Discrete DAC

Continuing on from the "favourite DAC" thread... at Jim's request, here's the basic schematic for my Discrete DAC.

A stereo I2S stream is used per DAC board to achieve up to 192K sampling rate from a 96K source (the '86 is used to fiddle the clock appropriately to make the '595s latch).

To get it to work with standard twos' complement from the likes of a DF1704/6, use one of the '86 gates to flip the MSB after the appropriate '595 output.

This will add jitter, though (and the Ron might be further out too), so not the best way of doing things - just the simplest if you don't have adequate control over your source.

- Tom.

Edit: Grr... had to resize it to fit within the 1000*1000 image limit... still just about legible, though...
 

Attachments

  • ddac18-002a.png
    ddac18-002a.png
    19.4 KB · Views: 3,788
Speed is not a problem

Peranders,

Please define "normal parts"!

The old matrix board version uses standard leaded metal films... yet runs plenty fast enough to cause significant intermodulation without an output filter. Inductance is reduced by paralleling resistors - a side effect of striving for tighter matching.

I've no doubt that the surface mount version is capable of higher speeds, but no issue has arisen.

I normally run at 176.4K (4X oversampling), and the DAC shows a minor roll-off at 20K... part of which can be attributed to the digital filter... part of it to inductance in the coupling cap.

I haven't tried putting any test tones above CD rate through, simply because it isn't very interesting, and it's a hassle to do.

Resolution is more of an issue, for the reasons stated in the "favourite DAC" thread. If you go to my website and see the -3dBFS sine graph, you'll see what kind of technical performance I've achieved without resorting to custom half bridges or fancy regulators to keep switch impedance under control. If it was feasible, I wouldn't be using any regulator at all... but battery operation is a pain, and didn't sound any better last time I tried it.

There are also various tricks you can pull to improve things further in the digital domain... namely noise shaping and bit calibration. The Ti48 transport already offers shaping, but not down to 14 bits. Without dither being fully resolved, quantisation distortion results.

The calculations I did before building the DAC suggested I'd only achieve 12 bits, so getting 14 or so isn't bad going at all.

When I first set to do the project I set myself one single technical objective... the measured performance has to be better or equal to vinyl, and I think I've achieved that.

Jim,

The output to ground resistor is to balance the ladder impedance, as I understand it (I have done LTSpice simulations in the past, but my memory is a little hazy now)... it's not strictly necessary to put one on the board to have it work, as your preamp input impedance will achieve pretty much the same result, but performance may be affected.

- Tom.
 

wimms

Member
2004-03-02 2:59 pm
home
peranders said:
Question: Have many bits is possible to achieve with a discrete DAC made with normal parts?

I don't have the answer but I'll suspect it's far less than 16 bits and max 50 kSp/s, probably 14 bits or so.
MSB must not induce error more than 1/2 of LSB. Very roughly, precision of resistors in 16-bit R2R ladder must be to 1 part in 65536, thats 0.001% precision to calculated ladder values. Laser trimming comes to mind.
Not only that, but also each switch must source and sink precisely same current, again with similar precision. In fact, total cumulative error must be below LSB.
In addition, temperature induced parameter drift must be handled. This might get even toughest part of discrete dac. If you pick R=1k then drift of 15 milliohms blows your 16-bit precision already.

There are few commercial discrete DACs, one of them very higly praised (memory fails me) that used controlled heaters to keep temperature under control, and iirc reached 19-bits resolution. At insane cost.

I think you can't go beyond 12-14 bits with discrete affordably, so its hardly worth it.

But building discrete dac can be very educational ;)
 
In addition, temperature induced parameter drift must be handled. This might get even toughest part of discrete dac. If you pick R=1k then drift of 15 milliohms blows your 16-bit precision already.

So don't pick R=1k... :rolleyes:

What makes you think any part of the network will be at enough of a different temperature than the rest to cause the required shift? This is a line level DAC with extremely low current flowing through the network. They essentially drift together, so the ratio stays the same.

There are few commercial discrete DACs, one of them very higly praised (memory fails me) that used controlled heaters to keep temperature under control, and iirc reached 19-bits resolution. At insane cost.

That'll be the Lavry. Incredibly complicated, which is exactly what I was trying to get away from in the first place. Regardless, a nice piece of engineering.

The price isn't insane. I have no doubt you can pay more and get a whole lot less, but that's "high-end" audio for you... or any esoteric field, for that matter.

