice projects regarding this with them).


Where were we ? ah, that "there is more";
The above ridiculous story is just to give the hunch that many things and phenomena are to be taken into account for the real interpretation of what is happening. The dither thing is almost like voodoo, but this is just what dither is to begin with. If one next feels a bit at home with what these chips do, you can even go further. So, where I suggested 0.0036% THD+N at -21dBFS, I am as far as being able to make that 0.0028% or so. Just by means of software, and I still wouldn't call that DSP.


Peter

[PeterSt]

And then more explicitly ...

I will summarize my complaints about advising use of digital volume as an absolute truth.

1)  Not enough evidence.

Eh, wait a minute. This is not (meant to be) a stupid theoretical debate like we see them everywhere (were it about preamp / no preamp stuff), but real practice from all angles. At this side at least it is.

For instance, I have a pad on the output of a waveform synthesizer to feed the external sync on my scope.

Which is a spectrum analyser ? If so, you talk practice too and we somehow seem to be even. Agree to disagree is vocabulary I never liked in my life. I am right or I am wrong, and never leave myself in the dark. Conclusion here ?
Something must be wrong.
You could show a picture of that flat 50MHz, but better make that the audio band first. It may unveil some things I predict and you never saw ? Besides that I am not so sure that a waveform synthesizer depicts reality. Just for fun think about how we should measure (for example !) loudspeakers; Do we take the software made for that and which includes the generation of waves / pules erc, or do we maybe need to used the software which performs best in its environment (which is computers !) anyway (think what XXHE is made for) ?

Yes, the NOS1 w/digital attenuation is heard to sound better than using an analog pad following.  But, that does not prove that the general idea is correct, only that it is correct with the NOS1.

Sorry, you are too fast. Not that you can help that Greg (you just can't know I suppose), but XXHighEnd was there for the public in 2007, and the lossless digital volume was introduced in that same year probably. Totally unrelated to the NOS1.
I am not saying that everybody unconditionally used the digital volume only, and merely the contrary. To this regard we must recognize that posts like in this Sauermann topic from my hand are the most rare on this forum. Not on others, but here they sure are (actually, sadly). With this I want to say that people choose for themselves and I never told to get rid of preamps outside the NOS1 (but people could have read it elsewhere of course). So ... what people do from nature (and some hints from me indeed) is applying a tad of digital attenuation (often 6dB) because it works out for the better. How ? the same thing as I said before ... because the analogue volume works (out !) so much differently than the digital one. This *is* what I told in this forum, and people start experimenting with it. Same thing with Arc Prediction which is a better example : everybody uses it, while I explicitly told not to use it because *that* was for the NOS1.
And why would people prefer a preamp ? because it nicely works as a huge filter for all what is wrong in front of it (DAC).

The above summarized a little : all what emerges for good sound is created from user experience and most often only afterwards I try to find some reasoning behind how things work out. Btw, I am such a user too. As is my wife. And yes, it sure happens that theories (in advance) don't work out. So, I listen. Not only to music, but also to you and my wife. Next comes the reasoning of the why, and I always manage to fiddle that reasoning my way. Not that you find much of it in this forum, as said. But see the previous post as such an example of a half-cooked reasoning (which I had to cut the other day because I can go on and on before it's really done).

Resistors with low noise and excellent voltage coefficients are available.

I see this as theory in a wrong context. Haha;
I count 3 active components in my whole chain which includes the 1704 (which latter I count for one) and when I count the 3 amps I use per channel as one too. For discrete resistors I come to 3 in the signal path, again in my whole chain. In *that* context I am talking. If you (can) do that too, we talk about the same. So :

I'm having a hard time believing that using a different range in the PCM1704 ladder and running the I/V at a different signal voltage has less sonic "character" than using a analog pad.

See my before post, and I think I already admitted that. Not that it is audible ... which only tells that this should be harmless. And :

why not have different fixed voltage outputs in the I/V.

Pray that you will never meet the day that you will be working on your I/V with the 1704 *and* can measure it ! (so, better listen only ... you will sleep better).
IOW, that doesn't work either, and people will know I certainly tried (hard). In at least two occasions (the one a year after the other) I tried this for a customer. The first time it couldn'e be done, and the second time I forgot what I actually tried the first time, but after hearing it I recalled it by the sound. Really.

