CLOCKS, what should we look for in next generation

[Miska]

Just don't try to draw out some "NN ps" figure out of it! That would be a meaningless number, one really needs to look at the plot in this case...

 

[jabbr]

31 minutes ago, Miska said:

This is static variable, so not related to jitter. And actually that static skew is a good thing, it makes it perform better! If you want to optimize it, you'd make 1/4th clock period skew between each register - which you could do with a rotator and 4x higher MCLK.

 

It causes skew which is a type of phase error. Agreed the sample passes through all of the elements and its all averaged together. But so is the random phase noise. One gives a path length dependent distribution, and the other a time dependent distribution ... so why is femtosecond clock precision a benefit when the samples are distributed in space and time?

 

Again, maybe this is a benefit because the shift registers don't all load at the same time, ala spread spectrum  

[Ralf11]

can we get a summary for p. 10?

[Miska]

33 minutes ago, jabbr said:

It causes skew which is a type of phase error. Agreed the sample passes through all of the elements and its all averaged together. But so is the random phase noise. One gives a path length dependent distribution, and the other a time dependent distribution ... so why is femtosecond clock precision a benefit when the samples are distributed in space and time?

 

Again, maybe this is a benefit because the shift registers don't all load at the same time, ala spread spectrum  

 

That phase skew is static sub-sample delay. This is actually quite common technique in DACs. The switching transitions get smaller when they are interleaved over time and don't all happen at once.  Just like the one clock cycle long delay between transitions in the register. Phase noise is random delay that changes over time, this skew is fixed and doesn't change over time. Imposing static phase delay doesn't constitute as jitter because it is time-invariant.

 

[barrows]

4 hours ago, Miska said:

With HQPlayer Embedded I'm much less worried since the OS doesn't run or have any other audio applications or sounds. But on Windows or macOS I would be more nervous.

How much less worried?  In my set up I have to run pretty high gain in order to accommodate some low average level files (like Channel Classics, for example), and I have to attenuate the signal by at least -5 dB to accommodate oversampling to DSD and avoid clipping there.  This is definitely a case where I could clip the amp.  I wish I could just run the HQP volume control, as I would prefer that to analog, but the analog volume does give me safety.  I do not want to be replacing (somewhat expensive) drivers...

 

I am using a Muses chip (laser trimmed resistor ladder on a chip VC) with my DSC board right now, I did listen without it (under controlled circumstances) and with it, and it it seems to be really good from a subjective POV (transparent or very close subjectively, i could not A/B it quickly), but still, I am strong believer in digital VC when reliable and possible, and do rely on it with my ESS based DACs, even with high gain.

[fas42]

Digital volume control is the way to go - IME, analogue gain adjustment and trimming is always a weakness, unless the most extreme quality devices can be used.

[Miska]

36 minutes ago, barrows said:

How much less worried?  In my set up I have to run pretty high gain in order to accommodate some low average level files (like Channel Classics, for example), and I have to attenuate the signal by at least -5 dB to accommodate oversampling to DSD and avoid clipping there.  This is definitely a case where I could clip the amp.  I wish I could just run the HQP volume control, as I would prefer that to analog, but the analog volume does give me safety.  I do not want to be replacing (somewhat expensive) drivers...

 

I am using a Muses chip (laser trimmed resistor ladder on a chip VC) with my DSC board right now, I did listen without it (under controlled circumstances) and with it, and it it seems to be really good from a subjective POV (transparent or very close subjectively, i could not A/B it quickly), but still, I am strong believer in digital VC when reliable and possible, and do rely on it with my ESS based DACs, even with high gain.

 

How far are you from full level, so what kind of total attenuations are we talking about?

 

You could alternatively have a resistor divider or just drop the total gain from DAC. But those Muses chips are pretty good, I have couple of DACs with those. It is essentially like an R2R DAC, but instead of reference voltage, you have sound going in.

 

[jabbr]

1 hour ago, Miska said:

Phase noise is random delay that changes over time, this skew is fixed and doesn't change over time. Imposing static phase delay doesn't constitute as jitter because it is time-invariant.

