The Multibit DSD debate

[The Computer Audiophile]

As more and more DSD-capable DACs become available ( see this wonderful database that Jesus from Sonore started, and I help co-ordinate......https://docs.google.com/spreadsheet/ccc?key=0AgVhKcl_3lHfdFVyenBBNjNpQ2lieG81WGpqQTNfVUE#gid=0) the debate rages on about what is really DSD and what is simply converted PCM (behind the scenes). Or more appropriately, does a DAC's architecture and chipset (or lack thereof) allow it to process DSD and/or PCM more effectively than others...i.e better sound? Because, frankly, who cares how it gets to the speaker, as long as its beautiful musically, right?

 

Well, as we all try to educate ourselves on hirez music, and find that the suppliers of said music want a premium for this stuff, we are more and more interested in making sure our DAC purchases are going to play back our favorite music properly and conveniently. We DO care how it arrives at our speakers; we want to be future-proof to some extent, yet are confused by all the options. One-bit, mutlibit SDM, multbit PCM, chipless, R2R, Ring DAC, FPGA, etc.

 

Well...enter a very well regarded DAC player, Berkeley Digital. They have a new $14k DAC that, according to their news release (which rumor says was not written by the principals, let alone Dr Keith Johnson, but possibly by a dealer of theirs) clearly, refreshingly (and likley erroneosuly) states it, like 99% of all DSD-capable DACs out their, is multibit PCM and does not process DSD directly, but instead converts to PCM...so they want the user to convert DSD to PCM outside the DAC, prior to its entry, to save the noise and heavy lifting from happening near the DAC chip.

 

OK about the heavy lifting, but it's the first part of that (99% of DSD-capable DACs are multibit PCM) that I and others question. So much so that I wanted to start a thread about this debate rather than hijack the Berkeley annoucment thread. I'd like other DSD-capable DAC mfhgers or knowledgeable folks to "dummy down" the tech talk so that many of us can understand. We have several threads here about math; this one needs to be about differences among DSD-capable DACs that potential purchasers can understand. I started with Michal Jurewicz, from Mytek. His comments are below.

 

First though, Here is the text from the Berkeley release comments:

 

"Careful consideration was given to providing the highest possible reproduction of DSD files by the Alpha DAC Reference Series. 99% of modern DAC’s, including the Alpha Reference Series use multi-bit D/A converters because they provide better performance than 1-bit converters – even those who advertise “native” DSD compatibility. So, at some point, the 1-bit DSD stream must be converted to multi-bit for all of those DAC’s.

 

We could, like many other manufacturers, convert 1-bit DSD to multi-bit within the Alpha DAC Reference Series and show “DSD” in the front panel display. That would be the easiest approach from a marketing perspective. But that would also mean increasing the amount of processing in the DAC during playback which would degrade audio quality, and audio quality is the reason the Alpha Reference Series exists.

 

Fortunately, virtually all reproduction of DSD files using external DAC’s occurs with a computer based music server as the source. If the 1-bit DSD to multi-bit conversion is done first in the computer it can be performed with extremely high precision and superior filtering that preserves all of the content of the DSD file. Computer DSD to multi-bit conversion can be at least as good as that performed in a DAC and without adding processing noise near or in the D/A converter chip. Another advantage of computer based DSD to PCM conversion is that if higher performance DSD versions such as DSD 4x appear in the future they can easily be supported with a software upgrade.

 

For all of those reasons, DSD capability for the Alpha DAC Reference Series is provided by an included state of the art software application that provides either real time conversion of DSD 1x and DSD 2x to 176.4 kHz 24 bit PCM during playback or conversion to 176.4 kHz 24 bit AIFF or WAV files. The software application is included in the price of the Alpha DAC Reference Series and is compatible with either Windows OS or Mac OS based music servers."

 

So, Michal Jurewicz, Mytek founder and chief designer, responded:

 

That quote (above) cannot be considered accurate.

 

1) Multibit DSD is not PCM.

 

PCM is typically 24 bit of the whole sample value while multi bit DSD is typically 5-6 bits of DIFFERENCE between adjacent samples. In 1 bit DSD this difference is binary(0 or 1) while in multi bit DSD it's the same difference between samples but quantized with 5-6 bits (Sabre has a 6 bit DAC). Multibit DSD is BETTER than 1 bit DSD, so there is nothing wrong with going 1 bit> 6bit which is what all these DAC chips do.

 

2) The " purity of 1 bit " is a marketing spin resulting from how DSD was marketed by Sony, but it's not 1 bit that makes it sound good, but it's nature. What makes DSD sound good is the conversion method (differential btwn samples- not the actual sample) and no digital filters , it's simply cleaner. 1 bit DAC conversion has the appeal of simplicity (perfect linearity theoretically) in the early 2000s. We are now well past that with performance requirement. We need more bits to resolve detail better.

 

3) All PCM conversion today (99% of ADCs) is a derivative of multi bit DSD.

PCM is always a subset of that, so naturally it can only have less information, not more.

 

In our latest experiments PCM has to have at least 32 bit to compete with DSD low level resolution and 352kFS to compete with lightness of DSD.

 

In the experiments with our newest prototype ADC, when we reduce wordlength of ADC from 32 to 24 bit, signal deteriorates.

 

Hope this clarifies things somewhat.

 

 

As I thought, it seems we have at least four categories of managing DSD within the DSD-capable DACs. People, please correct me.

1) one-bit architecture like EMM, Meitner and Playback Designs. The downside is that they upsample (or convert) PCM to a DSD multiple before it hits the analog stage. Michal's comments about one-bit DACs are, of course, his.

2) multibit-DSD like Mytek and other ESS SABRE designs.

3) multibit PCM like Berkeley. They convert everything to PCM before the chip..

