DEM (Dynamic Element Matching) - might it be audible?

[nanook]

I'd like to share a thought:

It's about DEM (dynamic element matching) in DACs.
The AK4499EX e.g. makes heavy use of this technique in "measurement mode". It's turned off (or reduced to a level that didn't get visible in my measurements) in "sound quality mode".
In "sound quality mode" measured performance is not up to the figures you obtain in "measurement mode" and the datasheet explicitly addresses this.

When measuring harmonic distortion or a multitone spectrum, the measurement time usually is long such that a lot of periods get averaged.

But what's happening short term? How big is the effect short term (for different DACs)?

Might we able to hear these short term (maybe even periodic) deviations in a signal due to the DEM process in case those (periodic?) deviations are big enough?
What would be "big enough" be and how would we measure this?

I assume that in a modern silicon process the matching of the current sources (without DEM) is still excellent and the performance still way beyond our hearing capabilitiy.

What do you think?

 

[Scytales]

Hi!

I cannot find by myself any reference about the settings you talk about in the AK4499ES's datasheet dated from 02/2019. Can you point to the relevant section of the datasheet or another release of the datasheet I am not aware of that would tackle the issue?

 

[nanook]

I was mentioning the AK4499EX.

https://e2e.ti.com/cfs-file/__key/communityserver-discussions-components-files/6/ak4499exeq_2D00_en_2D00_datasheet_2D00_myakm.pdf

Section 9.6 "Sound control".
"The AK4499EX has a function to control the sound quality by setting SC pin or SC bit (Table 9). The
analog characteristics of the AK4499EX are specified in Setting 1, and the characteristics are not guaranteed at Setting 2"

Table 9. Sound Quality Select Mode
SC pin (bit) Sound
L (0) Measurement Mode (Setting 1)
H (1) Sound Quality Mode (Setting 2)

 

[Scytales]

Thanks!

 

[nanook]

I have to apologize - I'm not sure if the effect that I observed is due to DEM or DAW.
I have to do some reading....

 

[Scytales]

After reviewing the datasheet you provided, I found no mention of the internal circuitry that the SC pin is supposed to control.

If you could share your measurements, it might help us understand the function of this setting.

 

[nanook]

The SMSL DO300EX seems not to have a higher order filter in or behind the I/V-converter. When operated at 5.6MHz (especially at Sound Color=1), there's quite some RF at the XLR outputs.
Stimulus: 1kHz -40dBFS from REW.
Signal at XLR output amplified by E1DA APU at 34dB gain.

The RCA outputs include an additional LP-filter that helps to remove some of the RF.

I assume that the 4 sound colors of the DO300EX are including the "Sound Quality" register mentioned above. The DO300EX documentation just reveals that Sound Color 1 and 2 are running at ca. 5.6MHz and SC 3 and 4 at 11.2MHz:

Since I didn't find any other register that would make sense in the context of "Sound Color", I concluded the other bit (-> SC=2 and SC=4) is the setting described as "Sound Quality" in the AK4499EX datasheet and that SC=1 and SC=3 refer to "Measurement" mode.
This is an assumption.

The spectrum at the XLR output shows a significant different distribution for the SoundColor settings of the DO300EX (spectrum with a tone at - 180dBFS. Without data stream the outputs get muted, so I played something close to nothing). Noise shaping seems to be different between SC=1 and SC=2.

I want to point out, that this "RF-noise" with SC=1 is quite visible on the XLR output because of the particular design of the DO300EX. As far as I could measure without desoldering components SMSL did not implement the filter proposed in the AK4499EX datasheet. With a nicely implemented 2nd order filter in the I/V-stage this would look a lot friendlier.

The signal at the XLR outputs measured in the time domain (note the different y-scale):
Sequence: SC = 1 (200mV/Div) / SC = 2 (100mV/Div) / SC = 3 (20mV/Div) / SC = 4 (20mV/Div)

Sawtooth pattern along time axis:
I have checked the output with an old storage oscilloscope and I was surprised to see a sort of sawtooth pattern in the Sigma-Delta modulation pattern when choosing SC=1 (5.6 MHz, assumed mode: "Measurement" mode). With SC=2 (5.6MHz, assumed "Sound Quality" mode), this pattern was not visible as far as I remember.
As far as I remember, this sawtooth pattern did not depend on the sample rate. Sample rate related processing seems to be handled by the AK4191.
The period of this pattern did scale perfectly with the "volume" setting. Decreasing volume (volume knob of the DO300EX) by 20dB did slow down the pattern by a factor of 10. Going to -40dB volume setting made another factor of 10 on the timescale. Going further down with the volume setting this pattern vanished completely.

Sequence: Volume = -99dB / -40dB / -20dB / 0dB

Please apologize the bad quality of the DSO screenshots - scope frame rate and exposure time just didn't want to "cooperate".

