Jump to content
IGNORED

New Berkeley DAC? This article implies as much but ... ?


Recommended Posts

Interesting article here:

 

pcm dsd

 

Any of the engineer types care to comment or enlighten us as to the article's insights / mistakes?

 

I think it's bad idea with multi-bit DACs. I have two noise shapers optimized for multi-bit PCM DACs, but those are not that aggressive. The increase of ultrasonics I use is only about 10 dB compared to TPDF, but it already gains about 40 dB extra DNR in audio band when combined with enough oversampling. So the operation stays around 2 LSBs.

 

My usual recommendation for choosing output wordlength is to pick one that corresponds to the DACs +-0.5 dB linearity range.

Signalyst - Developer of HQPlayer

Pulse & Fidelity - Software Defined Amplifiers

Link to comment
Evidence is though there's still significant noise modulation probably due to the difficulty of constraining the dither within a feedback loop into the required probability density function (TPDF).

 

If you dissect practically any S-D DAC datasheet (I've done this with AKM and ADI parts) you can't fail to miss the fact that the noise at full scale output goes up when compared to the -60dBfs stimulus - chances are Lynn's full-on dither will have the same effect, reducing the dynamic range significantly. The argument that 'there is dynamic range' to spare doesn't really wash when this effect is taken into account - for example the AKM part I looked at gives around -105dB noise putput at full scale, but -120dB at -60dBfs stimulus. -105dB is under 18bits of dynamic range.

 

To avoid noise modulation, it needs careful design of the modulator and dithering. Lower the modulator order, more susceptible it is to noise modulation. Noise modulation in D-S usually happens specifically due lack of dither. Keeping noise modulation away has been one of the most time consuming tasks when I design modulators.

Signalyst - Developer of HQPlayer

Pulse & Fidelity - Software Defined Amplifiers

Link to comment

Finally onto Lynn's 3rd paragraph -

 

'Spread spectrum' I'm currently unaware of as a way to improve performance of classD amps, though perhaps someone more knowledgeable on classD can enlighten me. Where I have come across this term is in EMC testing - if you jitter your clock (computer motherboards sometimes have this feature in the BIOS) then it looks to the usual measuring kit for EMC that you're radiating less. However you're not - you can pass EMC testing this way but its only a quirk of how EMC emissions are measured - you're distributing your emissions across more frequencies so there's less to see at any one particular frequency.

 

As for 'randomizing slewing' its a totally new concept to me and I can't see any advantage of it. There will be no 'uniform noise floor' from the presence of out-of-band signals interacting with amp non-linearities, rather noise modulation is the result - the noise floor most certainly does vary with signal level. That's one of the biggest reasons why traditional measurements (THD+N in particular) don't correlate well with sound quality, whereas 'noise loading' does fairly well.

 

So yes, this particular designer is indeed throwing up his hands in horror, but not quite for the reasons Lynn expects.

 

I rather reckon that the 'analog like smoothness' of DSD is a combination of factors, one being the noise modulation of DSD (uncontested fact from Stanley Lipshitz's 2000 paper) is somewhat similar in character to that of analog tape. DSD DACs are inherently low glitch so this is a contributor to 'smoothness'. Lynn's perfectly right about most analog electronics post-converters being too slow - perhaps his factor of 100X is rather an underestimate though. If it were true then rather than a 10MHz opamp (as recommended by TI for the PCM179x range) we should be using 1GHz ones. However I've tried a 2GHz GBW post-amplifier and still it wasn't enough without additional passive filtering on its input.

 

@Miska - how is it you are able to solve noise modulation issues in your modulators but ESS was unable to in their ES9018 flagship chip? Do you have some secret sauce or is it just your criteria for acceptable noise modulation are less stringent? How do you go about testing for it?

Link to comment
I rather reckon that the 'analog like smoothness' of DSD is a combination of factors, one being the noise modulation of DSD (uncontested fact from Stanley Lipshitz's 2000 paper) is somewhat similar in character to that of analog tape.

 

DSD can be made without noise modulation...

 

@Miska - how is it you are able to solve noise modulation issues in your modulators but ESS was unable to in their ES9018 flagship chip? Do you have some secret sauce or is it just your criteria for acceptable noise modulation are less stringent? How do you go about testing for it?

