• Welcome to ASR. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!

Modern Multi-Bit DAC vs Delta Sigma, specifically AKM's newest flagship, but also others

@mocenigo: I thought the PCM1791A used a version of van de Plassche's scheme that "rotated" among current sources (sinks)? Don't remember now, and am no longer sure how to reach the guys I knew at BB before TI took them over...

Segmentation, using unary (unit-weighted) cells for the MSBs and binary weighting for the lower bits, is very common in R-2R DACs. It is impractical to match (even with trimming) the MSBs otherwise (although other schemes using digital compensation are also popular).

Many devices have used charge pumps but they do have their own problems, including limited current, relatively high supply output impedance (various schemes to fix that are used), and limited current capacity. SMPS schemes are much more widely used these days.

This thread briefly discusses segmented DACs: https://www.audiosciencereview.com/...ital-audio-converters-dacs-fundamentals.1927/ An n-bit DAC requires roughly 2^-N matching among MSB cells, something very difficult to achieve for high-accuracy DACs. Segmented designs are one way, delta-sigma modulators another.
 
What does 4 level delta sigma modulation for the least significant bits mean?
I just looked through the 1791A's datasheet and didn't find any info alluding to its nature - how did you find out about this? And the resistor array?

Look at the 1791A data sheet
page 52. The same description is also in the 1793 data sheet.

You seem to know about the inner workings of DAC chips

Not really, I have read some introductory chapters in books about DACs, and some papers, and I understand the mathematics. I also have read a few data sgeets.
- have you ever come across a designs that used a charge pump to increase the source voltage for the most significant bits to increase accuracy?

Intuitively this risks to be quirte imprecise. The gain would have to be absolutely perfect.

Or one that uses extremely HF PWM to regulate the current of the least significant bits, which use the next LSB that's not PWM'd as reference to operate (set as a fraction of the reference)?
If you've never seen either thing implemented, I think they'd both be pretty cool to see in a design - as long as there's nothing inherently wrong with either lol

I think that the 1791/2/3 does something similar.
 
@mocenigo: I thought the PCM1791A used a version of van de Plassche's scheme that "rotated" among current sources (sinks)? Don't remember now, and am no longer sure how to reach the guys I knew at BB before TI took them over...

It is a bit more complex than the scheme I presented. The 6 MSBs are turned into ICOB, and the LSB are joined with the MSB and modulation is applied to that and reduced to 5 levels (not 4, i was wrongo). So the first set of values from 0 to 62 and the second one from 0 to 4 are added. Ao you have to represent values from 0 to 66, and this is passed through DWA. It may be rotating or randomising in some other way, so I do not know how DWA is performed.

[...]

This thread briefly discusses segmented DACs: https://www.audiosciencereview.com/...ital-audio-converters-dacs-fundamentals.1927/ An n-bit DAC requires roughly 2^-N matching among MSB cells, something very difficult to achieve for high-accuracy DACs. Segmented designs are one way, delta-sigma modulators another.

Thanks!
 
It is a bit more complex than the scheme I presented. The 6 MSBs are turned into ICOB, and the LSB are joined with the MSB and modulation is applied to that and reduced to 5 levels (not 4, i was wrongo). So the first set of values from 0 to 62 and the second one from 0 to 4 are added. Ao you have to represent values from 0 to 66, and this is passed through DWA. It may be rotating or randomising in some other way, so I do not know how DWA is performed.
Had to look it up again, wasn't sure what "ICOB" and "DWA" were (senility, and everyone invents their own TLAs). I see it does use a DS section for the lsbs (not the MSBs, forgot about that). IIRC the design was at least partly to reduce the oversampling ratio required for 24 bits by using a more conventional MSB segment with a DS lsb "segment". The block diagram is a bit confusing, probably intentionally. DWA is in the digital domain. Old design, still good.
NP; that was more for others since you seem to understand it already.
 
Look at the 1791A data sheet
page 52. The same description is also in the 1793 data sheet.
Cool DAC.
Intuitively this risks to be quirte imprecise. The gain would have to be absolutely perfect.
Before you replied, DonH56 saw and also responded, interestingly he thinks he's seen an implementation
I think that the 1791/2/3 does something similar.
the 1791 didn't mention in its "theory of operation" section you linked me to - are the 2/3 are different? I think it'd probably be a bit redundant for PWM + DWA on a 5 bit sigma delta DAC, but that's just my initial assumption
 
Last edited:
Cool DAC.