Unfortunately I didn't dig deeper into this observation because I decided to return the DO300EX. I made a listening session against a SMSL D-6s (calibrated to ca. 0.1dB level matching) and I ended up with "guessing".
--> Both of theses DACs are absolutely flawless sound-wise
I did try to hear differences switching the SoundColor modes - rather guessing than hearing

I did measure multitone spectra for the SoundColor settings, but I'm not sure if I can really trust them because of the amount of RF that may cause spurious tones on the ADC side. I should have used a LP-filter in front of the ADCiso, but i cannot repeat the measurements. The good news is - all of these look perfectly fine with SC=1 looking a tad nicer.

SC=1 SC=2
SC=3 SC=4

In a twin-tone measurement (41Hz, 89Hz), the differences got a bit more visible (SC=1 and SC=4 measured only).

I'm afraid I have to dig into the complex topic of DEM, DWA and Noise-Shaping a bit deeper in order to get a proper clue what the "sawtooth pattern" I observed might be.

 

[NTTY]

Hi,

5.6M or 11M refer to the oversampling rate of the DAC’s delta sigma modulator (128x or 256x). The higher number will better spread the quantization noise of the delta sigma modulator far in frequency, and lower it in amplitude.
For the two other options, my guess would be that it is related to the oversampling filter of the interpolator, prior to conversion.

 

[Redacted]

Are you choosing a chip, or ruling out this chip, to be part of a DAC unit you are designing?

 

[sq225917]

Do you think anything 110db below signal is audible with music?

 

[Redacted]

Do you think anything 110db below signal is audible with music?

Not for me!

 

[nanook]

Hi,

5.6M or 11M refer to the oversampling rate of the DAC’s delta sigma modulator (128x or 256x). The higher number will better spread the quantization noise of the delta sigma modulator far in frequency, and lower it in amplitude.
For the two other options, my guess would be that it is related to the oversampling filter of the interpolator, prior to conversion.

The other register setting (what I assume is measurement-mode vs. sound-quality-mode) seems to change the final sampling as well; the null in the spectrum is ca. 1.5MHz for SC=1 and ca. 3MHz for SC=2, although both are for a clock of 5.6MHz

 

[nanook]

Are you choosing a chip, or ruling out this chip, to be part of a DAC unit you are designing?

I'm not designing a DAC - maybe when I'm retired . It's just curiosity, since I saw this sawtooth pattern in the time domain. And of course the specification mentioning "measurement-mode" vs. "sound-quality-mode" made me curious as well.

I wanted a really good DAC for measurements and the DO300ex was not cleaner at intermediate levels compared to my D-6s (spurs) - at least according to my measurements. It may be possible that the amount of high-frequency noise (due to the very lax analog output filtering that is implemented in the DO300ex) stressed the E1DA ADCiso and that the "grass" that I have seen was on the ADC side. I should have used a LP-filter with a step roll-off.

This is why I returned the DO300ex and tried a DO400, which - to my surprise - is a lot cleaner than the D-6s (ES9039q2m). The DO400 (ES9039pro) shows no THD-hump at intermediate levels.

Let me say it again, there was nothing wrong with the DO300ex - it sounds great. I did a listening session (level matched to 0.1dB @ 1kHz) against a D-6s and I couldn't really distinguish them - I ended up guessing.

 

[Redacted]

I'm not designing a DAC - maybe when I'm retired . It's just curiosity, since I saw this sawtooth pattern in the time domain. And of course the specification mentioning "measurement-mode" vs. "sound-quality-mode" made me curious as well.

I wanted a really good DAC for measurements and the DO300ex was not cleaner at intermediate levels compared to my D-6s (spurs) - at least according to my measurements. It may be possible that the amount of high-frequency noise (due to the very lax analog output filtering that is implemented in the DO300ex) stressed the E1DA ADCiso and that the "grass" that I have seen was on the ADC side. I should have used a LP-filter with a step roll-off.

This is why I returned the DO300ex and tried a DO400, which - to my surprise - is a lot cleaner than the D-6s (ES9039q2m). The DO400 (ES9039pro) shows no THD-hump at intermediate levels.

Let me say it again, there was nothing wrong with the DO300ex - it sounds great. I did a listening session (level matched to 0.1dB @ 1kHz) against a D-6s and I couldn't really distinguish them - I ended up guessing.

OK - curiosity is good. Measurement beats the usual adjectives and adverbs we see even on ASR- so hats off to you.
So have you found any systems that use DAC chips that we ought to avoid?

 

[nanook]

Do you think anything 110db below signal is audible with music?

Absolutely no. I bet, I wouldn't even notice, if spurs were 60dB below the music.

It's rather that measuring with a long FFT (to get the noise floor down) may suppress effects that are a lot larger if you would look at them short term.
It is just that I saw this sawtooth like pattern and started thinking that such periodic phenomena tend to cancel out over a long measurement period—which may not be the case with our hearing.

The DO300ex is used here merely as an example, since I saw this periodic pattern when measuring this DAC.