 

I think they claim they have also solved it. I don't want to go into all the detail, but as I said it's mainly about modulator order and dithering and certain other design parameters. One reason wht they may have noise modulation is that if they are not using full output code space of the multi-bit modulator all the time, but DSD has to...

 

My acceptance criteria for most things is that it is undetectable to me. I use constantly changing signals specifically designed to test it. Your claim about different noise level at -60 and -0 dB is of course easy to test with. Here's a logarithmic level sweep in DSD64 from -200 dB to 0 dB. You can see that the noise level stays constant and the noise is random.

ramp.png

 

Multi-bit PCM has a severe problem with strongly correlated image frequencies at every multiple of the converter sampling rate. Since the sampling rates are lower, those pose much bigger problems. In addition, the glitch/overshoot level of the PCM converter may not exceed level of ½LSB which becomes significant problem on increasing the converter sampling rate to fight image frequencies.

Signalyst - Developer of HQPlayer

Pulse & Fidelity - Software Defined Amplifiers

Link to comment
DSD can be made without noise modulation...

 

Well this is significant news - I was previously unaware that Stan LIpshitz's math (in the paper I cited) had errors. What was his oversight?

 

 

 

Here's a logarithmic level sweep in DSD64 from -200 dB to 0 dB. You can see that the noise level stays constant and the noise is random.

[ATTACH=CONFIG]9848[/ATTACH]

 

I'm unclear what I'm looking at here, so please talk me through how this shows zero noise modulation?

Link to comment
What I like about in dCS is that they actually design their own DACs. There are not that many such companies. Meitner, Playback Designs and MSB comes to mind. Others?

 

I also liked the increadible sound and detail collected from the humble cd...

Simply amazing...

 

Still, not sure if they could not achieve the same result by going the integrated route...they really love their customers money...

Link to comment
Well this is significant news - I was previously unaware that Stan LIpshitz's math (in the paper I cited) had errors. What was his oversight?

 

IIRC, choosing oversimplified modulator design with lack of dither. If I wanted to make DSD look bad I could make a modulator that looks really bad. His paper has been challenged by quite a bunch of papers by others. And the people who know how to fix things definitely don't write out the details on papers but keep it secret.

 

I'm unclear what I'm looking at here, so please talk me through how this shows zero noise modulation?

 

It is 1 kHz tone level-swept from -200 dBFS to 0 dBFS and modulated into 2.8 MHz DSD and then converted back to 32-bit PCM for spectrum analysis. You can see that

1) Noise stays random regardless of the input signal

2) Noise level stays constant regardless of input signal level

 

This test is by no means challenging. This is full digital domain analysis. But similar test can be used with ADCs too, and some actually do show noise modulation due to lack of dither. PCM4202 ADC has dither, so it is not as susceptible to noise modulation as some others.

Signalyst - Developer of HQPlayer

Pulse & Fidelity - Software Defined Amplifiers

Link to comment
IIRC, choosing oversimplified modulator design with lack of dither.

 

But his argument was (IIRC) that its impossible to dither because there's no room - the modulator's overloaded with only a single bit.

 

If I wanted to make DSD look bad I could make a modulator that looks really bad.

 

So you're saying that's what he did? Set out to discredit DSD? Otherwise why raise this possibility?

 

His paper has been challenged by quite a bunch of papers by others.

 

So I've heard - I've even looked at some of the 'challenges'. None I've seen overturn the math they just change the subject....

 

And the people who know how to fix things definitely don't write out the details on papers but keep it secret.

 

But if SL is right, there's no way to fix it. And if he's wrong, still no evidence of that.

 

 

It is 1 kHz tone level-swept from -200 dBFS to 0 dBFS and modulated into 2.8 MHz DSD and then converted back to 32-bit PCM for spectrum analysis. You can see that

1) Noise stays random regardless of the input signal

2) Noise level stays constant regardless of input signal level

 

I can see 2) yeah but don't see how to work out 1). Also there seems to be some windowing evident around the 1kHz tone - it shows up with considerable width in the freq domain - plenty of places for noise to be hiding in the skirts.

 

 

This test is by no means challenging.

 

So you're happy to say 'My DSD has no noise modulation because I don't challenge it'. Like Peter Schiff's 'a mouse on a suspension bridge' for the banks' 'stress tests' a few years back. I get the picture..:P

Link to comment
But his argument was (IIRC) that its impossible to dither because there's no room - the modulator's overloaded with only a single bit.