So about that context a paragraph back, ... please listen or measure in such a context first, or otherwise it is one big mess to begin with, and nothing will matter anymore;
The "probem" I have is that I started out with only one resistor in the signal path (were it for the DAC) and no single other component in there (passive I/V). Better than that it won't get, if you can bear the low gain (which I can/could). That is, for distortion (and not so much for overall (buffer) performance). So, with all further developments to give it a decent output level, I only wanted to mimic that "pure" non distorted behaviour. So I know, it can't be done other than I did it (with this D/A chip of course).

Would you be playing with, say that native environment of a passive I/V only, you will learn soon (if you can measure) that there's one optimum and one only. Remember, we are talking about that same 0.0018% THD+N ever, be that passively or how I do it now. So, this is the base resistance (or impedance) the chip wants (empirically found). Make that higher or lower and THD gets worse.
Today ? today I indeed can change the output level to some extend without hurting the figures too much (changing that particular resistor), but the figures will still always get worse because of SNR already (assumed all is tuned to max output already, so I can only get lower).
And then the fun : Attenuating digitally instead of applying that lower output, always gives better figures (don't ask me whether *that* difference is audible, with a theoretical Yes because -as I said earlier- digitally attenuating is not audible (not to me, and not from anyone that I heard of even a single time), so then the figures start to speak (better use the good THD base).

I just want to point out that there are spots to pad that introduce no extra components at all.

Superfluously : not that I could find. But hopefully it is your time soon ! hehe

3)  Also, I do have a problem with the loss of bit depth.  It seems like you are asking for two incompatible things to be true:  Either the "filtering" action of upsampling and arc prediction is needed or not.

Besides what I said in my before post playing a role, I also said earlier on that not all the bits are utilized anyway. So, attenuating a reasonable 30 bits takes out 5 of them, while upsampling x16 needs 4. So, this example (with too much attenuation for my own liking) loses one bit on purpose. Now, I can refer to my -141dBFS story (with leaving it to you to explain *that*) for enough reason not to worry, but I could also say that by the time you will be able to peceive a dynamic range of something like 90dB something else very strange must be going on. Same like the -141dBFS story ... "not that you will be able to perceive that (of course highly distorted signal)". This losing one bit is different, because a. the decreased DR isn't utilized by your ears in the first place and b. the possible sticking out additional false harmonics are way more down (than 90dB) anyway.

Now Greg, you may get less sleep because of thinking that losing this one bit is harmful, but if so, consider two things :

1.
I say that possibly (!) HD rides on the music ("signal") itself, which thus would make the distortion audible at the playing level (say that the level is just "loud" in your room). I say (that I can imagine) this, but don't even know it (like from noise I sure do), so maybe it isn't even true.

2.
This is to be compared with not losing bits at all, and listening (or measure for that matter) to that nice analogue attanuator of which I know the result and which is completely audible and totally measurable.

And as a bonus :
This is all in the midst of using something like Minimize OS from XXHighEnd which is 100% totally completely audible but which is measurable by no comparable (FFT) means. What does this say ?

Well, that *when* things are measureable by means of FFT (outside of what can be (linearly) expected), it must be really really bad (different) for the audible result.
But remember my context, and you better show me your "context" by means of an FFT (noise line of the whole chain) before I guarantee you that all is moot in your situation because it is a big mess to begin with - and no additional poor resistor will harm (could even improve ).

Ok, long enough again.
Peter


PS:

We really agree on this stuff so much Peter, that I feel bad arguing.

I read that ! Therefore I see this as sparring and I can always be influenced by good argument (which leaves my ears to convince of course ).

[gsbrva]

Quote
1)  Not enough evidence.

Eh, wait a minute. This is not (meant to be) a stupid theoretical debate like we see them everywhere (were it about preamp / no preamp stuff), but real practice from all angles. At this side at least it is.

Lol, I deserved that.  I was trying to figure out how to say what I wanted to without making it very long.

You could show a picture of that flat 50MHz, but better make that the audio band first. It may unveil some things I predict and you never saw ?

Fair enough, I was referring to sine waves on a 50 ohm termination.  As far as flat, not really of course, but reasonably so.  There are always variations in the signal source (a DAC!).  The attenuated and unattenuated signals into the scope channels will track each other.   You could say (and be correct) that I'm just comparing two attenuators (digital in the scope preamp and analog in the 50 ohm line).  I think this ignores your major point about noise, hence spectrum analyser.

I really need to get my shop computer set up as a test instrument.  Those photo captures are pretty handy.  So, no photos to back up my possibly defeated argument......

Sorry, you are too fast. Not that you can help that Greg (you just can't know I suppose), but XXHighEnd was there for the public in 2007, and the lossless digital volume was introduced in that same year probably. Totally unrelated to the NOS1.