Before you replied, DonH56 saw and also responded, interestingly he thinks he's seen an implementation

the 1791 didn't mention in its "theory of operation" section you linked me to - are the 2/3 are different? I think it'd probably be a bit redundant for PWM + DWA on a 5 bit sigma delta DAC, but that's just my initial assumption

Page 53, not 52, of the linked document. It is described there.
 
To the mod who moved this OP to a massive thread about DAC sound signatures (and everyone else reading because I added more info to the thread in here too): this thread is not about sound signature, it's about the internal operation of recently manufactured multi-bit DAC chips (specifically the new AK4499 Velvet, but also others), which I've come to understand may be imposters compared to others. Imposters? Imposters! :not true multi-bit chips like the ones made in the past, but somehow a hybrid of delta sigma + multi-bit, which, being different, may affect performance. At the end of the thread I did have some words on the sound quality of my L70 Velvet, because to me, it sounded vastly superior to the [too many] other Topping DACs I have. Fun fact: the only reason I currently know the AK4499 Velvet in my E70 Velvet is a multi-bit design, is because I was looking for a possible explanation for my perceived [but maybe imagined..] difference in clarity between it and my DX1, E30 II, E30II lite, and E50 (I do have reasons for all of those except the E50, which I planned to return but missed the window for). ANYWAY! That bit at the end of the thread was very secondary to its primary focus, which, again (to be sure), is not discussing the implementation of the AK4499 Velvet in the Topping L70 Velvet, but the AK4499 Velvet itself: information on its design and basic operation, and maybe information on other, more recently manufactured multi-bit DACs if they're similar. I removed what I had at the end of the OP (regarding sound comparison) so there is no confusion.

Thread:
I've got AKM's 4499EX in my Topping E70 Velvet. I'm under the impression that the 4499EX is a two-chip solution - the 4499EX + comes with a 4191EQ

From AKM: "The AK4191EQ is a new concept Multi-bit stereo Premium Digital Data Converter employing VELVETSOUND™ technology. By using the AK4191EQ to process the digital signal for D/A conversion, we have minimized the effects of digital noise within the analog output, resulting in a perceived improvement of the ratio of signal to noise. The AK4191EQ has a built-in digital filter with multi-bit sigma delta modulator and 256 times oversampling processing. A wide variety of music can playback by inputting 1536kHz PCM data and DSD1024 data."

I understand this to mean the AK4191 is the chip that goes between the digital source (coax, spdif, USB, HDMI etc., etc.) and the AK4499.

What I don't get, though, is why there's a sigma delta modulator in it though... wouldn't that nullify the [alleged] benefit of the AK4499 being multi-bit?

Something else that's confusing is AKM says the 4499 and 4191 are separate to keep the analog filtering and digital processing in separate packages, but then they go ahead and say the 4919EQ processes the digital signal so that there's less noise on the 4499's eventual output... So if the 4191's input is already digital, what's this sigma delta going on for? T

And another thing: the low-pass filter on the analog audio output... you know, the thing that usually brings things down to -60dB to -100dB by 24kHz or so with a 44.1kHz... is that still in the 4499? Or is it in the 4191? So much of what AKM has said appears to be contradictory or just plain impossible!

Also, with the AK4499 Velvet (technically AK4499EX, and when packaged with the 4191EQ, called the AK4499EXEQ):
Top of Page 1 of datasheet for proof:
View attachment 352942
Where is its resistor bank? Is it in the chip???

I'm pretty new to exploring the intricacies of DACs past their advertised specifications and audible sonic qualities, so sorry for if any of the questions are pretty stupid... (Not an excuse but a reason: I had a desktop USB/toslink/coax DAC I was really happy with for almost 10 years, and it broke (accidentally, by my hand :( ), and now I'm interested in learning more about DACs. I do think I've been lucky to find my replacement as quickly as I have, but the journey isn't over yet -I'm still trying to learn more

And finally (& maybe most importantly...) : is the AK4499 a good example of a modern multi-bit DAC - basically made the same way but with slightly newer manufacturing technologies -, or is it an imposter - a delta-sigma DAC masquerading as multi-bit (like I said, I haven't been looking long, but I've briefly come across people saying some modern multi-bit DACs are weird hybrids or something... from what I've been able to extrapolate from what's been suggested here, there, and everywhere: a delta-sigma type DAC with unconventional output is run in a way (maybe at a higher frequency?) that it's able to output like a multi-bit DAC through multiple pins, but there are less pins and resistors and the clock is doubled up (well not doubled, but maybe 10x - it's like 5 bits instead of 20 or 24 bits). I don't know, it's messed up what I've seen and I'm tired of guessing so I'm hoping someone knows what's going on!
This is a very clear and enlightening discussion of the AKM velvet implementation! Thanks.
 