 

That's wrong, DSD is not 1-bit PCM so the entire dithering scheme is completely different.

 

So you're saying that's what he did? Set out to discredit DSD? Otherwise why raise this possibility?

 

It is clearly intended to discredit DSD.

 

So I've heard - I've even looked at some of the 'challenges'. None I've seen overturn the math they just change the subject....

 

I just checked the "math" offered by the paper which is very scarce and only for simplified 1st order modulator case.

 

I can see 2) yeah but don't see how to work out 1). Also there seems to be some windowing evident around the 1kHz tone - it shows up with considerable width in the freq domain - plenty of places for noise to be hiding in the skirts.

 

I think I used Blackman window. Since noise modulation would affect much wider bandwidth than width of the main lobe in the plot, it's nitpicking... Making a spectrogram plot of a million point FFT I also to to inspect my test signals, would mean image with over 524289 pixel vertical resolution is a bit too much for this site...

 

So you're happy to say 'My DSD has no noise modulation because I don't challenge it'. Like Peter Schiff's 'a mouse on a suspension bridge' for the banks' 'stress tests' a few years back. I get the picture..:P

 

You said multi-bit SDM DACs show different noise levels for -60 dB and 0 dB output signals and also the paper says it. Here you see output from -200 to 0 dB. Instead of just making two plots at two levels, I logarithmically swept the level over wider range.

 

One reason for multi-bit SDM to exhibit this noise modulation behavior is that in case the modulator doesn't all the time produce produce flat 0 dBFS output codes, the noise level will inevitably change when more levels kick in as function of input signal level... Especially with designs like ESS that try to utilize maximum modulation depths in order to reach highest DNR figures!

 

To some extent this is also likely to happen with PCM ladder DACs because percentage of thermal current noise of the total Iout varies as function of signal level while I/V stage self-voltage noise stays constant..

Signalyst - Developer of HQPlayer

Pulse & Fidelity - Software Defined Amplifiers

Link to comment
That's wrong, DSD is not 1-bit PCM so the entire dithering scheme is completely different.

 

Do go on to explain why Litshitz is wrong then. I'm curious to hear.

 

I just checked the "math" offered by the paper which is very scarce and only for simplified 1st order modulator case.

 

Lipshitz is a mathematician, a professor even - why the "scare" quotes? Intending to discredit him by any chance?

 

You said multi-bit SDM DACs show different noise levels for -60 dB and 0 dB output signals and also the paper says it. Here you see output from -200 to 0 dB. Instead of just making two plots at two levels, I logarithmically swept the level over wider range.

 

You only show up to around 85kHz - why the cut-off there? Why is what's found with multi-bit DACs relevant here?

 

One reason for multi-bit SDM to exhibit this noise modulation behavior is that in case the modulator doesn't all the time produce produce flat 0 dBFS output codes, the noise level will inevitably change when more levels kick in as function of input signal level...

 

Yes sure, but I don't see that it follows that the in-band noise has to change. After all the noise shaping loop is working to push noise out, up to higher freqs. If the designers didn't want that effect, presumably they were free to choose a single bit modulator. Does the fact that they didn't tell us anything about their design priorities?

 

To some extent this is also likely to happen with PCM ladder DACs because percentage of thermal current noise of the total Iout varies as function of signal level while I/V stage self-voltage noise stays constant..

 

The extent must be a whole lot smaller as I've not seen anything of the magnitude (15dB in the case of AKM for example) with PCM DACs. But if you have a DS to share that I've missed, by all means draw my attention to it.

Link to comment
Do go on to explain why Litshitz is wrong then. I'm curious to hear.

 

I don't have time, nor care.

 

Intending to discredit him by any chance?

 

About as much as you were trying to discredit me by referring to a single paper that is way over decade old... The only one you found to support your view?

 

How about putting out some of your own research material instead of referring to random pieces of other people's work? At least I do my own homework and don't take anybody's claims as granted, professor or not. I know number of professors who furiously disagree with each other.

 

You only show up to around 85kHz - why the cut-off there?

 

Because it was converted to 176.4/32 PCM for plotting out the graph. Of course I could have used wider bandwidth, but then anything within around 20 kHz would have been a thin band.