4) chipless analog-filtered-only designs like Lampizator. They are DSD-only (for this process). In this case maybe we do the converse of the Berkeley recommendation; we ask a player like JRIver to convert all PCM to some DSD multiple. I can't speak for the Lampi PCM side, but it must require a traffic cop or separate USB signal path?

 

I'm not sure where the esoteric "logic" designed ring DAC (dCS), R2R (MSB, TotalDAC) and FPGA (Chord) designs fit in here.

 

I'd love for this thread to help folks understand the clear design choice differences, and choose wisely among them. I am not asking that Berkeley be chastised; far from it. Someone in their network took the marketing definitions a little too broadly (or so it seems). However, if they can process a converted DSD-to-PCM piece of music and sound great, so be it!! Me, I've yet to hear that category make DSD sound as good as categories 1 and 2...but there's a boatload of incredible and mandatory PCM out there that we all own, so the choices are not black and white.

 

 

 

 

The following is relevant to this thread as well as the Berkeley thread where I initially posted it.

 

 

Hi Guys - I received a response from Berkeley Audio Design regarding DSD, PCM, and some items that have been said in this thread. What follows was written completely by Berkeley Audio Design and doesn't necessarily express the opinion of CA.

 

 

DSD versus PCM and the Berkeley Audio Design approach

12/19/13

 

 

 

 

In a recent thread on the Computer Audiophile site, The Multibit DSD debate, an obvious point is stated correctly; it is the sound quality of an overall system that really matters and that sound quality is judged by listening to analog signals. There are many factors that affect sound quality in any specific system, and attributing the differences to only one factor is always an oversimplification.

 

 

 

 

There is also a great deal of confusion between DSD and delta-sigma modulation evident in the discussion posts, with DSD being frequently used to refer to delta-sigma. For the purposes of an audiophile discussion, DSD and PCM are delivery / storage formats, and are distinct from the design of data converters and the processing that goes on as part of the conversion process. Delivery / storage formats and the design of data converters should be analyzed separately, since various designs use varying combinations of techniques.

 

 

 

 

Direct-Stream Digital (DSD) is the trademark name used by Sony for the raw data output of a 1-bit delta-sigma modulator (DSM), originally coming straight from the A/D, which can then be sent to a 1-bit D/A. This is an idea that is appealing in its simplicity, but in practice it is not so simple, and has its own problems and sonic signature that are different from those of conventional PCM. (More below)

 

 

 

 

In evaluating the merits of a delivery / storage format, it is useful to look at its information carrying capacity versus the information rate of the signal that it is delivering - in this case, high quality audio. The information carrying requirement is most easily calculated based on the required dynamic range and frequency extension of the audio signal. In the early days of digital audio, when digital bandwidth was very expensive, many people tackled this question with a goal of picking a minimum information rate that would be considered high fidelity, and the CD was the commercial result. Good, but not really good enough for audiophiles, hence the interim fixes and now hi-res.

 

 

 

 

The required dynamic range, as determined by human physiology, to reproduce the full range of audible sound is generally agreed to be about 120 dB, or 20 bit linear PCM resolution. The dynamic range that most transducers can handle well is less, so a delivery format that can provide 120 dB of dynamic range is sufficient. Note that professional formats used for editing, EQing, and processing need to be of higher resolution to produce a good 20 bit result in the final release.

 

 

 

 

The required bandwidth, or frequency extension to reproduce the full range of audible sound, is not as well agreed upon as the dynamic range. For steady tones, the physiological limit is around 20 kHz. There is some controversial research indicating the physiology responds to higher frequencies directly, but a more important consideration is that more bandwidth is required to reproduce realistic sounding transients. (Long, complex discussion) Most researchers conclude that a 50-60 kHz bandwidth is enough.

 

 

 

 

From the above, it can be concluded that 24-bit 176.4 kHz linear PCM has room to spare in both dynamic range and bandwidth to deliver all perceptible audio content. DSD64x falls a bit short, although it is better than a 16-bit 44.1 kHz CD. DSD128x is capable of hitting the goal with some fancy shaping of the noise floor, which may have audible consequences. (More below)

 

 

 

 

Regarding our approach of converting DSD to 24-bit 176.4 kHz linear PCM: We are simply converting one delivery format to another that has greater useful information carrying capacity. It is easy to see that 176/24 has more than enough information bandwidth to carry all of the information in DSD64x. It is possible to put the entire bit stream of a DSD64x signal in 2/3 of the bits of a 176/24 signal, as is done in DoP. It might be argued that because raw DSD128x will not fit in the bits of a 176/24 signal, information is lost. However, when the efficiency of the formats for carrying useful audio information is considered, it becomes clear that 176/24 is still more than sufficient to carry all of the audio information in DSD128x. DSD does not make good use of the available information bandwidth. The conversion process from DSD to PCM can be done with very high precision using digital filtering that is stable and predictable, and especially if done off line, it can preserve all useful audio information so that nothing is lost.

 

 

 

 

Since the only valid way that an end user can evaluate a given delivery format is by listening to the end result in an entire system, and since the adequacy of the various delivery formats has been discussed, we will now consider A/D and D/A converters and other system level issues. Here again, there seems to be a great deal of confusion in the discussion posts.

 

 

 

 

One of the problems with trying to evaluate digital audio systems, especially DSD, is the huge number of possible variations in implementation. When recordings are specified as 24-bit 176.4 kHz PCM for instance, at least the character of the delivery channel is well known, and the variations in quality from one to another can be attributed to the quality of A/D conversion and the quality of the source itself. The channel is generally understood to have a spectrally flat noise floor that is well below the noise of the converter and the source. The same cannot be said for DSD.

