How does a delta sigma DAC work?

Status
This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.
Bricolo said:
I've seen how a R2R dac works, but not for delta sigma ones
can someone explain me?
R2R DAC's are very easy to understand but delta-sigma is very complicated and some of the functions are also company secrets. You have to know digital math at a very high level in order to really understand but the first and second link Jean-Paul gave says a little.
 
Disabled Account
Joined 2002
Re: Re: How does a delta sigma DAC work?

peranders said:

R2R DAC's are very easy to understand but delta-sigma is very complicated and some of the functions are also company secrets. You have to know digital math at a very high level in order to really understand but the first and second link Jean-Paul gave say a little.


They sure are very complex ! I never bothered to understand them thoroughly ( to me they're all 1 bit :yuck: ) as I yet have to hear one that sounds better than those plain old R2R DAC chips :D

MASH/Bitstream/Delta Sigma : the names are nicer than the sound they produce...
 
Re: Re: Re: How does a delta sigma DAC work?

jean-paul said:



They sure are very complex ! I never bothered to understand them thoroughly ( to me they're all 1 bit :yuck: ) as I yet have to hear one that sounds better than those plain old R2R DAC chips :D

MASH/Bitstream/Delta Sigma : the names are nicer than the sound they produce...

The noise shaping changes the inter-sample sub steps, and the spectral nature of the consequent distortions.
Panasonic MASH drives me out the room, Sony is ok, and Philips quite good, and all are different to R2R.

Eric.
 
Koinichiwa,

Bricolo said:
I've seen how a R2R dac works, but not for delta sigma ones

can someone explain me?


I notice that so far no-one has really given a brief, concise explanation of both the ideal and the practical Delta Sigma DAC. I will try to keep this short, brief and sweet and on "popular science" level. In order to bring out the main points I will have to simplify many things and in order to throw some of the issues into sharp relief I will have to overstate some things.

So please, for all those who know it all and have not bothered to explain, note the context of the writing and save your extended criticism of specific points. I do know it just as well (but for simplicity will not cover it) and on most people who basically just ask "how does it work" the finer points will be mostly lost.

Okay. Now to approximate a given analogue waveform (you can never do more than that with digital) we can represent the waveform as a sequence of numbers.

An R2R DAC will take such a number and turn it into a current, by so to speak throwing a suitable number of switches that cause each a certain amount of current to flow, the switches being encoded binary, so that a given "number" can be translated directly into an analogue signal.

In order to make this work well our various forms of current switches and current dividers need to be jolly accurate, 16 Bit precision in micxed signal silicon is not that easy, 20 Bit becomes real hard, 24 Bit is definitly "G*d's domain".

But we can actually represent our waveform in another way. In the simplest way lets take a clock of 200KHz. If we divide this clock by a factor 4 we have now a square-wave with 50KHz frequency and each "Hi" or "Lo" lasts for four cycles of the 400KHz clock.

NOW, what if put in a locgic that allows ups to change the length of the "Hi" or "Lo" section in each cycle from 0 to 4 cycles of the clock.

In the most extreme case in each 50KHz full cycle the output of our logic remains "Hi" (or to say it remains "Hi" for 4 cycles of our original 200KHz clock), giving an output voltage for usual 5V Logic of around 4.5V and DC at that. Equally if our logic remains "Lo" for the full cycle our output voltage is around 0.5V and DC at that.

What if we chose an intermediate length of "Hi" and "Lo". Say we go back to equal cycle length for "Hi" and "Lo". Our output is now again a 50KHz squarewave, "Hi" for two cycles of the 200KHz clock and "Lo" for the other two cycles.

BUT, if we lowpass filter this squarewave out we are actually left in our case with an output Voltage of 2.5V and pretty much DC at that. If now make the "Hi" section 3 cycles of our 200KHz clock and the low section 1 cycle long and lowpass filter our 50KHz clock out, what will be the voltage measured on the output of our logic, yup, 3.5V. And if we keep the "Lo" cycle for three cycles of our 200KHz clock and "Hi" for one cycle we have 1.5V DC output.

