PCM vs DSD

[Sokel]

Did a little comparison and I was wondering if any friend with better gear here has tested this.
Results you're about to see is probably narrowed to the specific DAC but it's interesting to see how it handles things differently.
I will not try to decipher the results other that DSD seems much cleaner but to the cost of elevated noise.It seems like distortion is not hiding in it though,it's only lower.
But I really hope some experienced users try that properly.

1Khz PCM

1Khz DSD256

Multitone 32 PCM

Multitone 32 DSD256

SMPTE 60Hz/7k 4:1 PCM

SMPTE 60Hz/7k 4:1 DSD256


If any friend here wants any other test please let me know.

 

[MaxwellsEq]

Are you able to describe your methodology?

 

[Sokel]

Are you able to describe your methodology?

Same settings (shown at the left of the chart) same DAC,same ADC,I touched nothing during the recordings.
I trust signals are perfect as internally generated by Multitone Analyzer.
The only thing I have drawbacks about is level (anything above 1.5-2V is too hot for my ADC),as DSD can get a huge penalty if attenuated but I trust this has been taken cared of by Multitone.


Edit:
As for the order:
PC (no isolators and stuff,I prefer real life conditions) > Multitone Analyzer > Khadas tone1 Dac > 2m cables as I want it to be as-is > E-MU 0204 ADC > Multitone analyzer >PC.

 

[Sokel]

As we are used to see 1Khz measurements most of the time I thought it would be interesting to see a little lower so to compare.
I admit I was not expecting that:

100Hz PCM

100Hz DSD

both for comparison (Green is PCM - Yellow DSD)

DSD is much much cleaner down there and PCM's grass is visible since is not hiding in the higher than 20Khz limit of usual measurements.
Probably inaudible in that levels but still...

Also there's no excuse about mastering,signal in both cases is generated by Multitone.
Folks with measuring stuff (lots here) can try with better DACs (Khadas is not exactly on top with it's 110db 1Khz SINAD),I'm curious about the results.

 

[AnalogSteph]

It might be interesting to see whether this phenomenon relates to the "ESS hump" at all; if memory serves the "grass" at high levels was a sign of that. If so, results at -30ish dBFS out may be enlightening.

 

[Sokel]

It might be interesting to see whether this phenomenon relates to the "ESS hump" at all; if memory serves the "grass" at high levels was a sign of that. If so, results at -30ish dBFS out may be enlightening.

So,measure them both (PCM - DSD) at -30dbFS?

First thing tonight!

 

[Sokel]

So,@AnalogSteph you nailed something on the head here.

First of all that's Amir's measurement where the hump is evident:

So:

IMD vs level PCM 96K.

IMD vs level DSD128.


PCM hump - DSD no hump.
So,what's different about DSD processing in the DAC?
What's more or missing?

Interesting indeed.

Edit: As a friend pointed out to me in PM that the moded version (grounded,externally powered) will have little to no use to other owners,I remeasured it stock,as it came at first and adjusted the post accordingly.
Differences are far more evident,we're talking about 20db lower intermodulation distortion at some levels when playing DSD.

 

[KSTR]

So,what's different about DSD processing in the DAC?

My educated guess:

In PCM mode, the ESS chip first does the usual upsampling and digital filtering, and (sometimes) resampling. This PCM signal is then input to the Delta-Sigma modulator which is PCM in and PCM out, the output only having 6 bits for the final DAC. To get the resolution, dithering with noise shaping is applied, with the output strategically bouncing about +- 2 LSB's. The exact span depends on the current input amplitude and how it maps to the 6bit output transitions, giving a signal-modulated RF glitch energy, which affects the I/V-process -- even for a perfect I/V stage, it seems.

In DSD mode, the signal is 1 bit only and already has the noise shaping applied to it. The final DAC appears to be fed directly with a (even asynchronously) resampled version of that 1-bit stream and the output cells are now switched between large signals on average, not just a few LSB's. This de-correlates the glitch energy from the signal and thus reduces its impact -- a bit of noise is traded for reduced distortion.