I just don't get the need to use the term "imposter". A "pure" multi-bit DAC is not really desirable.

These modern hybrids are using multiple weighted 1-bit DACs instead of trying to give every single bit of the original information a weighting.

It gives you the benefit of neither having to run a 1-bit DAC at a stupidly high frequency, nor reaching the downsides of weighting more than a handful of bits.

There's really nothing to gain from "purity" in this case.
 
Because you’re looking at the wrong datasheet. The AK4191 does the modulation bits. It can do direct DSD. Which is rather silly if you have a 7-level DAC on hand… why not actually take advantage?
Sorry for my ignorante but how you convert 1 bit Delta Sigma DSD to 7 bits?
Also, the so called Sound Color on the AK4191+4499EX seem to be different ways to do x128 and 256 Upsampling. How is this done to 2.8 Mhz 1 bit DSD?
I own an SMSL D400 PRO that uses this AKM combo IC's, and I think It sounds outstanding for its price.
 
Sorry for my ignorante but how you convert 1 bit Delta Sigma DSD to 7 bits?
Also, the so called Sound Color on the AK4191+4499EX seem to be different ways to do x128 and 256 Upsampling. How is this done to 2.8 Mhz 1 bit DSD?
I own an SMSL D400 PRO that uses this AKM combo IC's, and I think It sounds outstanding for its price.
Convert to a higher bit depth first (so effectively make it PCM), or in case of DSD direct, you just don’t ;)
 
Sorry for my ignorante but how you convert 1 bit Delta Sigma DSD to 7 bits?
In modern D/A converter chips, the shaped quantization noise inherent of DSD signals is filtered out digitally, usually by some kind of FIR filter (earlier DAC chips often used a simpler comb-filter, which can be seen as an FIR filter where all coefficients are at unity), although at least one DAC maker (ESS) uses IIR filters. Either ways, the output of the digital filter is obviously a multibit code, since it is the result of addition of multiple samples. Sometimes, the resulting multibit output is re-modulated by a following multilevel sigma-delta modulator, whose output is also multibit (few bits). Along those successive digital processings, the sample rate remains at the rate of the input 1 bit DSD signal or becomes even higher.
 
Convert to a higher bit depth first (so effectively make it PCM), or in case of DSD direct, you just don’t ;)
Thanks for your reply.
I have my SMSL D400 PRO set to fixed output without DSD processing, or Direct, I don't remember the exact name of the setting as I'm not at home at the moment and I can't have a look at the exact name of the setting. The truth is having DSD unprocessed opposed to just fixed output, with DSD converted to 7 bit Delta Sigma (at least this is what I understand is done when the unprocessed setting if off), sounds better.
Getting DSD converted to 7 bits sounds a bit more PCM-like.
The second setting of my interest, is the so called Sound Color, which are four different setting to do Upsampling/Oversampling to x128 or x256, with two settings for each Upsampling/Oversampling "speed". I don't know how can DSD can be oversampled, but my knowledge on the matter are limited. And I've read Principles Of Digital Audio by Ken C Pohlmann, that is a bit book (800 Pagés, includding the Index), but my edition is from 2012 and It may have been updated since then.
 
The truth is having DSD unprocessed opposed to just fixed output, with DSD converted to 7 bit Delta Sigma (at least this is what I understand is done when the unprocessed setting if off), sounds better.

I have already pointed out in messages #307 and #313 in this other thread that DSD is never unprocessed in the AKM AK4191 + AK4499 combination. Whatever the user accessible setting, DSD is always low-pass filtered inside the AK4191 by a digital filter whose output is a multibit signal, as explained above.