 

 

 

 

With DSD, the delta-sigma modulator used to encode the audio in the 1-bit stream has a profound influence on the recording, and the variations in implementation are almost endless. The order of the DSM process is a major factor: the higher the order of the modulator, the greater the dynamic range that can be achieved in a given bandwidth, but at a price. The more one reduces the noise floor in the audio band, the faster the noise rises out of band. It’s like squeezing on a partially inflated balloon – you squeeze in one place and it pops out somewhere else. Also, the in-band noise floor generated by the modulator is normally not flat. Frequently, the noise floor in-band is shaped deliberately to put more dynamic range at frequencies where the ear is most sensitive. This has its own artifacts and sonic signature, most notably in level dependent shifts in instrument timbre. (Does anyone remember Sony’s Super Bit Mapping?)

 

 

 

 

Another factor that produces large variations in the system level reproduction of DSD is out-of-band noise. At least some of that noise must be filtered out – the question is how much, and D/A converter designers have a wide range of opinions on that subject. What complicates the decision is the fact that the analog electronics downstream have widely varying tolerances for high levels of high frequencies, and at some point they all become distressed. Typically, high levels of high frequencies cause some part of an amplifier, usually inside a feedback loop, to go into slew-rate limiting, which produces distortion. At onset, it may sound like a softening of the sound, which may be interpreted as euphonic, but it is also a loss of information and addition of distortion.

 

 

 

 

In ‘native DSD’ D/A converters, the filtering of the out-of-band noise must be done with an analog filter, and good quality analog filters are expensive and subject to drift with temperature, as well as sometimes requiring tuning during production. They also tend to introduce phase distortion near the filter’s corner frequency and hence also in the audio band. Because of this, ‘native DSD’ D/A converters often produce higher levels of high frequency out-of-band noise to keep the filter simple.

 

 

 

 

In contrast, conversion of the 1-bit DSD stream to multi-bit PCM in the digital domain can be done with digital filters, which are stable and, if well designed, free of most of the problems of analog versions. They can also be easily made selectable, even on a recording by recording basis. This is another argument for doing DSD to PCM conversion off-line. The noise floor of each recording can be reviewed before conversion using a spectrum display, often built into the converter, and then pick the optimal filter for the particular DSD delta-sigma modulator used to make the recording. It is only necessary to do this once. This is one of those things that can satisfy an audiophile who likes to tweak his or her system in a way that was common with analog sources but has largely gone away with digital.

 

 

 

 

We stand behind our statement that the vast majority of D/A converters currently on the market, including ours, use multi-bit delta-sigma converters. Originally, monolithic converters went from ladder structures at low oversampling ratios to single bit high oversampling ratios because better performance could be achieved at lower cost for mid-fi consumer CD players. It was not until the problem of element matching was solved that multi-bit high oversampling became practical. It has now taken over the high quality end of the market for both professional and high-end audiophile equipment because multi-bit delta-sigma converters produce very high performance without placing a large burden on the product designer.

 

 

 

 

One of the most important advantages of 5-6 bit delta-sigma converters is that the delta-sigma modulator can be low order and still meet dynamic range requirements. The result is that the noise floor rises very slowly as one goes up in frequency above the audio band, and therefore, the analog filter following the converter can be simple.

 

 

 

 

The above provides the answer to an intelligent question asked in the CA discussion; why go from single bit DSD to PCM to multi-bit delta-sigma conversion to analog. DSD has large amounts of high frequency noise, which can be easily filtered out digitally in the conversion to high bit precision PCM. The PCM can then be processed normally, including controlling level and up-sampling, and then a 5-6 bit lower order delta-sigma modulator drives the DAC with very slowly rising noise at the output and simple analog filtering.

 

 

 

 

Multi-bit delta-sigma audio very definitely is PCM, and represents coarsely quantized whole output sample values being sent to a linear PCM DAC. The fact that it is noise shaped does not negate the fact that it is linear PCM. The assertion that delta-sigma is the difference between adjacent samples of PCM is incorrect: that would be delta modulation, a precursor to delta-sigma modulation. Delta modulation for high quality audio was abandoned decades ago because it has a very serious limitation – it is slew limited by nature. The maximum value of the small difference word must be added repeatedly to the total to make a large level change. As a result, in order to reproduce high level, high frequency signals such as cymbal crashes, the delta modulation coder requires extremely high sample rates. For single bit delta modulation (the only version that was widely used in practice) achieving CD level performance requires multi-gigahertz clocking.

 

 

 

 

Further evidence that multi-bit delta-sigma data is PCM is the fact that recovering an audio signal after D/A conversion simply requires a low-pass filter with a flat frequency response in the pass band. If it were delta modulation, as has been claimed, an integrator would be required, which has a 6 dB/octave attenuation slope in the pass band.

 

 

 

 

It has been correctly stated that a single bit DAC has perfect linearity. However, eliminating amplitude linearity requirements has a side effect of increasing timing accuracy requirements. The high precision requirement has been moved from the amplitude domain to the time domain. 1-bit DAC’s are very jitter sensitive, and require much better clocking than multi-bit DAC’s. This is a perfect example of the design tradeoffs that designers face.

 

 

 

 

 

 

 

 

BTW, a misconception posted in the discussion needs to be corrected; Keith Johnson has no connection to Berkeley Audio Design and he was not involved in designing any of our products. He remains a close personal friend of ours, but that is the extent of his involvement.

[Superdad]

Chris:

Thanks for posting Berkeley's very clear dissertation on their viewpoint. Since it seems clear they are advocating for us CA folks to play our DSD tracks by converting (either offline or realtime) to 24/176.4KHz PCM, I wonder if they have researched the quality of the various s/w algorithms for doing so.

 

As a user and big fan of Audirvana Plus--running at 176.4 (352.8 after this weekend) into an NOS PCM1704--I already use its built-in DSD>PCM realtime conversion feature for the few DSD tracks I have. But I wonder if that conversion could be bettered.

(And for others who do this, I strongly advise against using the recent DSD conversion gain-compensation feature of A+--it ruins DSD tracks--just turn up the volume on your preamp!)