At least that's my modelling of what I see going on with the 9038q2m, actual behavior may or may not be completely different. ;-)

 

[Sokel]

My educated guess:

In PCM mode, the ESS chip first does the usual upsampling and digital filtering, and (sometimes) resampling. This PCM signal is then input to the Delta-Sigma modulator which is PCM in and PCM out, the output only having 6 bits for the final DAC. To get the resolution, dithering with noise shaping is applied, with the output strategically bouncing about +- 2 LSB's. The exact span depends on the current input amplitude and how it maps to the 6bit output transitions, giving a signal-modulated RF glitch energy, which affects the I/V-process -- even for a perfect I/V stage, it seems.

In DSD mode, the signal is 1 bit only and already has the noise shaping applied to it. The final DAC appears to be fed directly with a (even asynchronously) resampled version of that 1-bit stream and the output cells are now switched between large signals on average, not just a few LSB's. This de-correlates the glitch energy from the signal and thus reduces its impact -- a bit of noise is traded for reduced distortion.

At least that's my modelling of what I see going on with the 9038q2m, actual behavior may or may not be completely different. ;-)

You're one of the few who could know and thanks for that.

 

[voodooless]

My educated guess:

In PCM mode, the ESS chip first does the usual upsampling and digital filtering, and (sometimes) resampling. This PCM signal is then input to the Delta-Sigma modulator which is PCM in and PCM out, the output only having 6 bits for the final DAC. To get the resolution, dithering with noise shaping is applied, with the output strategically bouncing about +- 2 LSB's. The exact span depends on the current input amplitude and how it maps to the 6bit output transitions, giving a signal-modulated RF glitch energy, which affects the I/V-process -- even for a perfect I/V stage, it seems.

In DSD mode, the signal is 1 bit only and already has the noise shaping applied to it. The final DAC appears to be fed directly with a (even asynchronously) resampled version of that 1-bit stream and the output cells are now switched between large signals on average, not just a few LSB's. This de-correlates the glitch energy from the signal and thus reduces its impact -- a bit of noise is traded for reduced distortion.

At least that's my modelling of what I see going on with the 9038q2m, actual behavior may or may not be completely different. ;-)

That’s probably not correct. From what can be seen from diagrams of the ESS chips, the internal path is always PCM. DSD gets converted to PCM with the same sample rate (by just expanding the bits and applying a selectable low pass), and then the upsampling steps are just skipped. The various chips have some variation here, but they all work in a similar manner.

AKM chips are different. There you can actually ship all the PCM conversion and processing and feed 1-bit to the modulators.

Also: how does one do asynchronous DSD resampling?

 

[BR52]

That’s probably not correct. From what can be seen from diagrams of the ESS chips, the internal path is always PCM. DSD gets converted to PCM with the same sample rate (by just expanding the bits and applying a selectable low pass), and then the upsampling steps are just skipped. The various chips have some variation here, but they all work in a similar manner.

AKM chips are different. There you can actually ship all the PCM conversion and processing and feed 1-bit to the modulators.

Also: how does one do asynchronous DSD resampling?

Thanks this was the information I was looking for. I was not sure about it and could not find the time to figure it out. Better i ask earlier.

 

[Sokel]

I've been asked to measure how it does with native DSD (no DoP),so here it is played through foobar:
(Archimago's files,thanks for that)

Noise and SINAD is better than DoP and that with attenuation (from DAC's side via Ian's Canada controller),I expect to be better without it.

 

[Sokel]

That’s probably not correct. From what can be seen from diagrams of the ESS chips, the internal path is always PCM. DSD gets converted to PCM with the same sample rate (by just expanding the bits and applying a selectable low pass), and then the upsampling steps are just skipped. The various chips have some variation here, but they all work in a similar manner.
View attachment 297279


AKM chips are different. There you can actually ship all the PCM conversion and processing and feed 1-bit to the modulators.

