To DSD or not to DSD?

[Hiro]

Most of the time DAC chips perform badly when converting from PCM input because 1) they have inadequate oversampling filters, 2) they use low-order modulators.

 

[...]

 

Decimation is just a useless step because eventually you'll need to oversample and modulate again for the playback. Plus you get time domain effects of the decimation filter.

 

Similar to digital cameras, you should be capturing the raw converter output when recording. Similar to using DNG with digital camera instead of JPG. Taking PCM out of modern ADC is analogous to storing to a JPG with DSLR.

 

Exactly.

 

Give me a direct output of a sigma delta modulator over its downsampled and decimated copy any day.

[davide256]

Specs are essentially useless if they are not looking in the right area for what matters. Saw this in the 80's with direct drive turntable manufacturers advertising arguments vs. turntables like Linn Sondek and Sota Sapphire. Use ears and A/B swaps first, only go to specs when a clear trend emerges that allows you to correlate observed results vs. potentially related specs. Remember how THD was misused as a measure of amplifier quality in the 70's and 80's? Now we know better that the ear is more sensitive to odd order harmonic distortion than even order distortion.

[Miska]

This is a problem because, like I already said, an 1-bit quantizer cannot properly be dithered. If, on the other hand, the ADC uses a Sigma-Delta modulator that IS 1-bit, that means it also uses an 1-bit quantizer because an 1-bit Sigma-Delta modulator IS an 1-bit quantizer. So (again, please correct me if I am wrong) the problem persists both ways.

 

Number of output bits have nothing to do with modulator dithering. You need to get your thinking outside the PCM box.

 

If you think that 6-bit SDM (Sabre) would be like 6-bit PCM, then it would become essentially a 1-bit immediately when the level drops below -36 dB. With Resonessence Labs HERUS or exaSound DAC that would be quite normal volume setting for headphone listening so it would be always 1-bit. But as ESS emphasizes, the 6-bits are not in parallel, but in time.

 

While it is true that choosing a higher order Sigma-Delta modulator shifts more noise out of the audible band, it is also true that there exists an optimum order when designing such a modulator, due to practical constraints.

 

So far I've found 7th order to be pretty much perfect. It allows extremely good performance. 3rd order modulators used in many DAC chips are far from it, but they don't have enough master clock cycles to run better. So they try to compensate by adding bits, but that needs exponentially growing number of conversion elements.

[spdif-usb]

Number of output bits have nothing to do with modulator dithering. You need to get your thinking outside the PCM box.

 

If you think that 6-bit SDM (Sabre) would be like 6-bit PCM, then it would become essentially a 1-bit immediately when the level drops below -36 dB. With Resonessence Labs HERUS or exaSound DAC that would be quite normal volume setting for headphone listening so it would be always 1-bit. But as ESS emphasizes, the 6-bits are not in parallel, but in time.

 

 

 

So far I've found 7th order to be pretty much perfect. It allows extremely good performance. 3rd order modulators used in many DAC chips are far from it, but they don't have enough master clock cycles to run better. So they try to compensate by adding bits, but that needs exponentially growing number of conversion elements.

An Introduction to Delta Sigma Converters

 

In the article linked above, the section titled "Multi-Bit Converter" explains how it's possible to reduce noise by means of increasing both the internal signal width of the Delta-Sigma modulator and its output signal width from 1-bit to N-bit. The Delta-Sigma ADC that is pictured in Figure 12 uses a Multi-Bit Modulator (MBM), the N-bit output of which is transformed to PCM via a digital low pass filter / decimator. You could remove this digital low pass filter / decimator from the signal path. Doing so will enable you to store the N-bit output of the modulator directly, but it won't be DSD.

[Superdad]

In the article linked above, the section titled "Multi-Bit Converter" explains how it's possible to reduce noise by means of increasing both the internal signal width of the Delta-Sigma modulator and its output signal width from 1-bit to N-bit. The Delta-Sigma ADC that is pictured in Figure 12 uses a Multi-Bit Modulator (MBM), the N-bit output of which is transformed to PCM via a digital low pass filter / decimator. You could remove this digital low pass filter / decimator from the signal path. Doing so will enable you to store the N-bit output of the modulator directly, but it won't be DSD.