I think you can't go beyond 12-14 bits with discrete affordably, so its hardly worth it.

Rather depends on the sound quality you get, doesn't it? :scratch2:

I have heard plenty of DACs with vastly better linearity, but they don't sound as good to me. Modern multi-level delta-sigmas measure so wonderfully, don't they? :smash:

- Tom.
 
Jocko Homo said:
"Modern multi-level delta-sigmas measure so wonderfully, don't they?"

No. At least not by the standards that I favour. Measurements help, but they aren't the final arbiter.

Jocko

OK, so what's wrong with delta sigma chips?

By the way, Tom, my comment in the other thread about paralleling your DAC was tongue in cheek - the thought of building eight of the blighters and conecting them together ...

If you are happy with the linearity, what do you think is limiting the performance now? I doff my cap to you for your efforts.
 
Tom was being sarcastic. Naughty Tom!

Hi Carlos... :D

OK, so what's wrong with delta sigma chips?

Good question. I doubt anyone can give you a scientific answer, beyond some vague hand-waving about interaction with external components. The technical performance up to, say, 20kHz of a modern audio delta-sigma is excellent.

So should we care about ultrasonics? Even if they don't measurably intermodulate down? Perhaps...

I have a LynxTWO soundcard here with CS4396 DACs and AK5394 ADC for "lab" use... modified it for slightly lower non-harmonics (which are already extremely low compared to any other soundcard I'm aware of). But after hearing what an R-2R could do, I wasn't entirely convinced by the way the Lynx plays music.

It could be a whole bunch of things. Poor quality digital filtering (certainly a factor, IME)... lack of precision in the modulator (they will use the least they can get away with), disguised but still audible non-harmonic tones, the switched capacitor filter, correlation in the dynamic element matching, the use of global feedback through an internal A/D... oh, and yes, that old favourite... jitter...

Who can be bothered with worrying about all that? The datasheets tell you very little of use with regard to how it will sound. They might as well be black box devices to us mere mortals without a degree in advanced mathematics and insider knowledge...

They've already made all the important decisions for you. And I like to make my own, regardless whether judged by others to be wrong or not.

I do find it mind-boggling that the same people who insist on building amps out of individual transistors or valves will not think twice when it comes to their D/A about choosing a lump of silicon with possibly hundreds of thousands of transistors inside.

By the way, Tom, my comment in the other thread about paralleling your DAC was tongue in cheek - the thought of building eight of the blighters and conecting them together ...

Heh... the thing is, I *have* seriously considered it... yes, it would be hard work... unless I can pay someone else to measure and label the resistors to six digits... buying them off the shelf to this precision is simply not financially viable for someone like me.

If money was no object, I'd get Vishay to make me a custom network from laser-trimmed bulk foils. As if...

If you are happy with the linearity, what do you think is limiting the performance now?

My speakers and room. ;)

The DAC is nowhere near the limiting factor in my setup. At least by my own listening criteria.

To be fair, the speakers are now acceptable. I suspect the power amp is actually weaker. But the room is just awful.

So I spend a lot of time listening on headphones. I've had the opportunity to listen to quite a few digital front ends through the same headphone setup for many years... makes for a solid reproduction reference.

I doff my cap to you for your efforts.

Thank you. Most kind. :)

As I said to Carlos privately, I still bet no-one who comments on this thread is going to build one, though. I witnessed the last thread on this years ago on this very forum... a lot of talk... not a lot of action.

- Tom.
 

wimms

Member
2004-03-02 2:59 pm
home
tbrowne said:
So don't pick R=1k...
:rolleyes: Pick any number. That was example. What matters is that you need 15 ppm precision for 16-bits. 1 ppm for 20 bits.

What makes you think any part of the network will be at enough of a different temperature than the rest to cause the required shift? This is a line level DAC with extremely low current flowing through the network. They essentially drift together, so the ratio stays the same.
What makes you think its as simple as that? Switches are essentially part of the ladder, so you need to match thermal drift of ladder with that of switches, not speaking of switches themselves. For 20-bits even match thermal drift of PCB and wires. And why you assume that all drifting is equal for all discrete resistors? Why do you assume that ratio stays the same? At 1% precision that might be true, but at few ppm precision?

Resistors values drift over time too. In tens of ppm. Randomly. Manufacturing dispersion. Your DAC would show different linearity at every different ambient temperature, every different month. Yes you could in principle handpick and match all components and all their drift curves, but there is no repeatability in this. You need to auto-calibrate.