 

What are Windows users using and hearing for DSD>PCM software conversion?

 

I can't wait to read the responses to Berkeley (I am assuming it was written by Michael "Pflash" Pflaumer) from some of the other knowledgable folks here.

 

Happy Holidays!

 

Alex Crespi

UpTone Audio LLC

[Daudio]

they are advocating for us CA folks to play our DSD tracks by converting... to... PCM, I wonder if they have researched the quality of the various s/w algorithms for doing so.

 

Alex, Yes, thanks for that question !

 

I am getting tired of seeing software/algorithms treated as black boxes labled "Magic happens here". Berkeley Audio Designs response is another example: "The conversion process from DSD to PCM can be done with very high precision using digital filtering that is stable and predictable"

 

Time to look at them as you would our usual electronic machines, full of parts, and configurations, with different strengths and weakness's, subject to critique.

 

IMHO

[The Computer Audiophile]

Ah yes, software matters!

[The Computer Audiophile]

Hi Guys - I received one last response from Berkeley Audio Design. It addressed some items directly.

 

 

 

Berkeley Audio Design:

 

DSD versus PCM part 2.

 

 

 

Initial assertion:

 

Multibit DSD is not PCM.

 

PCM is typically 24 bit of the whole sample value while multi bit DSD is typically 5-6 bits of DIFFERENCE between adjacent samples. In 1 bit DSD this difference is binary (0 or 1) while in multi bit DSD it's the same difference between samples but quantized with 5-6 bits (Sabre has a 6 bit DAC). Multibit DSD is BETTER than 1 bit DSD, so there is nothing wrong with going 1 bit> 6bit which is what all these DAC chips do.

 

 

Our assertion in response:

 

 

Multi-bit delta-sigma audio very definitely is PCM, and represents coarsely quantized whole output sample values being sent to a linear PCM DAC.

 

 

We have discussed the fact that multi-bit delta-sigma modulation is not the difference between adjacent samples at length in our previous post.

 

 

Going to a discussion of specific DAC structures does not even address our assertion.

 

 

It is true that typical monolithic multi-bit delta-sigma DAC’s use what is known as a thermometer DAC structure in which a series of current-source elements of equal value are switched on or off and their currents are summed to produce the output current. It is also true that, because the current output of each element can never be exactly same as the others, the resulting errors are normally scrambled and noise shaped by various schemes of mapping the required number of on current sources to different elements for each successive sample.

 

 

The number of elements, typically current sources, is normally an odd number, but this is not inconsistent with binary numbers. There is an extra state in which all of the elements are off, giving an even number of output states. For example, a 4-bit DAC would have 15 elements whose output sum goes from 0 to 15, which is 16 states.

 

 

It is not even necessary that the number of elements be related to a power of two. The nature of the delta-sigma algorithm and the element scrambling algorithm can map a binary number to a larger number of elements than the largest number represented by the binary number.

 

 

The important point to take away from the above is that the DAC itself is linear. It is true that the input to the DAC elements is unary coded, but what is not even mentioned in all the discussion is the mapping of a binary 5-6 bit input to a 32 or 64 bit unary code that actually controls the DAC elements. It would be absurd to think that much processing is done on 64 bit unary coded data. That is just the end result.

 

 

The multi-bit delta-sigma data stream is a series of 5-6 bit binary values or in some cases 8 bit data, i.e. PCM with noise shaping. These binary values are then translated into scrambled unary code just before the output elements. That unary code determines which elements are turned on, and the sum of the output elements, typically in current form, is the output of the DAC. In most designs, that current is converted to a voltage and then low-pass filtered with a conventional filter to yield the audio.

 

 

The multi-bit delta-sigma data stream is generated by a delta-sigma modulator, which is typically used to re-quantize and noise shape an input with a different bit depth. Each multi-bit word is a binary number which represents an instantaneous value of the signal – in other words PCM.

 

 

At Pacific Microsonics, prior to starting Berkeley Audio Design, we were directly involved in the design of several multi-bit DSM DACS, two of which were commercially produced by Japanese companies. These DAC’s all used the principles described above.

[tailspn]

Hi Guys - I received one last response from Berkeley Audio Design. It addressed some items directly.

 

Nice explanation, but it glosses over the most single most important point; there are easily distinguishable sound quality losses converting from 1-bit or multi-bit (insert favorite acronym, I prefer PDM) DSD operating at a high sampling rate, to a much lower sampling rate 2's compliment binary linear PCM.

 

While Berkley is providing an argument as to why in their opinion it's superior to convert DSD to 24bit/176.4KHz linear PCM (2's compliment binary) in their DAC, I'm talking about the entire recording/playback process. To the degree that the sampling rate can be maintained at the original A/D converter front end rate throughout the complete process, be it 1-bit or multi-bit, the closer to the analog feed will be the resulting analog signal presented to the speakers.

 

Putting aside for the moment the debate over the definition of PCM, I believe Michael Jurewicz explanation in Ted's original post of this thread is accurate.

[Le Concombre Masqué]

Chris:

Thanks for posting Berkeley's very clear dissertation on their viewpoint. Since it seems clear they are advocating for us CA folks to play our DSD tracks by converting (either offline or realtime) to 24/176.4KHz PCM, I wonder if they have researched the quality of the various s/w algorithms for doing so.

 

As a user and big fan of Audirvana Plus--running at 176.4 (352.8 after this weekend) into an NOS PCM1704--I already use its built-in DSD>PCM realtime conversion feature for the few DSD tracks I have. But I wonder if that conversion could be bettered.

(And for others who do this, I strongly advise against using the recent DSD conversion gain-compensation feature of A+--it ruins DSD tracks--just turn up the volume on your preamp!)

 

What are Windows users using and hearing for DSD>PCM software conversion?