 

So what? We all know that multi-level SDM (less confusing to the twos-complement PCM-minded than calling it multi-bit) is and would be better (for noise, filtering, etc.) than 1-bit DSD. Miska's own DSC1 demonstration design is 33-level, and he could easily set his HQ Player SD modulators to output high-rate multi-bit streams instead of single-bit. But there are no popular standards, drivers, and interface buses currently to transmit such formats to yet-to-be-designed DACs for this (at least not in the consumer space). Not that it could not be done...

[spdif-usb]

So what? We all know that multi-level SDM (less confusing to the twos-complement PCM-minded than calling it multi-bit) is and would be better (for noise, filtering, etc.) than 1-bit DSD. Miska's own DSC1 demonstration design is 33-level, and he could easily set his HQ Player SD modulators to output high-rate multi-bit streams instead of single-bit. But there are no popular standards, drivers, and interface buses currently to transmit such formats to yet-to-be-designed DACs for this (at least not in the consumer space). Not that it could not be done...

Personally, I don't find it confusing the fact there exists a fairly huge difference between what comes out of the MBM and what comes out of the digital LPF / decimator (nor between what comes out of the MBM and what DSD is, for that matter...). So I still have absolutely no idea why I would, as Miska has suggested, need to get my thinking outside the PCM box. And no, Multi-level is not really less confusing than Multi-bit, as both 1-bit and Multi-bit are Multi-level.

 

So, in answer to your question "So what?", 1-bit Delta-Sigma Modulators are a bad idea because they cannot properly be dithered, and, similarly, conversion elements that transform the Multi-bit output of Multi-bit Delta-Sigma Modulators to DSD are a bad idea because it can be shown that they (the conversion elements) cannot properly be dithered either. Just because the ESS ES9102 chip uses a trademark HyperStream modulator, which is only a 5th order modulator instead of the "better" 7th order that Miska was talking about, doesn't necessarily also mean that 7th order Delta-Sigma Modulators can be good enough to fully compensate for the dither issue inherent of the DSD format. So you need to get your thinking outside the PDM box.

[Ipoci]

So what? We all know that multi-level SDM (less confusing to the twos-complement PCM-minded than calling it multi-bit) is and would be better (for noise, filtering, etc.) than 1-bit DSD. Miska's own DSC1 demonstration design is 33-level, and he could easily set his HQ Player SD modulators to output high-rate multi-bit streams instead of single-bit. But there are no popular standards, drivers, and interface buses currently to transmit such formats to yet-to-be-designed DACs for this (at least not in the consumer space). Not that it could not be done...

 

Start a crowdfounded project … That's the only way I know to go for a crazy idea ( unless you won a lottery ).

 

Have a nice day, Massimiliano

[spdif-usb]

John Siau: Benchmark Audio Guru | Real HD-Audio

[Hiro]

So what? We all know that multi-level SDM (less confusing to the twos-complement PCM-minded than calling it multi-bit) is and would be better (for noise, filtering, etc.) than 1-bit DSD. Miska's own DSC1 demonstration design is 33-level, and he could easily set his HQ Player SD modulators to output high-rate multi-bit streams instead of single-bit. But there are no popular standards, drivers, and interface buses currently to transmit such formats to yet-to-be-designed DACs for this (at least not in the consumer space). Not that it could not be done...

 

Miska's multi-level Delta Sigma Digital to Analog converter shows that you can play DSD in bit-perfect NOS mode on a multi-level Delta Sigma DAC (without the interference of PCM filters).

 

As for few-bit SDM output captured directly, it would still definitely be a Direct Sigma Delta format. And it is my understanding that ASIO is already capable of handling such stream.

[Miska]

Start a crowdfounded project … That's the only way I know to go for a crazy idea ( unless you won a lottery ).

 

Unfortunately that doesn't work in Finland, as crowdfunding is practically considered illegal. Couple of popular crowdfunding projects got stopped by the police and bureaucrats.

 

But I do what I can given the time and money I have, and have fun doing so.

[Miska]

As for few-bit SDM output captured directly, it would still definitely be a Direct Sigma Delta format. And it is my understanding that ASIO is already capable of handling such stream.

 

ASIO supports 8-bit "DSD-Wide", so it could be used for multi-level SDM. And more formats could be easily added as long as driver and application agree on the same ID.