What we have just made is the most primitive Delta Sigma DAC possible. It allows us to adjust the output Voltage of our logic from 0.5V in 1V steps to 4.5V, equivalent to a 2Bit "multibit" DAC.

Now if we increase the ratio between our main clock and the final sample clock (which in the above case BTW would be 50KHz) higher we can encode more bits. For example, doubling our main clock to 400KHz for a 50KHz clock gets us up to an effective 3Bit DAC, going up to 800KHz makes it a 4 bit equivalent, 1.6MHz makes it 5 Bit, 3.2MHz makes it 6 Bit, 6.4MHz makes is 7 Bit and 12.8MHz makes it 8 Bit.

Another way of describing our "primitive" delta sigma DAC is to call the operation "Pulse Width Modulation" or "Pulse Density Modulation". In fact, the recently again fashionable "Class D" or "Class G" Amplifiers operate in such a mode. I experiemnted with a fully discrete PWM Amp in the late 1970's, it has a way too slow clock (only 200KHz) and sounded pretty bad, but it "sort of" made music.

BTW, I hope that people have noitced that our "DAC" has a direct voltage output coming directly from a CMOS Gate or inverter. I hope also that it is noted that PSU voltage plays it's role here. And I hope it is very clear what influence Jitter would have on such a DAC.

Now, most commercial "delta sigma" DAC's operate at 256 to 512 times the sample clock. If we rely purely upon the pulse width modulation possible by adjusting the "Hi" or "Lo" period of the delta sigma modulators output to integer multiples of main clock, we can literally infer the the possible resolution of the DAC by comparing the main clock frequency with the Sample frequency.

If the Delta Sigma modulator operates at 256 times the sample rate we have 8 Bit effective resolution, if it operates at 512 times the sample rate we have 9 Bit and at 1024 times the sample rate we have 10 Bit.

To get the equivalent of 16 Bit resolution we would have to operate at a whopping 65536 times the sample rate, for 44.1KHz that would be around 2.89GHz, a clockspeed quite possible for modern Pentium and similar CPU's, but clearly outside the range for an "economy" DAC.

So, how do we reconcile a DAC capable of 8-10 Bit resulution and an Audio Output claiming 16, 18, 20 or even 24 Bit equivalent resolution? The "magic word" is called noiseshaping. It is literally "magic". How so?

What noise shaping does is to modulate the signal fed to the Delta Sigma Modulator in such a way that quantitasation noise from effectively truncating the data to 8..10Bit is pushed to outside the Audio range. So, the horrible distortion from the truncation is still there, but is no longer observable, after applying a suitably steep (usually 5th - 7th order) analogue filter after the DAC.

How exactly (their) noiseshaping works is one of key secrets of DAC makers. I'll not go deeply into noiseshaping, it's a fun topic and one that makes the Valves vs Solid state debates look harmless.

However, in truth, even with noiseshaping the results where not always so hot (measured and sonic), so the various delta sigma DAC's where for a long time consigned to "lo-fi" applications. Better noiseshaping algorythms and the introduction of "multi-level" delta sigma modulators changed the picture.

Oh yes, a "multi level delta sigma modulator" is nothing more that a multibit DAC being combined with clockcycle modulation and noiseshaping. But even a 2-Bit (4-level) multilevel Modulator would give 2-bit more real resolution and hence less agressive noiseshaping was needed. Certain manufacturers produced quite decent 2nd generation DS Chips, such as the NPC SM5864 and SM5872 or the Cirrus Logic (aka Crystal) CS4328.

While NPC provided a near raw PWM/PDM output Cirrus Logic integrated a switched capacitor filter on chip, making sure that even the daftest engineer couldn't balls up the design. The strong ultrasonic content in the output of a Delta Sigma DAC makes many a non-oversampling DAC look quite decent. Since then, just to prove that most EE's are daft, DS DAC's all come with on board switched capacitor filters and on-board buffer Op-Amp's making sure no-body can make a big cockup, but also disallowing the more competent engineer any control over the analogue filtering and analogue stage quality.