 

I can't wait to read the responses to Berkeley (I am assuming it was written by Michael "Pflash" Pflaumer) from some of the other knowledgable folks here.

 

Happy Holidays!

 

Alex Crespi

UpTone Audio LLC

you may look at my posts in [h=1]Thread: World’s First Valid Comparison of PCM versus DSD?[/h]

I wrote :

 

thank u so much mr Hanson !!!

 

 

your test led me to reinstate dsd in its full glory in my system

 

 

At first, pcm was a clear winner, dsd sounding dull, unfocused and boring in comparison. As a matter of fact, those days i noticed i preferred pcm hires downloads to dsd downloads/sacd rips though MFSL or Analogue Productions' signature appeals more to me than hdtracks’.

 

 

But then I took note of all the parameters I could play with (there certainly are zillions of others and other posters have listed some):

 

 

Src upsampling in A+ (the 30/.95/.99 flavor) in power of 2

 

 

Eq-ing in Alloy2 (baxandall bass +3.2 starting backward at 234 (peaks at +2.5 around 70 Hz), a dip bell of -3 Q 2.3 around 38.7, a large (Q 0.7) 2.5 dB dip around 859 HZ and a brickwall at 20K (Le Comcombre Masqué will remain masked but let say that my active Loudspeakers were top gear 20/25 years ago, i.e. the top of the line for monitoring at Radio France and other prestigious studios ; nevertheless they are claimed to be straight +-3dB between 20 and 20 KHz, but not beyond : why send hazards ?)

 

 

Last but not least : conversion to PCM (multistage, 0 gain of course). When I entered computer audio in May, I stated I preferred to convert and was displeased by DoP to my TEAC 501. Truly, that’s a good a way to get rid of harshness and get soft sound. Still works that way. Too much : straight PCM was actually closer to straight DSD than straight dsd to (over) processed dsd

 

 

But nowadays, thanks to Superdad, I’m running osx on a sd card, reading files from a ram disk, etc : my digital front is more analog than analog !

 

 

So I unchecked boxes….

 

 

Eq-ing (I had the room professionally measured and carefully, patiently designed the correction (might go Amarra/irc those days…) and upsampling are still good moves for my ears/system/room for PCM

 

 

But straight dsd beats even (at its best for me then) processed PCM

 

 

DSD is that good that now I prefer to be disturbed once in a while by harshness due to the 850 centered bump in my room. On the CSN track the echoes on the voices are almost lost with PCM or, at least, are far from being as conveying a sense of presence (my eq-ing does not mask them, on the contrary)

 

 

Thank you Mr Hanson

 

 

DSD is livier and more like live music, PCM is more for those who think how a good system should sound (actually i took note that some arrogant guys here argued in favor of PCM from their desktop chair not from a listening chair, just did not bother listening)

 

 

btw, OT i know , but it came along the way, to trained ears then : months of plugging/unplugging killed the 50 € Supra usb cable i got with the 501 ("that's the least you can do, you ought to blabla vendor speech) : the 2.90 € 1.8 m usb cable from supermarket x sounded awful and i feared i would have to order an audiophile one to be happy ; have to say that with the 5.50 € 1.5 m usb cable from hardware store y I don't miss the Supra...

 

and

 

I thanked Charles Hansen for his test led me to reinstate dsd in its full glory in my system, i'll back him there..

 

 

dsd bears an unbearable lightness of being that makes it different beyond level adjustment.

 

 

the test showed me how badly Audirvana behaves in converting dsd to PCM and leading SRC + EQing tasks.

 

 

At the end of the day that means i have to forget about my EQing for dsd files ; that's way beyond .1 dB and my mids and lows are in - and + 3 dB range difference.

 

 

I think that i.e. for listening to Pixies I might go the pcm route and EQ for I love my 36 to strike. But I dug that for acoustic stuff, i.e. the test exemples, there's a fluidity, a fastness, a liveliness that are unmatched by PCM. I think that people who chose ESL are more appealed by dsd for it's the same nature of difference (and it shows on my huge non ESL Loudspeakers)

 

 

 

 

(btw1 Korg with 0 gain, let it clip, + EQing in RX3 is better than A+ that kills and muffles;

 

 

btw2 RX3 performs a much better job at upsampling and EQing (same parameters of course) than A+

 

 

btw3 RX3 does not handle flac (nor up samples beyond 192) ; i always claimed i preferred wav over flac in my system/processes ; what happens in A+ when you go from flac 48 to flac 384 ?? here is an added possible cause, next to endianness)

[Superdad]

dsd bears an unbearable lightness of being that makes it different beyond level adjustment.

 

the test showed me how badly Audirvana behaves in converting dsd to PCM and leading SRC + EQing tasks.

 

Maybe you should try Miska's HQPlayer to convert all PCM to DSD…

 

Best to you, my ever enthusiastic comrade!

Alex C.

[Hiro]

I believe Michael Jurewicz explanation in Ted's original post of this thread is accurate.

 

Same here.

 

And BTW, does anybody on here argue that PWM amplifiers operating on a few bits are not PWM anymore?

[Hiro]

While Berkley is providing an argument as to why in their opinion it's superior to convert DSD to 24bit/176.4KHz linear PCM (2's compliment binary) in their DAC, I'm talking about the entire recording/playback process. To the degree that the sampling rate can be maintained at the original A/D converter front end rate throughout the complete process, be it 1-bit or multi-bit, the closer to the analog feed will be the resulting analog signal presented to the speakers.

 

And that's the whole point of the Direct Stream Digital format, or as some prefer to call it, Delta Sigma Direct - to capture the output of a delta sigma modulator without subsequent downsampling and decimation steps occurring in Delta Sigma "PCM" converters.

[Miska]

Simple definition without ambiguity, if there's a delta-sigma modulator, it is a delta-sigma DAC, but I get to that later.