 

DSC1 can be very easily modified to accept thermometer code inputs and I've been thinking about ways how to practically interface such with a computer. I have some ideas but have not yet settled on which way I would do it. One way would be to plug an ARM based NAA board straight to the DAC.

[Miska]

So, in answer to your question "So what?", 1-bit Delta-Sigma Modulators are a bad idea because they cannot properly be dithered

 

Again, it can be properly dithered and number of bits the SDM outputs have nothing to do with dithering it.

 

Class-D amps are also 1-bit, so if you don't like 1-bit, don't use any class-D amp either. Like Hyper nCore etc. Bruno Putzeys latest DAC design is also 1-bit, and surprise surprise it's also has 7th order modulator:

http://mola-mola.nl/index.php/dak

and the performance figures don't look bad at all. I also recommend taking a look at his blog:

http://mola-mola.nl/index.php/blog

I just leave out those DSP processors and use PC for doing the heavy lifting.

 

I just recently checked out AD's TigerSHARC processors (faster than the SHARC used above) and it's only 1/4th the speed of single core of my old Core i5...

 

P.S. And HyperStream modulator outputs 1-bit wide bitstream...

[Miska]

An Introduction to Delta Sigma Converters

 

In the article linked above, the section titled "Multi-Bit Converter" explains how it's possible to reduce noise by means of increasing

 

Yet another document where the author doesn't know how to do things and uses extremely over-simplified diagrams that have nothing to do with any proper real-world implementation.

[spdif-usb]

Miska's multi-level Delta Sigma Digital to Analog converter shows that you can play DSD in bit-perfect NOS mode on a multi-level Delta Sigma DAC (without the interference of PCM filters).

Does the DAC even matter? The dither issue is already present in the DSD data that comes out of the ADC.

As for few-bit SDM output captured directly, it would still definitely be a Direct Sigma Delta format. And it is my understanding that ASIO is already capable of handling such stream.

Unfortunately however, DSD does not stand for Direct Sigma Delta. Instead, it stands for Direct Stream Digital, and the "Stream" part is referring to a 1-bit serial bitstream, i.e. NOT a Multi-bit datastream.

 

Furthermore, the undesired artifacts that result from a decimation filter in an ADC (and from an interpolation filter in a DAC) are not necessarily more undesired than the undesired artifacts that result from the aggressive filter that is required to be able to remove the massive amount of high-frequency noise that plagues the DSD format. In fact, I believe the opposite holds true.

[wisnon]

Spdif, what DSD playback machines have you listened to so far? Not being disparaging, I really want to know.

[Hiro]

Unfortunately however, DSD does not stand for Direct Sigma Delta.

 

But it's essentially just that.

 

It's worthy of note that Prof Yoshio Yamasaki, the inventor of DSD, prefers to call DSD a Direct Sigma Delta format.

 

And even Andreas Koch notes the following:

 

"Delta-Sigma modulation: the analog signal is converted directly to DSD with a very high sampling rate. Various algorithms are in use depending on the application and required fidelity. They can generate 1-bit DSD or multibit DSD oversampled at 64x or 128x compared to regular CD rate."

http://www.audiostream.com/content/qa-andreas-koch

 

As for the "massive amounts of noise" canard, with DSD128 the noise is already above 50kHz. FYI human ears sense sounds at frequencies between 20 an 20,000 cycles per second (Hz).

[spdif-usb]

Again, it can be properly dithered and number of bits the SDM outputs have nothing to do with dithering it.

This AFAIK has been thoroughly debunked by Lipshitz, Vanderkooy & Wannamaker.

[Ipoci]

…

 

But I do what I can given the time and money I have, and have fun doing so.

 

If something will start as crowdfounded project … I will definitely support it …

 

Have a nice day, Massimiliano

[Miska]

For me, this is good enough performance for 1-bit SDM:

 

Compared to TPDF-dithered 176.4/24 PCM:

 

And this is IMD performance I get from Resonessence Labs HERUS with DSD128 input (see level of 1 kHz difference tone):

 

Compred to IMD performance of the same with 32-bit PCM input:

[spdif-usb]

Spdif, what DSD playback machines have you listened to so far? Not being disparaging, I really want to know.