I don't get where does all this anti-chipness come from. Don't you know the historical background? To get the thermal issues solved, you'd want to have ladder and switches manufactured in same technological process and on the same substrate, then you can expect ratios to stay the same, to some limits. And 1 ppm (even relative precision) is still hard to achieve.

That'll be the Lavry. Incredibly complicated, which is exactly what I was trying to get away from in the first place. Regardless, a nice piece of engineering.
Right. Do you think he used oven for R2R ladder and intense auto-calibration out of nothing better to do?

Rather depends on the sound quality you get, doesn't it? :scratch2:

I have heard plenty of DACs with vastly better linearity, but they don't sound as good to me. Modern multi-level delta-sigmas measure so wonderfully, don't they?
Right. Knew it will come down to this. 1543 NOS and friends. Emotions vs. reality. There are no tubes in your circuit ;) Linearity sounds so dull and lifeless... Reality looks so dull and colorless, not like those TV "quality" commercials..

You'd better have a DAC thats true to original and use DSP's to distort the sound the way you like it, instead of using emotional attachment to justify inherent imperfections.

Make a 14-bit dac and enjoy it if you do. To get 18-bit dac it takes quite abit more than just making 18-bit ladder with dogcheap shift registers, or to make it any better than off the shelf chip. Sounds good? Its in the ear of beholder. You might find that even 8-bit R2R sounds better than any modern chip.

Sound quality above LynxTwo? Give me a break..

No, its a nice project for somebody who wants to get to the meat of a DAC. I'm not discouraging at all.
 
"I do find it mind-boggling that the same people who insist on building amps out of individual transistors or valves will not think twice when it comes to their D/A about choosing a lump of silicon with possibly hundreds of thousands of transistors inside."

Well, as distasteful as it may be, there are good reasons for doing so.

There are even good reasons to use chip amps. Which doesn't really prove much. E xcept that when properly implemented, good results can be had through many methods.

Jocko
 
:sleep:

What matters is that you need 15 ppm precision for 16-bits.

The resistors I picked have a typical spec in this region.

Switches are essentially part of the ladder

Why don't you just read my own website back to me, eh? :rolleyes:

so you need to match thermal drift of ladder with that of switches

No, not if the switch impedance is low enough in comparison to network impedance after trimming. I'll give you a hint... the network values were not chosen at random. All known factors have been taken into consideration.

This is not a project that was undertaken lightly. I would have hoped that was obvious.

Resistors values drift over time too. In tens of ppm.

Precision resistors tend to come with appropriate data to indicate long-term drift. At the standard 25C with these ones, drift per year is not a particular concern. I can always do what I did with the old boards, and put trimmers in.

Randomly.

State your evidence.

DAC would show different linearity at every different ambient temperature, every different month

No, not significantly, it doesn't. We're not talking about some theoretical design here - we're talking about a real thing. Or do you not participate in such frippery?

I don't get where does all this anti-chipness come from.

Perhaps you should read my responses more carefully.

Don't you know the historical background?

I am aware of the "progression" of audio D/As all the way back to the very late seventies/early eighties, at a time when DACs were large packages of parts, and Burr Brown sold deglitching circuitry as an option - nothing earlier, unfortunately.

Do you think he used oven for R2R ladder and intense auto-calibration out of nothing better to do?

Why don't you tell us what you think? You seem to be very keen to offer an opinion.

Knew it will come down to this. 1543 NOS and friends.

I have no time for NOS @ 44.1 at all. Others feel differently, I know.

There are no tubes in your circuit ;)

Oh... should there be? This is about the point where a sense of humour failure is entitled from myself.

You'd better have a DAC thats true to original and use DSP's to distort the sound the way you like it, instead of using emotional attachment to justify inherent imperfections.

Perhaps three years ago, I was telling a current member of this forum essentially the same thing. Fortunately, I was mature enough to recognise a later time that I was missing the point of reproduction equipment entirely, and was also making some unfortunate assumptions. Even if what you (and me all that time ago) propose could be done, what is your rationale beyond a tedious academic exercise?

Jocko is entirely right regarding measurements, of course... and regarding good reasons for using an IC. Sure there are - cost, transistor matching and thermal coupling all immediately spring to mind... I even have some of BrianGT's LM3875 boards... and participated in testing and critical evaluation of a very fine PCM1704-based DAC. But that is not really the point.