 

The number of elements, typically current sources, is normally an odd number, but this is not inconsistent with binary numbers. There is an extra state in which all of the elements are off, giving an even number of output states. For example, a 4-bit DAC would have 15 elements whose output sum goes from 0 to 15, which is 16 states.

 

No, number of elements is typically even so producing odd number of levels. Reason is that you have zero-level in the middle and same number of levels above and below it. For example with typical "2.5-bit" has 5 levels; zero, two above zero and two below zero. In PCM you have one extra level at negative side, like CD has value range of from -32768 to +32767.

 

dCS has 24 elements = 25 levels and Sabre has 64 elements = 65 levels.

 

The important point to take away from the above is that the DAC itself is linear. It is true that the input to the DAC elements is unary coded, but what is not even mentioned in all the discussion is the mapping of a binary 5-6 bit input to a 32 or 64 bit unary code that actually controls the DAC elements. It would be absurd to think that much processing is done on 64 bit unary coded data. That is just the end result.

 

What ever is intermediate format doesn't matter, there are usually multiple different intermediate presentation formats. What matters in this scope is how the actual digital-to-analog conversion stage looks like - the conversion elements. Everything else is secondary.

 

The multi-bit delta-sigma data stream is a series of 5-6 bit binary values or in some cases 8 bit data, i.e. PCM with noise shaping. These binary values are then translated into scrambled unary code just before the output elements. That unary code determines which elements are turned on, and the sum of the output elements, typically in current form, is the output of the DAC. In most designs, that current is converted to a voltage and then low-pass filtered with a conventional filter to yield the audio.

 

Very typical configuration is "2.5-bit" binary coding that has four unary coded bits = five levels. And dCS has 4.585 binary coded bits which is 24 unary coded elements = 25 levels.

 

This has nothing to do with PCM DAC, if you use PCM bits to directly control these equally weighted bits (like you would do with R2R ladder PCM DAC) you get complete garbage out. Making a 24-bit PCM DAC this way would take 16777216 current sources...

 

Now give us DAC that can take 8-bit unary coded input and feeds it straight to output elements and we have a bit-perfect multi-bit delta-sigma DAC.

Leave out all the digital filtering, delta-sigma modulation and scrambling.

 

The multi-bit delta-sigma data stream is generated by a delta-sigma modulator, which is typically used to re-quantize and noise shape an input with a different bit depth. Each multi-bit word is a binary number which represents an instantaneous value of the signal – in other words PCM.

 

Wrong, having designed and implemented five different delta-sigma modulators that can output anything from 1-bit to 64-bit output I know this is wrong and doesn't have anything to do with PCM.

 

we were directly involved in the design of several multi-bit DSM DACS, two of which were commercially produced by Japanese companies

 

There you go, you said it yourself. It's not PCM DAC, it is DSM DAC and don't claim that it's PCM...

 

PCM DAC directly converts two's complement binary values to analog, like R2R ladder. Claiming anything else being PCM DAC is misleading. Especially claiming that scrambled array of 1-bit equal current sources would have anything to do with PCM. Such DAC can convert DSD directly without changing any of the source input bits, but it cannot convert PCM without changing the input bits. Period.

 

So as I originally said, if it needs delta-sigma modulator to operate, it is delta-sigma DAC and not PCM DAC. Period.

 

There is no need to jump the rate and bit-depth up and down to/from something like 24-bit 176.4k PCM. Let's keep it at what it originally is, end-to-end.

[elcorso]

Hi Guys - I received one last response from Berkeley Audio Design. It addressed some items directly.

 

 

 

Berkeley Audio Design:

 

DSD versus PCM part 2.

 

 

 

Initial assertion:

 

This looks to me more as a "Yellow Paper" that a White Paper.

 

Berkeley Audio Designs want to confuse readers even more?

 

As I said before DSD & PCM can survive together, but each one on their side.

 

If they don't want to built a DSD DAC, OK. But please, don't try to force anybody to conversions, isn't the best way for SQ...!

 

Roch

[Miska]

Berkeley Audio Design:

 

DSD versus PCM part 2.

 

 

 

Initial assertion:

 

Multibit DSD is not PCM.

 

PCM is typically 24 bit of the whole sample value while multi bit DSD is typically 5-6 bits of DIFFERENCE between adjacent samples. In 1 bit DSD this difference is binary (0 or 1) while in multi bit DSD it's the same difference between samples but quantized with 5-6 bits (Sabre has a 6 bit DAC). Multibit DSD is BETTER than 1 bit DSD, so there is nothing wrong with going 1 bit> 6bit which is what all these DAC chips do.

 

Small note on this.

 

Well, multi-bit SDM (DSD) is otherwise all the same than 1-bit SDM (DSD) except number of output bits. I can tell my modulators to output anything from 2-level (1-bit) to 65-level (64-bit) and it changes only one parameter in the process. All the math formulas stay the same, just one number changes.

 

If you would go directly from 1-bit to 6-bit without altering rate, that would be one thing. But converting sampling rate all the way down to 176.4k PCM just for it to be converted back up again in oversampling filter is pointless exercise.

 

Instead, I rather convert 2.8 MHz DSD directly to 5.6 MHz DSD!

 

But good that BAD now agrees about delta-sigma DACs. Now pretty much the only argument left is whether SDM and PCM are the same or not and some minor conversion stage details.

 

Sabre has 6-bit intermediate format (between modulator and element matching digital processing stages), but those bits don't drive the conversion elements. Since Sabre has 64 equally weighted conversion elements, it's a 65-level SDM DAC. To avoid confusion and due to inaccuracy of partial two's complement bits, we don't talk about bits in SDM scope, but number of levels instead. For example TI describes it correctly when they say that they have a "3rd-order 5-level sigma-delta" instead of "2.3219-bit sigma-delta".