In recent years the number of Bits in the Multilevel Delta Sigma modulators have continously increased. The dCS Elgar operates it's core IIRC at 1024 times sample frequency and has what amounts to a 5 Bit Multibit DAC, so without noiseshaping we already have 15 Bit real response, if we add the 8 Bit that earlier DS DAC's needed to "noiseshape" we have theoretically around 23 Bit performance.

The Burr Brown PCM1738 operates it's core at 64 times Fs and has what amounts to a 7 Bit equivalent Multilevel Modulator, while the Cirrus CS43122 "SOTA" DAC has what amounts to a 6-Bit Multilevel DAC section with the modulator core operating at 128 times Fs (for 48Khz or lower Fs), both yielding 13 Bit effective resultion before noiseshaping.

Sony's DSD used for SACD operates very similar too.

The "latest and greatest" PCM1792 from BB/TI seems to combine a multilevel DS Modulator with a further Multibit current output DAC after that (not sure how they make that work, might have to look at that).

So, I hope this gives some Idea of what "Delta Sigma" DAC's do and how, plus also the fact that few currently made socalled "Delta Sigma" DAC's are such, in the purest sense of the word.

The hybrid combination of high speed, medium bitdepth video DAC Chips (12-Bit ones are available) combined with pulse density /pulse width modulation may be the way forward for the "High End", if a high performance DAC is desired without agressive noiseshaping.

Using a single 12-Bit Video DAC with a main clock at 1024 times Fs (around 200MHz clock) would allow around 22 Bit equivalent resolution. By using multiple "interleaved" converters (luckily Video Converters come already in packs of three or more per chip) and/or massive parallel arrangements of such DAC's it should be possible to push close to the performance that for pure analogue means remains "G*d's domain" and to do so without requiring oversampling with digital filters in the traditional sense. Then of course we would need equally good ADC's...

Sayonara
 
Whao, thanks a lot!

That's very well explained :) (I suspect Kuei Yang Wang for being the writer from the books called "Idiots guide for ..." :D)


As delta sigma DACs are a kind of PWM, where does theyr name come from? Do they have a kind of mathematical operation in them, such as a derivation?


"The strong ultrasonic content in the output of a Delta Sigma DAC makes many a non-oversampling DAC look quite decent"
Wouldn't it be the opposite? Since you have HF output, non os would not be a good idea.
Where does this ultrasonic content come from? Isn't the 5-7th order lowpass enough to filter it?

Does this say that R2R DACs haven't any HF on the output? Making them more suitable to a non os DAC?


Thanks,
Alex
 
i can't help with the origins of the name but i know it's used for marketing.. product brochures will advertise "delta-sigma" like it's the key to good sound..

what i don't get is if multi-bit dacs are superior then why bother creating delta-sigma dacs anyhow? seems like a wasted time and effort.
 
HeadSh0T said:
i can't help with the origins of the name but i know it's used for marketing.. product brochures will advertise "delta-sigma" like it's the key to good sound..

what i don't get is if multi-bit dacs are superior then why bother creating delta-sigma dacs anyhow? seems like a wasted time and effort.


if I understood correctly, multi bit dacs are still delta sigma, with noiseshaping
 
I am perhaps on deep water here since I haven't read up on
these techniques, but Kueis explanation puzzles me a bit (joke
unintentional). I have always thought that delta-sigma is some
variation/extension on delta modulation, but now kuei says it
is actually PWM, which is not the same thing, at least it wasn't
20 years ago when I took the telecom course and we had
labs on delta modulation. So if it is PWM, where does the
name delta-sigma come from?? And in the case it is actually
some form of delta modulation, and not PWM, what does the sigma refer to?

:confused: :scratch:
 
Origin of the term "delta-sigma"

In the most basic form, a delta-sigma modulator consists of a subtractor (delta) followed by an integrator (sigma) followed by a quantizer. The input signal is connected to the positive input of the subtractor and the quantizer output connects to the negative input. This forms a servo loop that tries to drive the signal at the quantizer output to match the input signal. Since the loop has high gain at low frequencies (due to the frequency response of the integrator) the error in the quantized signal (quantization noise) is greatly attenuated at low frequencies, but can be quite large at high frequencies.

