Mola Mola DAC and other stuff (DAC techno-discussion)

This is quite an interesting article:
http://jugandi.com/ebooks/Digital/03 A&D Conversions.htm

especially these few diagrams:








Slope overload

 

On aliasing:


This additional aliasing noise occurs because the amplitude of the LSB components increase as frequency decreases. To minimize this, high order antialiasing filters are required to filter the baseband signal before sampling.



There are two different methods to integrate analog functions such as filtering. One very powerful technique is to employ DSP where the input signal is converted to the digital domain and the function implemented in either digital hardware or software. A second method uses switched capacitor technology. This is a sort of analog-digital hybrid technique.

One of the main applications for switched capacitor filters is in antialiasing filters. A DSP implementation would use over-sampling and decimation.


source:
http://jugandi.com/ebooks/Digital/03 A&D Conversions.htm

 

here's the video interview of the mola mola dac designer:

 

Regarding #3 (average filtering), I believe the individual current output segments of the dac-output are arranged as a shift register, or fifo. So the 32 current-output segments are active at the same time, each with 1 different sequential bit of the 100Mhz data. Then every tick of 100Mhz all the bits are shifted to the next current output segment, with one new bit in and one out the end. All these current outputs sum, so the effect is to do a moving average filter with a length of the fifo. Also every bit passes through every segment, so its total weight will be constant regardless of the precision trim of the individual current output segments. Anyway, comparing this approach to modulating all the segments together at 100Mhz: the results are essentially equivalent, but there is less high frequency noise (analog) that needs filtering out.

 

I think aliasing isn't as important on the playback side, compared to the capture side, assuming that it's the final output and not an intermediate DSP stage, and your analog output stages should be pretty immune to high-frequency signals on their inputs if you're going to let aliases through. Certainly, the upsampler needs to suppress aliases, and quantization error can also cause aliasing with the sharp edges caused by quantization introducing high frequency content.

Bruno was one of the people who worked on DSD at Philips when it was first being done. He did not then and probably does not now have a very high opinion about it, but he knows his stuff. It could be that his communication style or English is giving you the wrong impression of him.

 

I'll try to carve out some time at lunch or something to listen to your samples and give thoughts. What songs are you using? It might be worthwhile to grab something we can all use as a reference track.

I'd also love to dip my toe into this a little more technically.

Where'd you find out how to do all this? Do you have a step by step guide somewhere?

I've never used Matlab, but I used to be relatively fluent in quite a few languages (C/C++ being the most complex of them) and use Lua almost daily at work, so I assume I'll be able to pick it up.

I also bought JRiver while it was $10 off. Still kinda prefer foobar2000. Shrug.

 

The quote from the video is: "A lot of the filtering is actually done for free in this manner."
As far as I can tell he's only claiming it's a bonus feature of the arrangement. I think it fair to claim that reducing noise ~10-30x is "a lot". But certainly more filtering is done later.

 

I'm guessing the main purpose of the multiple output segments has to do with keeping the circuit noise (noise floor) low, while using components that operate comfortably with switching at 100MHz.

 

here's a internal pic of the prototype mola mola dac. It looks quite complex from a number of chips standpoint.

source:
http://puremusicgroup.com/cart/index.php?_a=viewProd&productId=545&review=write


It's (claimed) SNR is 140db which is even higher than Yggdrasil's(conservative 120db+/-) or ES9018S(Stereo mode 135db @ interface output).

 

over at his blog:
http://www.mola-mola.nl/index.php/blog

He mentions that the Audio Precision's ADC is insufficient to design high end dacs.

 

I was thinking of truncation without dithering, which would introduce an out-of-band signal that would alias.

 

Hi ultrabike, what does tap mean here? I remember Bruno Putzeys said although ONE single-bit DAC is enough (i.e. it's done, you could hear the music) but he chose 32, each DAC will carry the whole signal. 32 whole signal(s) will smooth something...

Is he using some secret weapons?

 

Who cares about his resolution and performance so far above ADCs and redbook? How about he builds something easy to implement that sounds relatively neutral and with a filter doesn't discard (or is at least able to accurately reconstruct) all the phase? Not that modern, sterile "studio sound" crap has any phase left in it after being mixed to a surgical ward. Sabre can mass produce shit that measures great but still sounds like a Sabre even with ridiculously overbuilt and filtered power supplies. Other brands of DACs like Wolfson, AKM, or TI do not have any measurable high frequency roll-off; Sabre just sounds like a saber.

 

Its basically the same kind of claims(possible mind you) that Chord makes. Making your own custom modulators to achieve better performance than the off the shelf stuff. Still wanna try his stuff. Also really like the stuff he did with Hypex n-core.

Its also true that for most dac design work, current ADCs simply don't provide the resolution needed without a Twin-T notch filter for some really fine measurement work. Then again it depends on what you goal actually is. For very low THD(if that is the designer's goal), Twin T and distortion magnifiers are required.