[The Computer Audiophile]

 

But good that BAD now agrees about delta-sigma DACs. Now pretty much the only argument left is whether SDM and PCM are the same or not and some minor conversion stage details.

I'm only speaking for myself here. It seems like you've declared victory and Berkeley Audio Design has conceded based on your statement. I think this is very misleading on your part.

[The Computer Audiophile]

having designed and implemented five different delta-sigma modulators that can output anything from 1-bit to 64-bit output I know this is wrong and doesn't have anything to do with PCM.

Hi Miska - Here is where I struggle to believe your approach versus Berkeley's approach. Have you created an actual DAC chip that others have used, have you created an DAC (complete) that others have used, are you just simulating all this stuff, are there measurements available for anything you've done?

 

I ask these questions in 100% sincerity. It would help many readers, those still reading anyway, take you more seriously.

[Superdad]

Hi Miska - Here is where I struggle to believe your approach versus Berkeley's approach. Have you created an actual DAC chip that others have used, have you created an DAC (complete) that others have used, are you just simulating all this stuff, are there measurements available for anything you've done?

 

I ask these questions in 100% sincerity. It would help many readers, those still reading anyway, take you more seriously.

 

Chris:

I have confirmed with John Swenson that most of Miska's facts, theory, and arguments are correct. And although Miska's commercial DAC creations are mostly software (though he apparently has built numerous prototype hardware DACs and ADCs as well), I don't think that takes much away from he says.

 

As you well know, computers are a great place to perform digital audio transformations--they are far less resource constrained than any $10 DAC chip, especially all the modern ones that serve both PCM and DSD in the same chip. Software can be made to do PCM>DSD (HQPlayer), DSD>PCM (Audirvana and others)--and both at high rates in such a way as to minimize the inherent limitations of both formats (large amounts of just-out-of-band noise in the case of DSD64; poor chip-based digital filters in the case of feeding Redbook to standard DAC chips).

 

So although I am not terribly interested in any discrediting of the words of the talented and ernest folks at Berkeley, I do find the discourse around the topology and methodology of DSD, DSM, and PCM to be HIGHLY educational.

 

Happy Holidays all,

--Alex

[Miska]

I'm only speaking for myself here. It seems like you've declared victory and Berkeley Audio Design has conceded based on your statement. I think this is very misleading on your part.

 

I don't think this about anybody's victory, but more about reaching some kind of common understanding of things.

[Miska]

Hi Miska - Here is where I struggle to believe your approach versus Berkeley's approach. Have you created an actual DAC chip that others have used, have you created an DAC (complete) that others have used, are you just simulating all this stuff, are there measurements available for anything you've done?

 

You can use my delta-sigma modulators in 1-bit mode with any DSD-capable DAC using HQPlayer. For example the Schiit Loki and Teac UD-501 measurements I posted have been made using my delta-sigma modulators in HQPlayer. If I could find any reasonably priced DAC that supports 8-bit ASIO DSD I could offer public multi-bit DSD support with measurements too.

 

Plus you can use my PCM noise shapers with PCM DAC's out there. I have also posted some results for these. For example if you know your PCM DAC has 120 dB linear region, you can set number of output bits to 20 and use noise shaping.

 

I don't design chips and don't care about chips, I believe better results can be obtained without DAC chips (like Chord, dCS, Meitner and Playback Designs for example are doing). However I have designed and created discrete DACs, but I don't see point in posting measurements for hardware that nobody else has, and thus nobody could reproduce the results. The main difference in my case is that I don't need FPGA or DSP chip to implement all digital processing, because I do the things in player software and use "bit-perfect DAC" instead.

 

P.S. And yes, I've used CS4398 DAC chip in "Direct DSD mode" too for converting DSD to analog, where the DSD is sourced for example from RedBook PCM through software conversion in HQPlayer (or as direct upsampling from DSD64->DSD128).

[jzahr]

Hi all,

 

There is a final thing that both camps seems to have it differently: BAD seems to imply that in multi-bit delta-sigma the stream of “low bit” samples do, in fact, represent directly signal amplitude.

 

I want to ask now about this: is this true?

 

I admit that, to date, Miska's explanations of what is delta-sigma modulation seems more convincing to me (being not an expert in the field, I have no other possibility than to try to analyze the clarity and the logic of the explanations; others rely on reputation based on having designed and sold succesful products in the market, something that both parties have in their curriculae, some in hardware, some in software. Having said that, I am willing to change parties if better explanation are provided...)

 

My reasoning in finding BAD's explanation somewhat confusing is as follows [i have education in structural mechanics, not in electronics, so forgive me if I don't use rigorous terminoloy, but, obviously, feel free to correct me where applicable]:

 

We know it is always possible to convert, for example, 24/44.1 khz PCM to 24/176.4 khz PCM; this conversion process makes use of some form of a mathematical procedure called interpolation. Even if electronic engineers call this filtering and perhaps they sometimes adds dither and noise shaping, on its basis, there is some form of mathematical interpolation. Now, in the example above, which converts PCM to PCM, no one would call this “modulation”. It is interpolation.

 

Also, in this example, in both, the starting and resulting rates, the “meaning” of a quantized sample is that of a direct representation of the “amplitude” of the signal. This applies to all possible conversions between the common PCM formats (48 to 192, 96 to 44.1, etc.): these are up-sampling or down-sampling procedures based on interpolation plus, eventually, dither and noise shaping.

 

Now, in multi-bit delta-sigma DACs, sampling rates in the khz range of PCM signals are converted to the Mhz range. Regarding this, I cite now the B.A.D. explanation (bold added by me for context):

 

...

The multi-bit delta-sigma data stream is generated by a delta-sigma modulator, which is typically used to re-quantize and noise shape an input with a different bit depth. Each multi-bit word is a binary number which represents an instantaneous value of the signal – in other words PCM.