You have me there again.  I can't remember, did I mention that I haven't used digital in my home until the past year.  I am trying to catch up now!  I did occasionally listen to pro gear and mixed a few shows on digital consoles/studio work or whatever.  But it's very hard to hear in those situations.  Too many variables.  Anyway, I did not know that lossless volume control is that recent for home playback.   

all what emerges for good sound is created from user experience and most often only afterwards I try to find some reasoning behind how things work out.

Amen!  I think this is the most significant statement you have made here.

I do believe that everything we hear can be explained by accurate theory.  I have yet to hear anything that didn't eventually make sense once I had all the information.  Can I measure everything I hear?  No, not very often.  I think also, that without theory, it's easy to get lost down the wrong road, following our ears down the easy path with immediate gratification only to miss out on much better sound by solving parallel, interrelated problems.  I know that seems non-specific theory , so for instance:  I've noticed many times people equalizing with component choices to overcome harmonic content causing what they perceive as a frequency response problem.  Don't even get me started on speaker design.  I would say that maybe "ears don't lie, but they also don't think for you".

I see this as theory in a wrong context. Haha;
I count 3 active components in my whole chain which includes the 1704 (which latter I count for one) and when I count the 3 amps I use per channel as one too. For discrete resistors I come to 3 in the signal path, again in my whole chain. In *that* context I am talking. If you (can) do that too, we talk about the same. So :

 

Yes, the same here.  A TDA1545 and two amplifiers.  But I do not want to defend my current setup.  It's still just for testing ideas.

Pray that you will never meet the day that you will be working on your I/V with the 1704 *and* can measure it ! (so, better listen only ... you will sleep better).

Would you be playing with, say that native environment of a passive I/V only, you will learn soon (if you can measure) that there's one optimum and one only. Remember, we are talking about that same 0.0018% THD+N ever, be that passively or how I do it now. So, this is the base resistance (or impedance) the chip wants (empirically found). Make that higher or lower and THD gets worse.

Yes, I notice the same issues on the 1545 with far less range!

See, I made another incorrect assumption.  I had expected your load (to the 1704) to be active (an emitter or a source) for lowest Z and the translation to voltage happen across an isolated resistor.  I totally agree that with passive I/V, there are no choices when you find the right value.  Noise lies in one direction and distortion in the other.  At least that is what my limited experience is showing.

Thank you for the longish post on bit depth, noise and dynamic range.  I haven't had time to digest it yet.

When you speak of full scale not being ideal on the 1704, are you referring to how it's loaded or to intrinsic properties of the chip itself?  I had meant to ask earlier if there is a sweet spot in the output level and does that give an advantage to digital attenuation in this case?

All for now,
Greg

   
   






 

[PeterSt]

Thank you for the longish post on bit depth, noise and dynamic range.  I haven't had time to digest it yet.

Let me merely thank you for a very nice post which was a joy to read to begin with (I won't be able to do that ).

Don't even get me started on speaker design.  I would say that maybe "ears don't lie, but they also don't think for you".

It is amazing how much one can work on e.g. crossovers and have one thing right but the other not. I did so for more than a year at finalizing the crossovers in my own horns, together with the manufacturer of these fine speakers. Not that the credit goes to me here, but the point I want to make is that after more than a year of balancing out things, I (we) came to the conclusion that all we tried to solve could be solved right in the player software - by means of having that better. So, just better. Nothing equalized or anything - better. This is why it appeared useless to measure speakers outside of this "better" playback software. It matters too much.

When you speak of full scale not being ideal on the 1704, are you referring to how it's loaded or to intrinsic properties of the chip itself?

I hope that the correct answer (regarding the question) is : both. So, first the maximum signal (always -3dBFS in my case), and next load it such that the highest output (in Voltage) comes from it. In that case the FS (-3) is not the best (IIRC this shows just in the curves in the datasheet). Attenuate this digitally with 10dB (up to 22 or so) and you are better off. The same can be achieved by a lower resistance to the chip (so it will have less output) but all what happens is that the curve becomes linear - hence at FS it then it as its best. Funnily enough (or logically I think) the effect is nearly the same, because you will have less output, and thus can use less digital attennuation. So, this (also) balances out for THD. But you can't be rough about it, because all is a very fine tuning thing. Here you see something of it (open this post in another browser instance and put the pictures next to eachother so you can do alt-tab to easily compare). Watch the yellow lines only :

The first one ends (FS, at the right side) at something like 0.015%. The curve is linear as you can see; each more digital attenuation will degrade THD linearly. Now compare this to the second picture, which doesn't show linear at all. THD at FS is now a tad worse only, but as you can see it is better in quite some attenuated range (some -4dBS to -20dBFS for sure). FS is still quite OK too, but attenuating is better. And what you also see is that the second picture is better all the way, up to -60dBFS or so, where things collapse. This is okay with me, assuming we can perceive a DR of 70dB max only anyway, but, of course this is relative to the attenuation to begin with (notice that this -60dBFS is not only happening because of attenuation, but already with softer music (or the more micro detail of it if you want). But this is the chip, and remember, this is passive and only tried to squeeze out the most of it, which was 315mV in this case (in the end I achieved 630mV by another setup).