Also: how does one do asynchronous DSD resampling?

I tried to find the same for the chip under test here but it's a little blur with all the bypass modes (by choice) that I can't understand when it's applied and when not.

 

[voodooless]

Why would there be a difference between DoP and DSD? DoP is just a transport mechanism for backwards compatibility.

 

[Sokel]

Why would there be a difference between DoP and DSD? DoP is just a transport mechanism for backwards compatibility.

You're right,it shouldn't.
But the test shows otherwise.

As I often say,it comes down how the device handles it,that's my only explanation.

Edit:Here's a comparison,look at the noise floor.

 

[MRC01]

... In PCM mode, the ESS chip first does the usual upsampling and digital filtering, and (sometimes) resampling. This PCM signal is then input to the Delta-Sigma modulator which is PCM in and PCM out, the output only having 6 bits for the final DAC. To get the resolution, dithering with noise shaping is applied, with the output strategically bouncing about +- 2 LSB's. The exact span depends on the current input amplitude and how it maps to the 6bit output transitions, giving a signal-modulated RF glitch energy, which affects the I/V-process -- even for a perfect I/V stage, it seems.
...

That’s probably not correct. From what can be seen from diagrams of the ESS chips, the internal path is always PCM. DSD gets converted to PCM with the same sample rate (by just expanding the bits and applying a selectable low pass), and then the upsampling steps are just skipped. The various chips have some variation here, but they all work in a similar manner. ...

The ESS hump was eliminated (or reduced) with changes to the I/V conversion. Yet that doesn't mean the root cause is in I/V conversion. The I/V conversion changes that eliminate the hump could be a band-aid to a root cause further upstream in the chip. Which is essentially what @KSTR says above. Now whether that root cause is signal modulated RF glitch energy, or something else, who knows? But @KSTR's educated guess seems plausible to me.

@voodooless if not this, what do you think causes the ESS IMD hump? And why/how can the hump be reduced with changes to the I/V conversion?

On a related note, what do you think about DAC chips that do the I/V internally and provide voltage output pins? Like the Wolfson/Cirrus 8741?

 

[Sokel]

The ESS hump was eliminated (or reduced) with changes to the I/V conversion. Yet that doesn't mean the root cause is in I/V conversion. The I/V conversion changes that eliminate the hump could be a band-aid to a root cause further upstream in the chip.

That's probably true,I followed the fix thread and it seems that the energy only went into the noise and thus not visible.
That's not elimination but reduction.

 

[Sokel]

I did a comparison with lots of averages to reveal every little spike so we can see that energy:



50Hz DSD



50Hz PCM

So,the energy IS there but greatly reduced,seems like DSD does something similar with the fix applied in I/V stage.

(all that is WAY above my pay grade,interpreting the results is a shot in the dark for me )

 

[voodooless]

@voodooless if not this, what do you think causes the ESS IMD hump? And why/how can the hump be reduced with changes to the I/V conversion?

I have absolutely no idea

On a related note, what do you think about DAC chips that do the I/V internally and provide voltage output pins? Like the Wolfson/Cirrus 8741?

I think in general, if you want the very best performance, a current output DAC is the way to go. That does not mean voltage output DACs are bad. Many come really close to SOTA performance. It’s usually the more budget friendly option as well. Chips are cheaper, and you don’t need IV conversion. Usually the designs should be easier to implement.

 

[voodooless]

I did a comparison with lots of averages to reveal every little spike so we can see that energy:

View attachment 297298

50Hz DSD


View attachment 297303
50Hz PCM

So,the energy IS there but greatly reduced,seems like DSD does something similar with the fix applied in I/V stage.

(all that is WAY above my pay grade,interpreting the results is a shot in the dark for me )

The main difference here is obviously the filtering. I wonder if the PCM noise spectrum would be different with different filter selections. Which one did you use in this case?