"3x80 A/D Converter" PCIe Idea..

Ixahe

Member
2016-04-30 12:19 pm
Hi.

I´ve had several DAs over the years, and started thinking about myself how to do an optimal D/A.

It seems to me even today, people cannot satisfy clocking requirements to do 1-bit converters well. Even some high-end manufacturers seems to want to try and hide this behind designs and colours.

While back in the 80s nobody really complained about the clocking of 8-bit converters, and even still today, many people listen to these kinds of records, and think they sound good.

The clocking requirements of an 8-bit is ofcourse much less, than the insane speeds of 1-bit streams.

What if one tried to align several 8-bit converters, to make them sum to 24-bit?

And would it be easily feasible to fit it on a low-latency PCI-e card? If noise is an issue, can we get it below -110dB matching high-end converters? (/Balanced Circuit might be good with 8-bit converters that probably are not expensive..?),

I would like to hear your comments.
 

Ixahe

Member
2016-04-30 12:19 pm
No comments? :)

Peter Gabriel also used 8-bit samplers like Fairlight, as many in the 80s. Later many also liked 12-bit samplers, but I think 8-bit is really where it was. As many people are nostalgic about the"amiga" homecomputer sound aswell, and there is a whole 80s movement around sounds like that, typically combining it with analog filters and/or envelopes aswell. Add a 2x2-voice chorus, and you have 24-bit stereo already.

I am going to do some more thinking on this.
 
Last edited:

Ixahe

Member
2016-04-30 12:19 pm
Well, the whole discussion is around whetver Sigma-Delta purely is a theoretical optimal that cannot be achieved in real-life conditions.

For instance even the quite recent Konnekt 6, sounds overall worse than a 20 year old 8-bit converter.

Some devices though, such as RME Fireface UCX, uses multibit variants, and sound quite good. It would seem to me that 8-bit would be something of an optimal. My original idea was trying to sum 3 8-bit converters to 24-bit though :) As a kind of a conclusion on the "vintage" debate that has been going for so long. In the end really whatever sounds the best, but definately the 1-bit theory is not working in practise. And even some multibit devices aren´t really working that well either, so it seems the bits need to go up atleast, and requirements to clocking down.

I am reading that last one now which talks about >3Mhz rates, and I am thinking that if the bits are too few, the harmonic distortion will be hidden in error correction feedback, but it will degrade the quality significantly.

Also some seem to overlook, that if you use a filter on the output with gibbs, then you will change the waveform, usually limited to 0dB in these days, and get distortion at the output stage, if not compensated.

The best thing here if using noiseshaping etc, is getting the noise in high frequency, and just using a gaussian filter to filter it out sufficiently.
 
Last edited:

Ixahe

Member
2016-04-30 12:19 pm
An easy approximation would ofcourse be summing 3 8-bit converters, with error feedback in a quantizer before them. A much lower samplerate can be used also. And on the output one needs an analog filter, preferrably non-gibbs to filter out the error feedback again. It is a similar principle to 1-bit on several bits. That is my much idea anyway. Taking clocking requirement from an impossible "accurate pulses" at 3Mhz to maybe 3-4 oversampling at 44.1khz.
 

Ixahe

Member
2016-04-30 12:19 pm
I'm unsure how to reconcile the "beats everything else in the chain" nature of top-flite audio DACs (much less ones designed for infinite other uses) and the fact that's Sigma Delta topologies aren't working. Plus, the entirety of your premise is that they don't sound as good because feedback. That basis is unreliable at best and leaves us wondering if implementation is at play.

Seriously, have you looked at the specs on newer devices? Anyhow, seems you have found a solution that works for you.
 

Ixahe

Member
2016-04-30 12:19 pm
Thanks.

I thought more about it, and lineary simply summing the three converters for a channel, and lineary dividing it at say 0.33, 0.66 should work just fine I think. And reduce the requirements for clocking so that quite readily accessible components can be used. And a minimal of analog components on the output.

This will be a fun project I think, I will be excited to learn more about.

Peaceful Greetings,
Ixahe.
 

Ixahe

Member
2016-04-30 12:19 pm
Or that might be a bit inefficient use of bits, so one should consider summing for optimal use of bits. Using Dither would also just be negative feedback paths around the quantizer, and more again around for higher order etc.

One can probably think of some clever schemes there. :)

And one would get the 8-bit paradigm, in 24bit resolution.

Peaceful Greetings.
 

Ixahe

Member
2016-04-30 12:19 pm
I emulated a diode the other day also. It seems it needs ~172x oversampling, for Vintage diode sound, like on a Roland System 100. Which really says much about how good transient response it must have. Probably why these were so popular.
Which led me to thinking, maybe a one-bit paradigm is better realizable on diodes, and thus should be considered still.
 

Ixahe

Member
2016-04-30 12:19 pm
Fun related miniproject: The parallel 8-bit thing I tested in practise also, by rom-readouts of a TR-909 cymbal sound, and layering them in parallel with offsets to get more bits. It sounds very good. - Did you know the 909 cymbals were actually jazz played? That is why they work so well in dance music.