Now also look at the purple line in the first picture, which clearly does better at the low end. Way better in fact. What does it tell you ? that this isn't alowed to be used with less than -10dBFS of attenuation, and when that is applied THD is at its optimum from -10dBFS to -30dBFS. Attenuate more is allowed, but less sure is not.

In fact this is what I was talking about all the time and here you see it happening, once the load is so that this emerges indeed. And of course you start to feel how complex this is because in the end any curve is possible and it really can be so that a worse looking curve performs way better if only some guidelines are followed (like that purple trace would need a guide clearly).

Below you see two more where the first one reaches almost 0.0012% at Full Scale. The second one makes ~0.00125 of that but can sustain it to -6dBFS, which the first one can not. So, better than the before pictures. Small problem : this is 115mV output only and even for me unusable.

I showed you the last two pictures, because it should make clear that this is the base of it all, and once we are going to amplify this, the base - hence those curves stay. Still it is not said that I use one of the latter two for the base, because the overall THD (at more attenuation) is worse again. There is no free lunch anywhere (we say it must come from the length or the width, which is almost literal here), but we still can chose the best and apply some rules more down the line.

Peter

[manisandher]

Hey Peter, nice explanation. Looking forward to reading Greg's response.

But just to dumb things down to my level, I have a question. In my main room, I use XX's vol control usually between -18dB and -9dB (although -6dB for very, very quiet pieces). And 'Peak Extend' gives me a further -3dB on top of this, right? So it looks like I'm usually the 1704's optimum range. But in my office, the NOS1 is connected to a preamp (the Pass X1), which although far from ideal is useful because I like switching between different sources as I work (NOS1 for PCM, Mytek for DSD and my FM tuner for radio). Peak Extend is always on (off for HDCD of course), but should I set XX's vol control to say -9dB anyway, rather than the 0dB that it's currently set to?

Like I said, dumb question because the real answer is "get rid of the preamp". But this would just make my life a lot more difficult.

Mani.

[PeterSt]

Hi Mani,

No dumb question at all;
This is related to what I talked about a few posts back, and users by themselves (without me telling them) coming to that 6dB or so digital attenuation while using a preamp. So, a. this should be better, but merely b. those applying it found out by themselves.

Notice though that this was a kind of "hot" when the digital volume was introduced and for psychological reasons people start to play / test with it. Today it's always there already, and people just don't bother when they use a preamp (etc.). But the same "guideline" obviously still applies.

Please notice also that I myself don't bother about it too much, already because I myself don't use that preamp (etc.), meaning : all I could be bothered about is whether my gain matches the output of the DAC and next whether I play in that better range. But the latter is a stupid thing of course, and the former applied (ever back).

Next notice that indeed this ever was not about the NOS1 at all, although for sure more D/A chips will be subject to this matter. A next thing could be that when the DAC as a whole is not all that good at the slew rate (merely back down - hence preventing overshoot) it would be possible to go beyond the allowed voltage (what the voltage rails are supposed to deal with) which will incur for clipping. This is a DAC thing of course and shouldn't happen, but now look at our little nice Weiss experiment ... So, here something *else* incurred for the too high voltage, and it would just clip because the DAC isn't made for the thing (bad input).
(others : allow me to not elaborate further here)

More interesting (for testing the merits of the difference between analogue and digital attenuation) would be to play very soft (like people in the house sleeping already) while using a preamp. You can go either way of a balance between huge attenuation in digital vs. analogue, and I sort of claim that when the balance is towards "more digital" you will perceive way more detail at the soft levels. A better balance in the sound too.
But just try it.

Lastly, everybody with an NOS1 will use it in its optimum range. Why ?

1. Because I used the "guidelines" reversed, meaning :
2. People should allow themselves to play the rare very soft albums too while at the same time they won't incur for huge digital attenuation (like 48dBFS would be huge, and by now hoping that people will understand this, hence adjust their gain where possible).