The second setting of my interest, is the so called Sound Color, which are four different setting to do Upsampling/Oversampling to x128 or x256, with two settings for each Upsampling/Oversampling "speed". I don't know how can DSD can be oversampled, but my knowledge on the matter are limited. And I've read Principles Of Digital Audio by Ken C Pohlmann, that is a bit book (800 Pagés, includding the Index), but my edition is from 2012 and It may have been updated since then.
To the best of my knowledge, oversampling DSD can be as crude as a simple sample and hold function. For instance, by oversampling DSD64 2x to obtain DSD128, the sequence of samples "1", "0", "1" separated by roughly 0.354 µs intervals in DSD64 becomes the sequence of samples "1", "1", "0" , "0", "1", "1" separated by roughly 0,177 µs intervals (half 0.354 µs) in DSD128, each original sample beeing hold once. It has sometimes been done that way, for example in the Cirrus Logic CS4397 DAC as explained in the patent relevant to this chip (column 6, line 1 to 10 of the description).

Realizing a more sophisticated oversampling requires a digital filter that is called an "interpolation filter", exactly as in PCM (I think it is explained in your book). Digital filtering implies going to a multibit representation of the resulting samples.
 
Last edited:
Have you actually used any basic controls in determining this? PCM or DSD do not have a sound...
I think, and is my humble opinion after having SACD players since 2003, that DSD DOES have a sound. It uses, if there hasn't been any PCM intermediate process when mastering an SACD, to have a softer and more analogue-like sound, despite being digital.
 
DSD is just a special form of PCM (1-bit but at a very high sampling rate). However, the advantage of a pure DSD path is that there os no decimation or upsampling filtering. Such filters have typically pre echos and pre echos are not masked well.
 
DSD is just a special form of PCM (1-bit but at a very high sampling rate). However, the advantage of a pure DSD path is that there os no decimation or upsampling filtering. Such filters have typically pre echos and pre echos are not masked well.
But I'm not sure the AK4191+AK4499EX despite having a direct DSD Mode there are two things to keep in mind:
1- The AK4191 receives all the digital data, regardless if It IS PCM or DSD. This does what it's called Sound Color, two different flavours of x128 and another two of x256 oversampling, for both PCM and DSD. These Sound Color cannot be defeated.

2-The AK4499EX is a Delta Sigma 7 bit D/A Converter. So in the AK4191 1 bit DSD has to be converted/remodulated to 7 bit, even if the "unprocessed DSD Mode" is ON.
 
pre echos are not masked well.
But pre-echo don’t really show up in actual program material. They show up when you try to capture signals that cover greater than half the sample rate. Music don’t really have those. Most likely candidates are then around clipping and maybe some artificially generated sounds. In practice though, pre-echos are not a concern.


However, the advantage of a pure DSD path is that there os no decimation or upsampling filtering
Almost no such material exists because you cannot digitally process such material very much.

I think, and is my humble opinion after having SACD players since 2003, that DSD DOES have a sound. It uses, if there hasn't been any PCM intermediate process when mastering an SACD, to have a softer and more analogue-like sound, despite being digital.
So you think you would pass a double blind test?
 
I think, and is my humble opinion after having SACD players since 2003, that DSD DOES have a sound. It uses, if there hasn't been any PCM intermediate process when mastering an SACD, to have a softer and more analogue-like sound, despite being digital.
There is more than the nature of the source digital format that may explained your opinion.

For one example : it is stated in the Cirrus Logic patent I provided an hyperlink to in my previous message that one implementation of the DSD digital filter have a gain less than unity. That means that, all other things being equal, that the DAC output level would be less when playing back DSD with that filter on than when playing back PCM, even when the DSD and PCM source materials have exactly identical levels to begin with. Thus, replaying DSD without precaution to align the output level with an identical PCM source in order to compare the former to the latter might mislead one to believe that DSD has a "softer" sound than PCM, whereas it simply has a bit less level.

That is the type of reasons than voodooless insists on controlled listening tests: to avoid such basic mistakes, when they exist.

In order to compare at home DSD and PCM playback in equal terms, perhaps the most convenient way is to procure oneself with a preamplifier having a level trim feature for every input. With that kind of preamp, it would be possible to switch instantly between two identical source devices, one playing DSD and the other PCM, whose potentially different output levels when replaying one of the digital format (which can be a few tens of a dB to several dB depending on any particular device) have been compensated in the preamp. Unfortunatly, preamplifiers that incorporate that feature are usually not cheap.
 
Last edited:
Back
Top Bottom