 

Yes the two are different. My point is that if you take 100,000 samples in a Gaussian distribution based on random phase variations, and take 100,000 samples which form something like a Gaussian distribution (only 8 samples) but fixed variations, this variation is all at 22 MHz — how does a few femtoseconds of Gaussian variation outweigh many picoseconds of an approximation of Gaussian variation? 

 

This is is an honest question.

[mansr]

1 hour ago, Ralf11 said:

can we get a summary for p. 10?

Get the latest superclock. Lifted veils, tighter bass, and better dynamics, all can be yours.

[Miska]

3 minutes ago, jabbr said:

Yes the two are different. My point is that if you take 100,000 samples in a Gaussian distribution based on random phase variations, and take 100,000 samples which form something like a guassian distribution (only 8 samples) but fixed variations, this variation is all at 22 MHz — how does a few femtoseconds of Gaussian variation outweigh many picoseconds of an approximation of Gaussian variation? 

 

Delay is not variation and it doesn't show up in the phase variations because it's distribution is 0.

 

Example; just invert the clock signal. That is 180 degree phase shift. Keep it like that throughout the measurement. Does it change the phase noise plot?

 

[jabbr]

Just now, Miska said:

 

Delay is not variation and it doesn't show up in the phase variations because it's distribution is 0.

 

Example; just invert the clock signal. That is 180 degree phase shift. Keep it like that throughout the measurement. Does it change the phase noise plot?

 

 

Right but I mean the reconstructed audio signal — why isn’t a process centered around 22 MHz filtered out?

[barrows]

28 minutes ago, fas42 said:

Digital volume control is the way to go - IME, analogue gain adjustment and trimming is always a weakness, unless the most extreme quality devices can be used.

Agreed, but with DSD there are few options for such, HQP is one, but controlling the volume via computer based software may be risky.  In my case I am talking about running the DAC(s) always at DSD 256 or 512.

[Miska]

Just now, jabbr said:

Right but I mean the reconstructed audio signal — why isn’t a process centered around 22 MHz filtered out?

 

This only changes the filter behavior...

[jabbr]

38 minutes ago, Miska said:

 

This only changes the filter behavior...

The digital FIR but the subsequent analog filter should diminish the 22 MHz +/- component to the point where - 100 dBc/Hz phase error becomes ??? voltage variation ... below any reasonable noise floor ... no?

[PeterSt]

2 hours ago, Miska said:

That phase skew is static sub-sample delay. This is actually quite common technique in DACs. The switching transitions get smaller when they are interleaved over time and don't all happen at once.

 

I like to see chips with skew specifications which are "as is" instead of "40fs min, 60fs max and 45fs typical". Skew itself bears jitter. Only when you use low frequency stuff which doesn't have skew specified, the skew is fixed. In your mind ... 

 

3 hours ago, jabbr said:

Agreed the sample passes through all of the elements and its all averaged together.

 

Averaged ? Maybe if you average it. But in practice it isn't. Think carefully how this works out from input to output and what might happen to your complete sample before the logical (one-sample e.g. 24 bit) transition is completed.

 

Skew is a given but what we can do is "utilize" (exploit) the facts of it (see Miska's quote which could be regarded that).

Skew depends on current usage, which is a (multi)bit not what we want in our AC audio (again Miska's quote).

 

The subject of skew (in the realm of lowest jitter systems) is quite large and the design completely depends on the configuration as a whole which needs to anticipate the physical possibilities (which chips to use) Nothing for a few forum posts. 

[jabbr]

22 hours ago, PeterSt said:

Averaged ? Maybe if you average it. But in practice it isn't. Think carefully how this works out from input to output and what might happen to your complete sample before the logical (one-sample e.g. 24 bit) transition is completed.

 

 

The averaging is done by the post I-V analog LPF filter (but also by the digital FIR in DSC so let's limit this to the post I-V LPF. Perhaps LPF is better terminology than average...