FWIW, for years there was a debate over whether the correct name is "delta-sigma" or "sigma-delta", owing to the tradition among control theorists to name their systems from the inside out instead of from front to back.
 
Koinichiwa,

Bricolo said:

That's very well explained :) (I suspect Kuei Yang Wang for being the writer from the books called "Idiots guide for ..." :D)

Please note that the explanation was EXTREMELY simplified, mostly to drive the basic points home.

Bricolo said:

As delta sigma DACs are a kind of PWM, where does theyr name come from?

I seem to remeber that it was based on the Delta (Difference) of Sigma (Time). Note that I said that DS DAC's CAN be viewed as PWM/PDM, the operation is not exactly identical, especially not once multi-level modulators are included.

Bricolo said:

"The strong ultrasonic content in the output of a Delta Sigma DAC makes many a non-oversampling DAC look quite decent"
Wouldn't it be the opposite? Since you have HF output, non os would not be a good idea.
Where does this ultrasonic content come from? Isn't the 5-7th order lowpass enough to filter it?

As siad, all this ultrasonic noise is basically the result of reducing the bit-depth. Only, instead of being liberally smeared across the
Audio Range it is instead "pushed" outside the audio band.


Bricolo said:

Does this say that R2R DACs haven't any HF on the output? Making them more suitable to a non os DAC?

I would not want to try to make a non-oversampling Delta Sigma DAC, unless the sample rate is very high (DSD/SACD is an example of Non-Ovesampling Delta-Sigma).

So for CD, if you don't want the Digital Filter - Multibit only.

HeadSh0T said:
what i don't get is if multi-bit dacs are superior then why bother creating delta-sigma dacs anyhow? seems like a wasted time and effort.

A decent quality Multibit DAC costs a lot of money to make.

The Burr Brown PCM1704 is rated as 24-Bit/96KHz DAC. It requires two DAC Chips and an external digital filter. Total cost in qanteties of 1000 around USD 40. It also needs a complex analogue stage, with a high performance Op-Amp as I/V converter and another as Filter and several seperate, well filtered and regulated powersupplies.

If you do it well I'd think the cost to implement a conventional 96/24 DAC subsystem using PCM1704 will be near $ 100 US for a manufacturer who makes at least 1000pcs annually.

The Burr Brown PCM1728 is also rated as 24-Bit/96KHz DAC. It is a single chip including Digitial filter, 2 channels of DAC, part of the Analogue filtering, needs a single supply and a single Op-Amp as Buffer/Filter per channel. Budgetary pricing is 3 Bucks for the DAC Chip and maybe another 6 Bucks for the needed passive parts, analogue stage with regulators etc.

So the price differential between shipping your "Gizmo" with a 96/24 DAC based on DS and a proper Multibit DAC is around 1:10. Or to turn this around, the contribution to the endcustomer retail for the "Gizmo" (Say a "High Perfomance DVD/CD Player) will be around $80 - $160 for the DS DAC but $800 - $1,600 for the multibit, assuming they sell in equal quanteties.

Does that answer this question?


Bricolo said:

if I understood correctly, multi bit dacs are still delta sigma, with noiseshaping

You understand incorrectly. You CAN combine Dleta/Sigma technology and Noiseshaping with Multibit DAC's. Commonly the low level linearity of Delta/Sigma is better than in multibit DAC's and more consistently so with production, at the same time multibit has less out of band noise and seems to do better at higher levels.

The best modern DAC Chip's (according to specifications anyway) usally combine the two fundamental technologies. I have heard the Burr Brown PCM1738 sounding rather decent, in fact about as good with CD as a fullblown implementation of the PCM1704 with Parallel DAC's and HDCD Filter.

Christer said:
I am perhaps on deep water here since I haven't read up on
these techniques, but Kueis explanation puzzles me a bit (joke
unintentional). I have always thought that delta-sigma is some
variation/extension on delta modulation, but now kuei says it
is actually PWM, which is not the same thing, at least it wasn't
20 years ago when I took the telecom course and we had
labs on delta modulation. So if it is PWM, where does the
name delta-sigma come from?? And in the case it is actually
some form of delta modulation, and not PWM, what does the sigma refer to?