... With the result of people playing in exactly that range, except for this way soft album, which is rare anyway.

Sneaky ...

[gsbrva]

I envy your capability to post those analyser pics Peter.  I see what you mean about the distortion at FS.  A lot of things make more sense to me now.

I'll try to attach a file.......

Maybe other readers, even if you aren't technical will find this interesting. 

Here is the inside of the PCM1704 (see attached jpeg).  I thought it might be helpful to see where the business of producing the sound is happening.  The diagram shows only two bits due to space considerations.  I circled the MSB (no, should be highest current bit) section in red.  The other bits are represented by the single section circled in magenta and are typical for the remaining 22 bits.

You can see that the highest current bit switch (in red) is connected directly to the output. With a low Z load, nearly all of it's current will sink into the load.  The other bits are summed progressively through the rungs of the R2R resistor ladder and summed to the output.  The rungs of the ladder divide the currents from the lower bits to give the decreasingly less current for each lower bit. 

I think it's interesting that the NOS1 sounds better with some attenuation.  If you digitally attenuate, each -6db, abandons the use of another significant bit.  First the highest current bit is unused and then the adjacent sections as further attenuation in 6db steps.  For each additional -6db, the summing point for the bit currents is moved further down the ladder.

The lower bit switches that connect into the R2R ladder load are already working into a higher impedance than the highest current bit section by design.  They already swing voltage regardless of the load on the DAC, unlike the highest current bit.  (back to the HF filter arguments again, lol) 

Thanks BTW Mani for the report on your past volume control tests. 

I wonder if the slight advantage in dynamics (for digital attenuation) you heard was due to voltage swing issues in the MSB.  Having the load non-ideal (for completely understandable reasons) might make that bit current slightly out of calibration and perhaps affect adjacent bits as well.  It depends on operating conditions of those switch transistors and we can't easily know that.   

On another related topic:

You might see my point now why I tried to say that the resistors in part of the ladder are working in a similar way to an analog attenuator. 

When you are not using the higher bits, those switches are left in a static condition and that unswitched section of the R2R ladder is just a resistor acting as one leg of a voltage/current divider out to the load.  The bit currents have all been summed previously in earlier steps of the ladder and the signal flowing is the complete analog signal.  The resistive divider is attenuating this signal voltage.  That is the definition of an analog volume control.   All IMHO of course and I'll wait the response.

Back into theory heh.  With an ideal load (maybe not possible at this sound quality) then the advantage of digital volume should disappear, yes?

Sorry Peter, but I couldn't resist having one more go at this attenuator sparring (as you say)!  It was just too tempting.  

BTW, I do have 4 of the K spec 1704 chips, but I thought I'd try damaging the regular ones first   I would tell you guys the I/V design, except it might sound like cr*p and I'd prefer at least some of my failures be private!

Anyway, another can of worms is open .

(post edited to correct misuse of term MSB by me)

Cheers,
Greg


 

[CoenP]

I think it's interesting that the NOS1 sounds better with some attenuation.  If you digitally attenuate, each -6db, abandons the use of another significant bit.  First the MSB is unused and then the adjacent sections as further attenuation in 6db steps.  For each additional -6db, the summing point for the bit currents is moved further down the ladder.

I thought the msb is allways involved no matter how much you attennuate since it is determining the dc value of the output. Iow you only attennuate the ac component made by the other bits in the r2r ladder.

Am I missing something here?

Regards, Coen

[gsbrva]

Hi Coen,

You are right, in a regular ladder DAC that is how it would be.  However, look at the bit switches and think about the Colinear feature.  In this chip, the MSB in the word determines which bank of switches are used (positive or negative) and depending on which side of zero all bits are processed on either upper or lower section.  If the total value of the word is more than 50% of full scale then the highest current bit of that "bank" is switched on.  This way you never cross zero with MSB currents.  You can see the two banks in that every bit section contains two differential pairs of switches.  Very clever, these guys   At least that is how I think it works.  I have to be careful here in case someone who actually knows this stuff chimes in.

Greg

[gsbrva]

I should not be using the term MSB to talk about this.  Yeah, no attenuation on that bit as it always must exist in the word.  Inside the dac is different.  I'm not a digital guy, so I struggle for the words on this.  I should maybe say, the highest current bit instead?  I'll edit my posts so they make more sense.  Thanks for pointing that out.

Greg

[PeterSt]

Yo Greg, thanks for that post, and adjusting it on that "MSB" part. But maybe it wasn't so much wrong after all. And maybe it isn't all *that* important ...