 

Consider the single bit SDM (DSD) signal. The shift register based FIR outputs each supply current before I-V conversion, so 31 bits are averaged.

 

Consider a simple single bit SDM (DSD) signal at 22 mHz which varies 22 Mhz +/- 1 Hz over a second ... and 22 mHz +/- 10 Hz over 1/10 second.

How does this affect the filtered audio signal? Surprisingly little.

 

Now pass this through an LPF with corner frequency of 100 kHz. -40 dB/dec attenuation ... so about -90dB attenuation at 22 Mhz. Now if the +/- 1 Hz components are attenuated 100 - 120 dB to start, or 120 - 130 dB at +/- 10 Hz to start with and then further attenuated 90dB by the LPF how is it physically possible that this is audible?

 

Of course if the signal is at 44kHz the jitter doesn't get filtered out to a ridiculous degree. But jitter at 22 mHz?

 

What specifically/mathematically am I missing?

[rickca]

Oh yeah, I'm following this discussion with perfect comprehension. 🙄

[mansr]

12 minutes ago, jabbr said:

What specifically/mathematically am I missing?

The nonlinear aspects.

[jabbr]

1 minute ago, mansr said:

The nonlinear aspects.

 

That's what it would seem to have to be. There is a well described non-linearity whereby voltage noise results in phase offset noise i.e. a 10 Hz voltage variation in the oscillator supply will result in 10 Hz offset phase noise (e.g. 22 Mhz +/- 10 Hz).

 

The Leeson Effect:

1) http://rubiola.org/pdf-slides/2011T-IFCS-Leeson-effect.pdf

2) https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=7464875

3) https://arxiv.org/pdf/physics/0602110.pdf

 

I have not seen the converse non-linearity described.

[PeterSt]

36 minutes ago, jabbr said:

What specifically/mathematically am I missing?

 

Maybe nothing. But:

 

38 minutes ago, jabbr said:

Consider a simple single bit SDM (DSD)

 

You set your own rules. While I did too:

 

22 hours ago, PeterSt said:

the logical (one-sample e.g. 24 bit) transition

 

and

 

22 hours ago, PeterSt said:

Skew depends on current usage, which is a (multi)bit not what we want

 

So ... I considered to add to it all "this seems to incur for DSD" (1 bit) but I thought to leave that for Miska to add. OK, he didn't ...

 

I don't remember whether it was in this thread, but there has been the question whether DSD would require a lower jitter system (you using the same or similar math to attest it does not), but now it is about skew. And that seems to tell that this puts DSD in favor of PCM ...

[mansr]

13 minutes ago, jabbr said:

That's what it would seem to have to be.

A phase error in the clock results in a corresponding phase error in every frequency of the reconstructed signal. A linear low-pass filter can't do anything about that.

[PeterSt]

32 minutes ago, mansr said:

The nonlinear aspects.

 

In the end maybe that too. Also notice how DSD is not a continuous up-up-up or down-down-down (when the "music" depicts that) - see Scarlet Book, and that this in itself adds non-linearity. This could be nit-picking though.

[jabbr]

7 minutes ago, mansr said:

A phase error in the clock results in a corresponding phase error in every frequency of the reconstructed signal. A linear low-pass filter can't do anything about that.

That would be the inverse of the Leeson effect, which to my knowledge has not been described. Have you seen this described? I'd love a reference.

[PeterSt]

5 minutes ago, mansr said:

every frequency of the reconstructed signal.

 

Hard to wrap my mind around that. Would we still talk of "reconstructing" with DSD as such (and then regarding the "in every frequency") ? not sure. This would be about that other averaging (like over 22 etc. DSD bits). I think this too was discussed earlier on (in this thread ? but maybe over 6 months ago).

[mansr]

10 hours ago, jabbr said:

That would be the inverse of the Leeson effect, which to my knowledge has not been described. Have you seen this described? I'd love a reference.

The Leeson effect has nothing to do with this. I'm talking about a direct mathematical consequence of phase noise in the clock on the reconstructed signal in a D/A converter.