You might find that even 8-bit R2R sounds better than any modern chip.

A less ignorant soul might understand that an 8-bit R2R can achieve far more than 8 bits of audio reproduction.

To get 18-bit dac it takes quite abit more than just making 18-bit ladder with dogcheap shift registers, or to make it any better than off the shelf chip.

Who gets to say what is "better"? You? What are your credentials?

Sound quality above LynxTwo? Give me a break..

I see "in the ear of beholder" has already gone out of the window. I've actually listened to both and you haven't, but don't let that stop you from voicing an opinion.

After all, not everyone can judge a book by its cover, can they? That takes special talent...

I'm not discouraging at all.

Could have fooled me. Maybe it's a language thing.

Never mind, I'm sure we'll be hearing about something you've actually built one of these days...
 
Tom:

In its final version, is this fully enclosed? From my experience, air movement can upset things when you are trying to resolve things with that many bits of resolution. There can be enough temperature differential from one end to the other to throw things off quite a way. Even with tight tolerance parts.

Mind you, I have never tried this sort of beast, but having worked with MC phono preamps, I have had my share of wrestling thermal problems.

Jocko
 
Hi Jocko,

In its final version, is this fully enclosed?

Do you mean the individual modules? No - wouldn't bother without filling with fluroinert or similar (had a dream of OEMing them out in a sealed, swish copper box at one time). I don't even put the case lid on anymore for serious listening.

I'm slightly embarrassed to say it, but I could swear it sounds better with the lid off... significantly more dynamic.

Before the ridicule starts, I would note that with the Revision 1 and 2 boards, this actually correlated to a significant difference in harmonic distortion (presumably down to RFI) depending on lid (significantly lower with it off - from memory, as much as 3dB or so for some harmonics!).

In contrast, no significant measured difference was found today with the swish, teensy Rev.4s in some quick testing (they've been running for many months, but I've been too engrossed in listening to measure for a while). Yet I'm still hearing a difference... or at least I think I am!

I probably need a second or third opinion. Only one member of this forum other than me has heard the Discrete to date... revisions 1 and 2 (which uses individual buffers for each bit) in different houses, but not Rev.4.

Some might be interested to know that I never meant to run the network directly off the logic, for good reasons. It came about by accident, when the intended circuit (using HC4066s, one per bit) wasn't working as well as had been hoped (quite noisy). It was suggested to me to try just hooking the '595 outputs direct, and what I heard blew me away...

I've no idea why it sounds so good. I can only assume that '595s have been continuously been refined over the years, as I thought the crosstalk would be a huge problem.

There can be enough temperature differential from one end to the other to throw things off quite a way. Even with tight tolerance parts.

Absolutely, but I think I've taken this into account adequately with the design for both resistors and switches through various simulations and calculations... and also sporadic measurements to check the linearity is remaining consistent.

As well as the sound, I love the simplicity of the signal path... half bridges -> network -> passive filter -> attenuating transformer -> power amp... if only it could supply enough current to drive a transducer directly...

But that's a project for another day... I've got three different ways in mind to create a multi-level PowerDAC (no, not 16 bits plus - I'm not a complete masochist), and none of them are a walk in the park. One of them may not even be feasible at all.

It's something I've wanted to build for years (have had an oversampler, filter and 8-bit modulator running on Blackfin DSP for a similar length of time), but sometimes you have to recognise when something could be a bottomless pit for your money and time...

Speaking of which, WTH am I doing up after 3am on the computer! Now I remember one of the reasons I didn't join the forum for so many years... :)

- Tom.
 
Hi Tom,

which values do you use for R & 2R ?

Would 50K & 100K be ok ?

I like to built discrete too and collect 0,1% resistors for that, even have a few with 0,003%.

My plan is to connect 50ohm & 100ohm pots in series with each resistor and adjust to minimal distortion on spectrum analyzer.

What about power DAC ? Multibit ?

I started with it last year but had no time to proceed.
And did not publish the method here for some reasons including the usual discourageing comments like "You can't do this and you can't do that, because this will not work and that will not work".
 

dddac

Member
Paid Member
2003-07-02 10:10 pm
Wiesbaden
blog.dddac.com
Bernhard said:
And did not publish the method here for some reasons including the usual discourageing comments like "You can't do this and you can't do that, because this will not work and that will not work".