 

Anyway, here's different kind of view of halfway through the same signal.

dsd-1k.png

 

Yes sure, but I don't see that it follows that the in-band noise has to change. After all the noise shaping loop is working to push noise out, up to higher freqs. If the designers didn't want that effect, presumably they were free to choose a single bit modulator. Does the fact that they didn't tell us anything about their design priorities?

 

It's not visible at higher frequencies either because it is randomized. And those high frequencies are removed by reconstruction low-pass filter in the D/A section. Output reaches -120 dB level typically between 200 - 400 kHz. While many multi-bit R2R ladder converters with 8x oversampling are only -60 dB down at the same point with their strong images and still above -100 dB at 5 MHz.

 

The extent must be a whole lot smaller as I've not seen anything of the magnitude (15dB in the case of AKM for example) with PCM DACs. But if you have a DS to share that I've missed, by all means draw my attention to it.

 

It is just because most of the R2R ladder datasheets have barely any plots compared to delta-sigma ones. :D

 

But you get some indication from the bumpy THD vs level plot there is in the PCM1704 datasheet. And the linearity manages to 16-bit level within 1 dB while PCM1795 manages to 20-bit level within 1 dB. But I can squeeze some more with oversampling and PCM noise shaping.

 

But that's anyway one of the reasons why I run software modulators and don't rely on hardware ones. In software I can afford more processing power, already DSD64 stereo is around 3.2 billion calculations per second and that's just around 25% of CPU load on my old desktop machine. :)

Signalyst - Developer of HQPlayer

Pulse & Fidelity - Software Defined Amplifiers

Link to comment
Do go on to explain why Litshitz is wrong then. I'm curious to hear.

 

He's wrong because he uses oversimplified first order modulator model and then assumes that it applies to all other models. Overall, delta-sigma modulators are not dithered same way as PCM, this is recognized even in his paper, although he doesn't go any deeper on relation of the modulator and dither designs.

 

But there are others who have written hundreds of pages about the topic before and after his paper... Audio is just one of the niches of delta-sigma modulators, those are used across the board in all kinds of applications.

 

If we take same modulator and just increase rate to DSD256, this is the result:

ramp-256-2.png

 

Other authors have clearly said that noise modulation is primarily problem of modulators order of 3 or lower.

 

But yes, I'm getting tired of arguing with people who are trying really hard to find things that are not perfect with DSD while completely ignoring all the PCM problems and the fact that almost all audio converter chips are delta-sigma design with limited computational resources.

 

Bring a multi-bit SDM DAC to market that can take multi-bit SDM in and I'll support it, no problem. So far input choices are a) DSD or b) PCM.

 

But please don't show me any of the PCM-input delta-sigma DAC chips that use 8x OS filter + S/H or linear interpolation.

Signalyst - Developer of HQPlayer

Pulse & Fidelity - Software Defined Amplifiers

Link to comment

You are correct - it is pointless because Miska is essentially correct in every point he has presented here, and you are taunting him and taking advantage of his good nature. Not many people are willing to share tech information as he is.

 

I am not saying he is always correct, but he always tries to back up his ideas with hard data. And doesn't take offense when someone needs a bit more explanation to understand. Even when perhaps, he should.

 

-Paul

 

 

Continuing this is pointless as I note you've become defensive, personalizing the discussion. Also I'd be going counter to Upton Sinclair's dictum.

Anyone who considers protocol unimportant has never dealt with a cat DAC.

Robert A. Heinlein

Link to comment
So perhaps you could explain - why he's engaged in messenger shooting (both myself and Stan Lipshitz)?

 

Would you please provide something new and more comprehensive, that one short paper has been discussed to death in past 13 years.

 

Every time someone wants to smash DSD, they go to wikipedia and then paste link to that same paper. There has been countless messengers repeating that same message. And no matter what, that one piece of paper is regarded as ultimate truth about everything DSD and nothing else needs to be taken into consideration.

Signalyst - Developer of HQPlayer

Pulse & Fidelity - Software Defined Amplifiers

Link to comment
Would you please provide something new and more comprehensive, that one short paper has been discussed to death in past 13 years.

 

Happy to go into depth with the paper when you single out specific points of SL's where he's made an error. You'll need to quote chapter and verse as so far your responses have been characterized by bluster, hand-waving, obfuscation and deflection onto irrelevances (as in the text below which I shall now snip).

Link to comment

Please - either you are trying to taunt me or else you failed to read any of his responses. On the off chance you are just in some kind of emotional state, why do you not simple address the quote below?