Actually, PWM is a little easier to understand than Delta/Sigma, so I used it as the explanation. While different in certain aspects, in the end you end up with a pulse width or pulse density modulated signal that look very similar on the 'scope and achieve very similar results in similar ways.

Other than for certain practical application issues and academia I would lump PWM, PDM, DS and co all together as special cases of the same principle in DA (or AD) conversion, as distinct to PCM conversion, which also has several subclasses of operation.

Okay?

I did warn in the beginning that I had to brutally simplify for clarity and understandability to "Joe average" (sorry Joe, nothing demeaning is implied - we all are specialists in something).

Cool Dudes?

Sayonara
 
A better expression would be PDM (for pulse density modulation).
While PWM uses a fixed frequency rectangular output waveform whose duty-cycle is signal dependant.
D-S modulation uses a signal representation with a constant sample-rate. This results in an output switching waveform whose transitions take place at integer multiples of the sampling frequency. But you don't have a constant switching frequency anymore.
It is high for low output voltages and low for high output voltages.

I have attached an example of a 20 kHz signal. The green trace shows the unfiltered D-S output. One could theoretically see that all the transitions take place at integer multiples of 250ns. The red trace shows the signal after passing a 2nd order 120 kHz "reconstruction filter". The switching residual can be clearly seen as high-frequency noise.

Regards

Charles
 

Attachments

  • ds_output.zip
    36.1 KB · Views: 425
mikewu99 said:
Origin of the term "delta-sigma"

In the most basic form, a delta-sigma modulator consists of a subtractor (delta) followed by an integrator (sigma) followed by a quantizer. The input signal is connected to the positive input of the subtractor and the quantizer output connects to the negative input. This forms a servo loop that tries to drive the signal at the quantizer output to match the input signal. Since the loop has high gain at low frequencies (due to the frequency response of the integrator) the error in the quantized signal (quantization noise) is greatly attenuated at low frequencies, but can be quite large at high frequencies.

FWIW, for years there was a debate over whether the correct name is "delta-sigma" or "sigma-delta", owing to the tradition among control theorists to name their systems from the inside out instead of from front to back.

Well, then it is identical to what was called delta modulation
when I took the telecom course long ago. It seems delta-sigma
(or sigma-delta) is just a more modern term for the same thing.
That clears up the issue for me. Thanks.
 
Kuei,

OK, it seems you did not really mean that delta-sigma is PWM,
only that it is similar in some sense. I agree to that, but I don't
think it appropriate to mix them up to the point of saying that
delta-sigma "is" PWM even in a simplified presentation as you
did, which is why I was confused.

Anyway, issue cleared.
 
There is a significant difference between delta-modulation and delta-sigma (or sigma-delta) modulation.

The delta modulator has the integrator in the feedback path whereas the delta-sigma modulator has the integrator(s) in the forward path.

The result is that with delta modulation the pulse density is proportional to the signal slew-rate, where the pulse density is proportional to the signal amplitude for delta-sigma modulation.

Regards

Charles
 
phase_accurate said:
There is a significant difference between delta-modulation and delta-sigma (or sigma-delta) modulation.

The delta modulator has the integrator in the feedback path whereas the delta-sigma modulator has the integrator(s) in the forward path.

The result is that with delta modulation the pulse density is proportional to the signal slew-rate, where the pulse density is proportional to the signal amplitude for delta-sigma modulation.

Regards

Charles


I have been thinking about this now and then, trying to figure
out what you mean, but I don't get it. This is my understanding
of it. In both cases we have analog in and binary out. In the case
of delta modulation the input goes to in+ of a comparator, the
output from the comparator is clocked and used as output. This
output is also fed back to in- via an integrator. This seems to
agree with what you say. However, for delta-sigma you say
the integrator is in the forward path, and I fail to see where
to put it unless the whole design is fundamentally different.
Since we want binary out, I cannot quite envision any other
place to put the integrator than in the feedback path. What do
I miss? Or are my fundamental assumptions wrong??
 
Status
This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.