Building the ultimate NOS DAC using TDA1541A

[SSerg]

<font color="#000000">

[SSerg]

Hi, colleagues jameshillj, studiostevus and oshifis,
As judged by your postes, you bad visualize working of 1541. Do not study with divination, read the firsthands.
Here is else pair of the references.

http://www.vegalab.ru/forum/attachme...2&d=1340644012


http://www.vegalab.ru/forum/attachme...4&d=1340694699


I shall say only that frequency of the generator FDEM must be not lower 176 kHz (4Fs) that each current of the passive divider of the DEM cell has time to participate at averaging and at the bit current shaping.

[studiostevus]

Yes, that's what i said.

[-ecdesigns-]

Hi SSerg,

I shall say only that frequency of the generator FDEM must be not lower 176 kHz (4Fs) that each current of the passive divider of the DEM cell has time to participate at averaging and at the bit current shaping.

- TDA1541A typical application runs in 4 times oversampling.

- TDA1541A sample rate in a typical application equals 176.4 KHz.

- DEM clock frequency is specified between approx. 150 and 250 KHz.

- In a typical TDA1541A application the DEM switch advances roughly one time during each sample.

[studiostevus]

Tda1541 dem has a 4-step pattern, which is visible when you remove the passive filter caps. It is my understanding that you want to run through at least all four steps during one word (fDEM = 4fWS) to achieve optimal averaging over these 4 steps.

Total averaging is a function of averaging the inacuracies in timing dt (jitter of the dem oscillator) and averaging the inaccuracies in current di (inaccuracies of the internal resistors in the tda1541, which are not laser trimmed) according to van der Plassche. Refer to the earlier link(s).
My understanding is that, if you run through -say- 16 dem steps during one word (fDEM = 16fWS), you would achieve better result as the inaccuracies in timing dt (jitter of the dem oscillator) would be averaged 16 times per word. The inaccuracies in di would be the same though as in the case above, as it would cycle through the same 4 steps, just 4 times per word this time. I understand from John though, that there are drawbacks in spikes when moving from one dem step to the other, which probably outweigh the advantage of 4 vs 16 steps per word (with a reasonable low jitter dem oscillator).

Still, cycling through 2 steps per word (fDEM = 2fWS) would give low averaging across the di axis (2 steps per word only) as well as across the dt axis. The advantage is only 2 spikes per word though, but i am unsure as to whether this outweighs the lack of averaging (otherwise, why implement dem at all?).

I could be certainly way off here, as i am by no means an expert, but i still don't see how a 22khz dem clock (fDEM = 2fWS when using no oversampling) would give a satisfactory result.... If anyone can explain where my understanding is flawed?

[brubeck]

Originally Posted by studiostevus

Tda1541 dem has a 4-step pattern, which is visible when you remove the passive filter caps. It is my understanding that you want to run through at least all four steps during one word (fDEM = 4fWS) to achieve optimal averaging over these 4 steps.

Total averaging is a function of averaging the inacuracies in timing dt (jitter of the dem oscillator) and averaging the inaccuracies in current di (inaccuracies of the internal resistors in the tda1541, which are not laser trimmed) according to van der Plassche. Refer to the earlier link(s).
My understanding is that, if you run through -say- 16 dem steps during one word (fDEM = 16fWS), you would achieve better result as the inaccuracies in timing dt (jitter of the dem oscillator) would be averaged 16 times per word. The inaccuracies in di would be the same though as in the case above, as it would cycle through the same 4 steps, just 4 times per word this time. I understand from John though, that there are drawbacks in spikes when moving from one dem step to the other, which probably outweigh the advantage of 4 vs 16 steps per word (with a reasonable low jitter dem oscillator).

Still, cycling through 2 steps per word (fDEM = 2fWS) would give low averaging across the di axis (2 steps per word only) as well as across the dt axis. The advantage is only 2 spikes per word though, but i am unsure as to whether this outweighs the lack of averaging (otherwise, why implement dem at all?).

I could be certainly way off here, as i am by no means an expert, but i still don't see how a 22khz dem clock (fDEM = 2fWS when using no oversampling) would give a satisfactory result.... If anyone can explain where my understanding is flawed?

I tend to share your thoughts, but the size of the external capacitor plays also a role i suppose. If it has higher capacitance, it will act like a voltage buffer and equalizes the (DEM) ripple to a moving average voltage? Viewed that way i can imagine that lower fDEM doesn't much affect the results of DEM. Then it is not related to 1 word, but time moving average.

In my personal experience (with 1uF cap) the optimum for my ears is around 60khz DEM. If i go higher (>100khz) the sound seems more refined but tends to irritate, if i go lower (<25khz) it becomes a bit lifeless. I find it hard to correlate the results. There's a lot going on in the chip

[Bernhard]

Originally Posted by brubeck

In my personal experience (with 1uF cap) the optimum for my ears is around 60khz DEM.

Optimum ? Or maybe best compromise ?

[oshifis]

Originally Posted by brubeck

I tend to share your thoughts, but the size of the external capacitor plays also a role i suppose. If it has higher capacitance, it will act like a voltage buffer and equalizes the (DEM) ripple to a moving average voltage? Viewed that way i can imagine that lower fDEM doesn't much affect the results of DEM. Then it is not related to 1 word, but time moving average.

Moving average ~ low pass filtering
The internal resistance of the bit current generator and the external capacitor does exactly this. This arrangement simply can't take 4 DEM periods, average them (how? calculating the mathematical average?), take the next 4 DEM periods, do the same, etc. The only important thing is to have a precise 50% duty cycle of the DEM clock - as the cited article states. John also concluded that it can be best achived by decreasing the DEM frequency (and increasing the external capacitor at the same time).

[studiostevus]

Alright, thanks guys, it makes more sense to me now....

The key is that averaging happens continuously (moving average) and not in a word-by-word fashion.

In that notion, fDEM and fWS are completely unrelated.

[brubeck]

Originally Posted by Bernhard

Or maybe best compromise ?

I can't correlate 60khz as optimum fDEM in technical point of view.

It's subjective judgement based on listening assessment.