Bernhard,

As a DIY developper, it is very simple. You have build designed and build something which works, then you publish. Regardles what other people may say about what works or not. It worked didn't it ?? You can stand above that...

If the design leads to the expected result is not so important. We do not only want to read in this forum (at least I do) about the next ultimate reference, but also about new ways (to Rome :D ) which even has not been so succesfull, but we can LEARN something about this and it might help other people find other ways....

So, if you have built something show it to us..... don't be afraid

Doede
 
Hi Bernhard,

which values do you use for R & 2R ?

Revisions 1-3 use 23.5K and 11.75K (from 47K resistors).

Revision 4 uses better-quality 22K resistors, also paralleled two and four (three and four including the high ohmic 2R switch trim).

You need to be very careful with the values chosen for a whole bunch of reasons... but one major one if looking for precision is that you need to know your switch Z first... otherwise your trim might not be a value you can find off-the-shelf... making life much harder (I prefer to avoid the use of pots).

With Rev.4, I made the mistake of assuming that AHC and VHC logic would have roughly the same Z when ordering parts (Rev.2 used VHC), as they are so similarly specified on the datasheets... not even close... fortunately, it turned out there was a reasonably close resistor match for the actual Z.

I like to built discrete too and collect 0,1% resistors for that, even have a few with 0,003%.

The actual "as sold" value precision of the resistors isn't of ultimate importance, though more tightly specified parts do tend to be formulated for a lower drift coefficient.

What needs to be done if you're on a tight budget (and I didn't enjoy it either) is to measure and label every resistor individually - ideally in one sitting where the temperature is fairly constant.

And you *will* need more than you use... at least if you want useful linearity (I would suggest aim for at least 0.005% in each position, preferably better). Which is why I bought a thousand 1% resistors for 1p each instead of a fortune on far fewer 0.1%s.

For revision 1, I had to make to do with a 4-digit DMM. This really isn't good enough, IMO... and things are made more complicated by the fact that DMMs drift too! It takes a while to get the hang of getting repeatedly consistent measurements, working with the liimtations of the equipment you have.

For revision 2, I was loaned a superior DMM with five digits, making life a little easier. I also developed a rough software application which would give the optimum pairings for lowest overall deviation when parallelled, tightest for MSB... vastly easier than working it out manually.

For revision 4, I bought my own precision DMM second (or third) hand from Ebay. I wanted a HP, but the prices were out of my range. Eventually ended up with a Solartron 7150+... the insides are quite a treat, it goes up to six digits with a lowpass filter, and it works pretty well, though drifts more than I'd expect it to (perhaps due a calibration, as if I could afford it...).

My plan is to connect 50ohm & 100ohm pots in series with each resistor and adjust to minimal distortion on spectrum analyzer.

I wouldn't recommend that... pots are generally noisy, generate more distortion, and also drift like crazy compared to a precision resistor. It's better to use high ohmic value of resistor or pot+resistor in parallel than a small value in series for matching, IMO.

What about power DAC ? Multibit ?

I've not built any hardware for that project as yet... only software, to reduce a 16/44.1 source down to 8/358.2K. It would involve a large amount of time and money to do it "right"...

I started with it last year but had no time to proceed.

I understand...

And did not publish the method here for some reasons including the usual discourageing comments like "You can't do this and you can't do that, because this will not work and that will not work".

There will always be people like that. I try to adopt a practice of offering an alternative if I feel something is incorrect, rather than just leaving someone with a dead end. What would be the point of that? I might as well just not comment, leave them to it and let them learn something from the experience.

As a long time reader of this forum, I have seen plenty of individuals over the years who habitually offer nothing better than unconstructive, ultimately worthless criticism (this is one of the reasons I didn't join until now). It's a bad habit to get into, and I have indulged in the past at points, I'm sure... but at least I'm aware of it.

How long did it take until someone took it upon themselves to "school" me in a project I've been researching and developing for many months, and considered for years? Nowhere near long enough, sadly. But don't let the b*****s grind you down... :angel:

- Tom.
 

rfbrw

Member
2001-10-26 11:51 pm
.
There is a difference between a well thought out idea from one who has done the research and is clearly on top of their subject,as in the case of the Discrete DAC, and a daft idea that won't withstand a second's analysis. This is an open forum and if you are not prepared to do the research or even make the slightest effort to learn the subject, as some clearly are not, you should expect to be ridiculed.