In your perception of how "analoguely" it works, you may forget about something, and this is the continuous change of that "analogue attenuator". So, it seems that you approach this as a static set of switches, while they switch all the time (let's say 44100 times per second (or 705600 times per second in the case of the NOS1)).
This is already important when you look at the bipolar setup of the chip, and your mentioned "zero crossing". So, zero crossing in a virtual fahsion all right, but it merely is the switch from positive to negative from one digital sample to the other, and what that implies for current change normally (current goes from zero to "full scale") - and how this is solved in the chip (a bit of another subject, but you seemed to touch it).

I wonder if the slight advantage in dynamics (for digital attenuation) you heard was due to voltage swing issues in the MSB.

Better turn this into : ... we measure is due to the voltage swing issues in the MSB.
Because remember, I measured it first (depending on the created "slope") and next people came up with it, without me pointing it out.
But it also goes beyond just "stupid DAC output" since there is so much more going on. Not qualifying it really, but consider this :

Something like the NOS1, and especially its means of filtering, allows for relatively super transients towards the analogue part of the chain (starts in the gain stage of the DAC itself of course), which theoretically could be too high for whatever is behind it. For example, imagine that pulse train to reach the speaker diaphragm, and maybe wish the "dynamic range" would be somewhat lower. So, distance of the pulses remain, but tops are lower. The "sharpness" of them has gone, and next you will be amplifying it with a (more) noisy amplifier. Could work out for the better ...
Remember, just one angle, while many more can be thought of.

But this was an example of how less DR can work out for the better, while I merely point out that there just *is* more DR at that little attenuation (see that -6 to -22 or so in one of the pictures (slopes)). And never to forget, in this case (NOS1 specific) the noise goes down linearly.

But still from the same angle, think about what too high micro dynamics may do (with emphasis on "too high") : they will imply distortion and that by itself will be looking as "feshness". It is a pitfall so easy to fall in. It may even color things like cymbals a nice natural way (the transients of them - which are gradients - riding along a slope which nicely works out in "sibilance" which is natural to the cymbal itself).
So much dangerous stuff here ...

Anyway :

The bit currents have all been summed previously in earlier steps of the ladder and the signal flowing is the complete analog signal.  The resistive divider is attenuating this signal voltage.  That is the definition of an analog volume control.   All IMHO of course and I'll wait the response.

So, if a normal analogue volume control would be changing itself 10s of thousands of times per second, then maybe. Haha.
But on this part an R2R D/A chip wil only make things worse, because now there's also the "glitch distortion" (there's a more formal pehenomenon for that, but I forgot it at this time). But let's say this emerges by the current changes themselves, a bit similar to the "zero crossing" you taked about, though way smaller. You can (also) well say that the current the chip needs is changed by herself to it deteriorates herself.
But this is only pro normal analogue volume, so never mind.

We can all sum it up by stating that there's no way to compare the both with easy theory because too many parameters play a role and we always have to look at the whole chain. Or think of how the passive I/V suddenly is subject to the reactance of the amplifier which is out of our control (thinking random apmplifiers). So many parameters ...
This is why (endless) measuring tells all. That is, once you believe in better figures sounding better, and while at first (long ago) I myself thought that better sound not necessarily measured better, that indeed is a long time ago now. It just does. Always.
Did I say always ?

Regards and thanks,
Peter

[gsbrva]

Please bear with me, I am going somewhere with these posts and thanks for the reply Peter. 

Still not sure if I should use MSB and LSB for the ends of words that get truncated.  I ended up confusing myself with that.  From now on I'll use the bit numbers as shown on the datasheet.  MSB of the original word is 1 and the LSB is 24 (for 24 bit audio).  Sorry all for the sloppy way I worded that post.  I'll try to clear it up now .

One of the reasons I posted the schematic is I never saw a good explanation of the 1704 internals and what things there affect the sound in which ways.  I did quite a few searches and only found that Japanese application note and I can't read it either BTW   At least the schematic is clear.  It has been posted elsewhere on the net, but with little or no explanation. 

I think I understand the point you are making about dynamic range.  However, I'm still having trouble with valuing bit depth based on the noise floor.

I'm also still wondering about filtering benefits if using more than 4 bits for upsampling.  I'd love to see a spectrum analyser, but maybe it wouldn't resolve.  I suppose you have done listening tests and I cannot do that yet.  I also question if small transients and noise riding up near the high frequency limit, might get imperfectly filtered at 4 bits.  I mentioned in another post that I couldn't understand how adjacent words that only differ by one LSB (at 16bit) can be filtered to an arc with 4bit upsampling at x16.  That is still confusing to me, since the math doesn't work. 