...

 

From the cited text, I understand that for Berkeley, the “meaning” of each quantized sample is the same in the starting and resulting signals in this khz to Mhz conversion – namely, a direct representation of signal amplitude – . Then I ask: why in this case the conversion is commonly refered to as “modulation” instead of simple interpolation?

 

What is then the meaning of the “modulation” term in this khz to Mhz conversion in multi-bit DACs if the meaning of each quantized sample is not modified in this conversion?

 

On the other hand, from an explanation recently given by J. Swenson to reader Teresa in the other thread (the “world's first valid comparison...” thread), I understand the contrary – namely, that the “meaning” of each quantized sample is, indeed, changed during the khz to Mhz conversion in multi-bit sigma-delta DACs, from the direct representation of signal amplitude to the representation of magnitude of an “error”, whatever that means.

 

I could agree with B.A.D. in that there is a certain similitude between PCM and multi-bit delta-sigma in that both consist of a stream of quantized samples - hi-bit count at khz sampling frequency in the case of PCM and low-bit count at Mhz frequency in the case of multi-bit delta-sigma; however, the "meaning" of each quantized sample is, I feel, not the same.

 

This all could be just semantics, but at this point, and given that assertion P cannot be true at the same time that assertion “not-P”, I ask now: can the involved camps recommend bibliography of standard textbooks on the subject, I mean the reference books used in courses in electronics?

 

Please note that I am not advocating here in favor of one format over the other (in fact, I believe this is pointless), I'm just trying to understand how things work.

 

Now, merry Christmas to all and have happy holidays

 

Jorge

[tailspn]

Hi all,

 

There is a final thing that both camps seems to have it differently: BAD seems to imply that in multi-bit delta-sigma the stream of “low bit” samples do, in fact, represent directly signal amplitude.

 

I want to ask now about this: is this true?

 

The short answer to your question is no, in comparison to the absolute amplitude represented in each sample of a 2's complement binary PCM data word. Each word of multi-bit (Pulse Density Modulation) represents a quantized digital value of the difference or change of value from the preceding sample. You can see two immediate advantages; the first in requiring fewer bits to have the same or higher resolution, making it much more efficient code, and secondly, it's now mathematically process-able in a computational engine verses 1-bit PDM (DSD).

[tailspn]

On the other hand, from an explanation recently given by J. Swenson to reader Teresa in the other thread (the “world's first valid comparison...” thread), I understand the contrary – namely, that the “meaning” of each quantized sample is, indeed, changed during the khz to Mhz conversion in multi-bit sigma-delta DACs, from the direct representation of signal amplitude to the representation of magnitude of an “error”, whatever that means.

 

Set aside for a moment the additional confusion of upsanpling and converting PCM to DSD, and just consider the action of a simple Delta-Sigma Modulator. It's a feedback system, gated by a clock. Like any feedback system, there's an error generated representing the difference between the input to the system, and the output, multiplied by the feedback factor. The system generates an error (signal) to null out the input/output discrepancy.

 

In the case of a simple first order Delta-Sigma Modulator:

 

The basics of sigma delta analog-to-digital converters | Embedded

 

that error signal is clamped between two levels (1 and 0) because it's the output of a comparator, and is the modulator's output.

 

The operation of the modulator includes a delay (the integrator), causing the system to "hunt" for equilibrium at the clocked rate (sample rate). Table 1 of the above link shows the 16 iterations of that modulator necessary for it to achieve null/equilibrium. Those 1's and 0's are the "error" mentioned in your quote, and more importantly, the modulated digital (1's and 0's) representation of the analog input.

 

As you can see, there's no absolute value represented as there is in a 2's compliment binary PCM audio sample. Just a pulse train who's density represents the instantaneous valve of the input.

 

For this system to be accurate, the clock/sample rate must be high enough such that at the highest frequency to be sampled, the 16 iterations occur (for this example DSM) without the input changing value.

 

A further investigation of the operation of these modulators leads me to the conclusion that DSD (Pulse Density Modulation) is not digital at all. It's analog! It's as analog as AM of FM carrier modulation of a signal, but in this instance, it's the analog signal modulating the density of pulses (clock). That's the primary reason DSD is described as sounding analog.

[firedog]

A further investigation of the operation of these modulators leads me to the conclusion that DSD (Pulse Density Modulation) is not digital at all. It's analog! It's as analog as AM of FM carrier modulation of a signal, but in this instance, it's the analog signal modulating the density of pulses (clock). That's the primary reason DSD is described as sounding analog.

 

Yes, which is why PDM can actually be played back directly over analog equipment, without being converted to "analog". You will hear music - not optimally reproduced music, but music nonetheless. Not true for PCM. Without A/D conversion it really is just a bunch of numbers.

[Miska]

As you can see, there's no absolute value represented as there is in a 2's compliment binary PCM audio sample. Just a pulse train who's density represents the instantaneous valve of the input.

 

And with multi-bit DSD you can have higher and lower pulses, but otherwise it's exactly the same.

[Sam Lord]

I hope this isn't too far off topic, but...in single-bit or multibit DSD, when is a ground reference made, and is noting of a zero crossing ever made? Obviously start and finish are at ground level, but the codes don't float along a whole track for many minutes without a ground reference, do they? Thanks.

[tailspn]

My guess is it's the same effect as coupling an AC signal through a capacitor. The receiving circuit establishes the new "ground" reference centered about its operating point. The is no ground reference in a PDM stream, unlike PCM, which is made of absolute magnitude samples.

 

One of the difficulties in editing a DSD file is choosing a point to make a slice. If you don't hunt around on the timeline for a "zero crossing", you'll get a click where the new segment begins, or ends. There's a DC offset specification for the DSD file for SACD manufacturing that limits the maximum DC offset to -50db. That's very often exceeded.