I have lost count of them. The latest one that I have heard was the PS Audio DirectStream DAC.

[spdif-usb]

For me, this is good enough performance for 1-bit SDM:

[ATTACH=CONFIG]14751[/ATTACH]

 

Compared to TPDF-dithered 176.4/24 PCM:

[ATTACH=CONFIG]14752[/ATTACH]

 

And this is IMD performance I get from Resonessence Labs HERUS with DSD128 input (see level of 1 kHz difference tone):

[ATTACH=CONFIG]14753[/ATTACH]

 

Compred to IMD performance of the same with 32-bit PCM input:

[ATTACH=CONFIG]14754[/ATTACH]

Whenever I judge the quality of sound, I do not rely on measurements alone as they often are wrong or misleading, or both. The simple fact those AES papers by Lipshitz, Vanderkooy & Wannamaker regarding DSD have mostly fallen on deaf ears is one of the things that keep reminding me of why I think it is naive to trust any such marketing ploys.

[Miska]

This is the kind of stuff that goes as input to the modulator in many DAC chips, such as AD1955 for example. Digital interpolation filter (sharp roll-off) takes input to 352.8/384 kHz and from there on, oversampling is done using sample-and-hold.

 

Or with typical kind of "slow roll-off":

 

And this is what goes as input to my modulator (poly-sinc-2s):

[Miska]

Whenever I judge the quality of sound, I do not rely on measurements alone as they often are wrong or misleading, or both. The simple fact those AES papers by Lipshitz, Vanderkooy & Wannamaker regarding DSD have mostly fallen on deaf ears is one of the things that keep reminding me of why I think it is naive to trust any such marketing ploys.

 

I do a lot of listening too. I demand both, good sound and good theoretical and technical performance (measurement results).

 

If measurement results are considered a marketing ploy, I'm fine with it. Anybody is free to listen with their own ears and make their judgement. I'm just sharing the information I have. I don't have a particular agenda, I support both PCM and SDM outputs and conversion from any input format to any output format, so I'm kind of format agnostic in that sense. HQPlayer can do 1536/32 TPDF-dithered (or noise shaped) PCM output too.

[spdif-usb]

But it's essentially just that.

 

It's worthy of note that Prof Yoshio Yamasaki, the inventor of DSD, prefers to call DSD a Direct Sigma Delta format.

 

And even Andreas Koch notes the following:

 

"Delta-Sigma modulation: the analog signal is converted directly to DSD with a very high sampling rate. Various algorithms are in use depending on the application and required fidelity. They can generate 1-bit DSD or multibit DSD oversampled at 64x or 128x compared to regular CD rate."

Q&A with Andreas Koch | AudioStream

 

As for the "massive amounts of noise" canard, with DSD128 the noise is already above 50kHz. FYI human ears sense sounds at frequencies between 20 an 20,000 cycles per second (Hz).

The "human ears sense sounds at frequencies between 20 an 20,000 cycles per second (Hz)" canard does not change the fact noise above 50kHz is massively amplified by 7th order Delta-Sigma modulation. In the case of an 1-bit system, which any system that is outputting an 1-bit serial bitstream (DSD) still is, the high-frequency noise is so much of a problem that it needs to be very aggressively filtered, i.e. by using a filter that is even more aggressive than the decimation filter DSD fans always keep crying about.

On top of that, proper dither of an 1-bit system is simply not possible in any way at all. (And I really very much hate to keep sounding like a broken record on this, but it has long been established by Lipshitz, Vanderkooy & Wannamaker in a series of IMO fairly groundbreaking AES papers, so I still think it is safe for me to completely ignore pretty much whatever it is that Andreas Koch would like me to believe instead).

[Hiro]

This is the kind of stuff that goes as input to the modulator in many DAC chips, such as AD1955 for example. Digital interpolation filter (sharp roll-off) takes input to 352.8/384 kHz and from there on, oversampling is done using sample-and-hold.

[ATTACH=CONFIG]14755[/ATTACH]

 

Or with typical kind of "slow roll-off":

[ATTACH=CONFIG]14757[/ATTACH]

 

And this is what goes as input to my modulator (poly-sinc-2s):

[ATTACH=CONFIG]14756[/ATTACH]

 

Very interesting and eye-opening.