So, if a normal analogue volume control would be changing itself 10s of thousands of times per second, then maybe. Haha.
But on this part an R2R D/A chip wil only make things worse, because now there's also the "glitch distortion" (there's a more formal pehenomenon for that, but I forgot it at this time). But let's say this emerges by the current changes themselves, a bit similar to the "zero crossing" you taked about, though way smaller. You can (also) well say that the current the chip needs is changed by herself to it deteriorates herself.

I don't think we are seeing it the same way.

I'll try to explain my mental image of it again.  This is what I see when I look at the schematic:

At zero, the two switches for each bit are in opposite states.  Each controls a separate current for that bit.  One is normally open and the other is normally closed. 

Depending on the state of bit #1 MSB, closed switches can open or open switches can close, for either positive or negative waveforms (from zero).

Starting from silence, as the signal level builds, increasing large bit currents are used starting at bit #24(LSB) and working towards #2 (max current).  The state of the #2 bit switches never changes unless the signal exceeds -3db.  The state of the #3 bit switches never change unless the peak signal exceeds -9db, with bit #4 at -15db and so on.

As TI / Burr Brown says in the datasheet "The sign-magnitude architecture, which steps away from
zero with small steps in both directions, avoids any glitching
or large linearity errors"   (edit:  They call this Colinear in the PCM63 datasheet)

The high level bit switches don't do anything at all unless there is high level signal.  When you digitally attenuate, you are removing the possibility the high level signals can exist, so the high level bit switches sit idle.  The output end of the ladder is just sitting there as passive as can be.  Even the closed switches are a very high impedance current source and have no effect.  There is no switching or glitching, just analog attenuation.

This is very different than my lowly TDA1545 or other dac designs.

Each time the digital attenuation reaches another -6db boundary, another bit switch section can no longer be actuated since those word values cannot possibly exist.  So if I'm right about this .  We have the first boundary at -3db (since there is a possible +3db for peak extend).  Any digital attenuation of -3db makes it impossible for data to exist that could actuate the bit #2 switches, so that section of the ladder then becomes passive.  When digital attenuation of -9 is called for bit switch #3 is likewise idle and another section becomes passive.

Those now passive sections of the ladder each attenuate 6db (when properly terminated) The signal is already analog at this point inside the dac.  It is analog because it is complete and contains all the bit currents.  The bits have been summed (1 bit per rung) through the R2R ladder.  By digitally attenuating, you are progressively moving the active digital portion back up the ladder away from output end in 6db steps (1 ladder rung at a time).  I'm guessing that the rungs are composed of 500 and 1000 ohm resistances, since they claim 1000 ohm output impedance.  As you digitally attenuate, those resistances are still dividing current/voltage, even if there are no bit currents being added at that ladder rung. 

If you took that same resistor network and put it outside the DAC what would you call it?   

I now view this digital attenuation as an analog stepped pad inside the dac with 6db steps, in series with a 0db to -5.999999......db digital pad.  This is inspired by our debate

The "analog" steps should be slightly wrong (not 6db) for the first couple steps since the ladder is intentionally terminated out of spec(for passive I/V), if I understand correctly the NOS1.  This also explains the distortion graphs you gave since the current division in the R-2R should develop small errors as the output end is approached.  The last couple bit currents in the ladder would not get divided as accurately.

The distortion graphs are actually strong evidence that my explanation is correct, I hope, haha.

Maybe this view seems crazy and If I'm wrong I would certainly appreciate a better explanation. 

Best regards,
Greg

[PeterSt]

Haha, say that !

Your logic about the bipolar (as how it's called for the 1704 to my best knowledge) seems correct to me, although I have problems with its working anyway. So, not with your outlay, but with its working. But, as it still seems to me, you let guide yourself by its good intentions (of its working) to derive other things which now seem wrong. Seem - to me;

Your perfect example is the Peak Extend (smart digging btw) which points me to how you actually think. So, with Peak Extend - and 3dB attenuation always for the necessary headroom - the whole bipolar feature can't work. That's what you say, right ? Ok. I agree. But about there the agreement stops.

You talk about the signal slowly rising, and thus the bipolar feature won't be active in that stage. Correct again in my view. But where it goes wrong is that somehow you pose that at any wave cycle or something, bipolar gets active and then ... then what ? then only helps distortion for that highest possible output level ? that by itself is something I have problems were it about the design itself (that is, what I logically derive from it for its merits), but merely : that highest output level may not happen at all. So, play a nice classical piece and only a relative few seconds of the whole piece the highest MSB may be "active"; all the other minutes - not.