 

I looked at it and I see reasonable justification for this point of view. If you do to, then post your reasons why you - personally - do not.

 

He's wrong because he uses oversimplified first order modulator model and then assumes that it applies to all other models. Overall, delta-sigma modulators are not dithered same way as [/Quote]
PCM, this is recognized even in his paper, although he doesn't go any deeper on relation of the modulator and dither designs.[/Quote]

 

And the evidence for this claim is where?

Anyone who considers protocol unimportant has never dealt with a cat DAC.

Robert A. Heinlein

Link to comment
Please - either you are trying to taunt me or else you failed to read any of his responses.

 

Both options are wrong. Why personalize this though - its a discussion about a technical matter, there is absolutely no need for any personality stuff to enter into it. But so far you have rushed to the defense of Miska as a person whilst I've not been attacking him personally, merely explaining why his arguments don't hold water.

 

On the off chance you are just in some kind of emotional state, why do you not simple address the quote below?

 

When Miska provides the reference to exactly where SL does what he's alleged to be doing, preferably directly pasting in the text from his paper then I'll take note.

Link to comment

Hi Miska:

 

I am trying hard to keep up but I do get lost.

Quick off-topic question:

If you had a 24bit/24.xMHz parallel PCM stream (upsampled from Redbook with a quality filter) and were going to DSM it to a 6 bit/7 level format, what order of modulator would you choose?

(There is a reason for my question, but I'll not post that here just yet.)

 

Thanks,

Alex C.

Link to comment

OIC - you are willing to criticize and dispute what he thinks about a the math in a paper *you* posted a link too, but don't want to defend the paper yourself?

 

Perhaps you have not been attacking - in your mind - but it sure is coming off exactly like you are doing just that. However, I am certainly willing to give you the benefit of the doubt and say I might have misjudged *you*.

 

So try again, go back into the paper *you* posted a link too, look at the math, and then tell us why you dispute Miska's opinion of the same.

 

Bigger man, take the first steps, and all that stuff. Besides, since I rather agree with Miska's interpretation of that, if not necessarily with all his other conclusions, I am interested in what you think.

 

Paul

 

 

Both options are wrong. Why personalize this though - its a discussion about a technical matter, there is absolutely no need for any personality stuff to enter into it. But so far you have rushed to the defense of Miska as a person whilst I've not been attacking him personally, merely explaining why his arguments don't hold water.

 

 

 

When Miska provides the reference to exactly where SL does what he's alleged to be doing, preferably directly pasting in the text from his paper then I'll take note.

Anyone who considers protocol unimportant has never dealt with a cat DAC.

Robert A. Heinlein

Link to comment
OIC - you are willing to criticize and dispute what he thinks about a the math in a paper *you* posted a link too, but don't want to defend the paper yourself?

 

Incorrect - he has only waved his hand in the direction of what he thinks the paper is saying so far. Rather than take issue with the text of the paper. Until he takes up and addresses the text, there's nothing to defend.

 

Perhaps you have not been attacking - in your mind - but it sure is coming off exactly like you are doing just that.

 

Incorrect again in the first claim - but I accept your second claim. In your mind the words I've typed on CA are indeed being interpreted as an attack. The falsity of your first claim is demonstrated by your not taking up my requests to provide evidence. If there was evidence in the text I've typed of my attacking, surely you'd be able to find it and point to it, no? So the reasonable conclusion is - the attacks are solely in your (and also it seems, in Miska's) mind and not in the text.

 

However, I am certainly willing to give you the benefit of the doubt and say I might have misjudged *you*.

 

Thank you - appreciated.

 

So try again, go back into the paper *you* posted a link too, look at the math, and then tell us why you dispute Miska's opinion of the same.

 

No thanks, already explained why not.

 

Bigger man, take the first steps, and all that stuff.

 

Now you're reverting to playing the man rather than the ball. Let's not go there again.

 

Besides, since I rather agree with Miska's interpretation of that, if not necessarily with all his other conclusions, I am interested in what you think.

 

Sorry - you're interested in what I think of what? In the main I'm not interested in saying what I think, I am interested in giving observations though. One of my favourite sayings comes from Dirty Harry 'Opinions are like a**holes, everybody's got one'. if you'd like my observations then by all means ask.

Link to comment

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now



×
×
  • Create New...