So, while you say "when the signal starts to rise" you somehow seem to imply that at some stage it will reach the maximum digital level, while this just is not true.
For a test signal maybe. But even there, only for a fraction of the wave cycle.

We better say that our both perception of how the bipolar feature works is wrong. As a matter of fact I already know that at least I am wrong (so you too hehe) because I can prove that by some means of measuring. But that doesn't imply I know how it works. Merely that it's always active, and isn't hung op to the MSB.
A clue could be that servo part (see your datasheet) which IIRC is described differently (in words) in the normal datasheet; combine the two and your insight migyht be different (I didn't look for ages, but this is what I recall).


In between the lines and maybe funny :
At this time there is no single person within TI to be found who really knows how all is working, and whether it's at it best or could be better. Maybe the person exists, but he can't be found. I mean, by the TI organization herself. It also seems that I may know more about the stupid chip than anyone else within TI. As I said in between the lines earlier : I know how they test the chips, what they are tested for *and* what is wrong with it. This latter comes from myself only, because no feedback about it is available, or maybe I should learn Japanese. So, we spent some nice posts on it by now, but imagine it to be ten times more/longer when the chips are really down.

I mentioned in another post that I couldn't understand how adjacent words that only differ by one LSB (at 16bit) can be filtered to an arc with 4bit upsampling at x16.  That is still confusing to me, since the math doesn't work.

The introduction you have in your last post, where this is part of that, shows some great deal of confusement, while something like this quote doesn't make it better.
At least in that other post it was clear to me what your thinking was. WAS, because now you seem to explain it better.
Oh.

What you say, I think, is that we have a stream of samples and each over the other differ one LSB, BUT, going up and down. So, the general level stays the same.
Now think practically ...
If this really would be the case, it would be about a frequency of 44100 - ehm, no, 22050 (up *and* down). So, 22050 indeed is resolved by two samples only, and here you'd have it. Okay, at -90/-96dB.
When this is upsampled 16x, we have 32 samples instead of 2. No problem. The wave gets virtually 16 times longer, but, it's also passed 16 times faster. Same result (for frequency).
(btw, count out where the sample points are for the 2 sample base, so this is *not* a square but a wild monster (urging for sinc filtering).

Referring to my earlier little talk about this kind of stuff ...
the 32 samples have sharp edges now. The 2 also had sharp egdes, but it was in balance with the (time) length of it. So, analogue could smoothen the 2 samples, while with 32 the story will be different (never mind this, and it may be hard to justify, but it makes clear the nest better, hopefully);
We need to smoothen those edges, which happens by adding more level resolution. So, adding 4 bits (2 x 2 x 2 x 2) just does that. Less is still out of balance, and too many is a kind of useless (also not in balance).
The edges now have slopes, where btw the slopes are "eaten" from the samples in time length. It is here where your sine emerges, while at first it was a square (but don't forget the monster because of where the sample points are).

If you call this math, then I don't see where it's wrong.
If this isn't math at all, it's also no wrong math.

We now utilize 20 of the 24 bits. 4 more to go for 24dBFS of attenuation without losing anything;
The only thing where the "algorithm" could be off somewhat, is -like I said earlier- that I just use all the bits there are for new level calculation, and with 4 less because of the filtering itself, that level will be less accurate. However, think about that balance which has some truth in it. So, you can have 64 bits for the level if you want, it is still the sampling speed which determines where those levels fall, and although more accurate, no "adjacent" line of level steps of 1 will be found anywhere. So, you could say that 64 bits (or whatever) leads to a more accurate level for where the samples are, but it is a bit of a moot thing because the samples themselves are too few to do it right (to justify the granularity in the levels now).
Filtering is always the product of sampling speed and bit depth but the one with the least granularity determines the accuracy of the product.

To make you feel better, maybe turn this the other way around. So, we skip the filtering but have 4 more bits for better level accuracy. Below you see two sets of pictures; for each set the first unfiltered, and the second filtered/upsampled 8x. First set is 5KHz, second set is 2KHz.
It's a no-brainer what to use of course - but which doesn't imply that digital attenuation is or is not to be used; it only tells that we always want to trade the 4 bits for the filtering. What's left (another 4 bits) and what to do with them, is something else.

Peter

[PeterSt]

PS: I grabbed these older pictures from somewhere, but I think the first of the sets are not right for how the one step emerges into the other. The steps are sure there though, and it is about that.