Prism Sound: DSD technology review

DSD is the format of the audio data used on SONY's Super Audio Compact Disc (SACD) format. DSD data was formerly only available on SACD discs, but more recently has been made available in computer readable files. The technology has maintained a devout following and hence its inclusion in the range of formats supported by CALLIA.

On the subject of formats: In addition to the 'standard' DSD bit rates that are multiples of 44100 Hz, we have implemented support for the slightly unusual DSD bit rates that are multiples of 48000 - namely 6.144Mbit/s (48000 equivalent to DSD64) and 12.288Mbit/s (48000 equivalent to DSD128). Note that as with the 44100Hz rates, only the lower rate is available on the S/PDIF interface (also as DSD over PCM).

As there are many myths and legends surrounding the DSD format we felt it worth providing some explanation for those that might be interested, both about the format itself and about our implementation of DSD support in CALLIA, designed to provide the absolute best, no compromise transcription to the analogue domain.

Prism Sound Chief Technology Officer Ian Dennis takes us through it!

A bit of History

DSD was conceived in an age when monolithic sigma-delta converters were relatively new. The quantisers in the converter chips of the day were one-bit and ran at 64x the eventual PCM rate ('64fs'); their digital decimation and interpolation filters were real-estate-limited and so had very non-ideal behavior in terms of flatness and stop-band attenuation - as well as being confined to PCM rates below 50kHz. The low sample rate meant that brick-wall filters were needed, with their attendant time-domain dispersion problems.

Although the idea of transacting and storing pulse-density-modulation (PDM) signals (like DSD) directly, without down-conversion to PCM ('decimation') was not new, it had not been used for audio until the mid 1990s. Experiments were made by connecting the 64fs one-bit product of the front end of a sigma-delta ADC directly to the one-bit back end of a DAC, bypassing the decimation and interpolation filter stages. The result, unsurprisingly, sounded better than with the filters in circuit as there were no interpolation stages. No processing, be it analog or digital, or even up-sampling, can make a signal more like the original - but the filters of the day were also very far from perfect.

So it was that DSD and the SACD were born.

By the time the SACD format came to fruition, higher quality monolithic PCM converters had become available with longer wordlengths and higher PCM sample rates. These formats were specified for DVD Audio (DVD-A), and a format war ensued which killed them both (or perhaps it was the lack of enthusiasm amongst punters for ANY new higher-cost format whose main selling point was better sound quality than the CD, which most people found quite acceptable). Although SACD and DVD-A never caught on, DSD and high-resolution PCM audio are now enjoying a renaissance amongst quality-conscious listeners in download and streaming form.

DSD or PCM? Which is better?

So which is better? Although Prism Sound is a manufacturer of no-compromise ADCs for the music recording industry, it is not necessarily for us to say: whilst we may have an opinion on the best way to convert analog audio to digital, and maybe even an opinion on the best format in which to store it, we must give our customers the means to produce whatever format they want: both! And as a manufacturer of no-compromise DACs we must strive to give them the best conversion to analog from either format - after all, there are some excellent recordings in each format! In the case of Callia this involves converting incoming DSD streams to something with a longer wordlength and a lower sample rate prior to conversion. But hang on - surely that sacrifices the wide bandwidth which is the whole point in DSD? No, wait, hear me out:

Since the days when DSD was conceived, converter technology has moved on - dynamic performance and linearity have greatly improved, largely owing to the use of multi-bit front and back end circuits running much faster than 64fs - the limitations of one-bit quantisers at 64fs were surpassed not long after the initial DSD experiments.

This presents us with an awkward choice in converting a DSD stream to analog: do we present the DSD directly to a lower-quality 64fs one-bit back end, or do we decimate the signal to produce a longer wordlength which we can then re-process for a faster state-of-the-art multi-bit back end? We might prefer the latter.

Is there a difference in performance specifications?

All this is borne out by studying the datasheets of today's highest-performing audio DAC chips: they nearly all support both DSD and PCM inputs, so surely we will see that the dynamic performance and linearity is worse for DSD than for PCM? Actually, no! Most of them seem to have exactly the same performance for both! So, problem solved: we can use them in DSD mode without losing all those DSD advantages like super-wide bandwidth and inherently linear one-bit quantisation. Sadly though, further study reveals that these devices are actually decimating the DSD to a longer wordlength and lower sample rate internally - that's why the performance is identical, since the decimation process can be done with minuscule degradation compared to that of the D/A conversion process itself. There, you see - I knew that the digital volume control working in DSD mode was a dead give-away!

Few DACs actually go direct to analog from DSD ..

So I'm afraid that there are few DACs which really do convert DSD to analog without PCM in between, although most who adopt them will not know it. Like many things lurking in the audio woodshed, this reality is anathema to the purist. But we should remember that nearly all DSD releases have been processed somewhat in the recording and mastering studio, even if it's just a tad of levelling or a smidgeon of EQ - and that NONE of this can be done without going via PCM or (worse) back to analog, so maybe we shouldn't be prissy about the DAC chip. Having said this, there are a few rigorously-produced minimalist DSD recordings which haven't been processed at all; I wonder if their ADC had a one-bit 64fs front end, and if not... No, no, I mustn't go there.

Direct DSD to analogue is often worse ..

Actually, there are a few DAC chips out there which DO allow the DSD stream to be presented directly to the back end DAC, without prior conversion to some form of PCM - but their datasheets tend not to quote the direct-mode performance, only the via-PCM figures - which are usually (you guessed it) the same as for PCM. The rest tend either to admit to worse performance in direct DSD mode, or else their performance may not be such that we care much either way.

DSD via PCM

So, having decided that we may prefer to sacrifice some bandwidth for linearity by going via PCM, what will be the price? Even a basic DSD64 signal, which is sampled at 2.8224MHz, could represent frequencies up to nearly the Nyquist frequency of 1.4112MHz, whereas the current state-of-the-art PCM signals are only sampled at 192kHz (with a bandwidth up to only about 90kHz), or perhaps 384kHz (with a bandwidth up to maybe 180kHz). We'd better hope that what we're sending to the back end DAC isn't sampled anything like that slowly!

How might the presence of high frequencies affect my music?

In reality, though, even my asthmatic puppy can't hear much above a few tens of kHz so perhaps the higher PCM rates might not sound too awful? Also, it turns out that we really don't want super-high frequencies entering our other items of analog audio equipment - which are so rarely capable of either rejecting them thoroughly or transmitting them intact, so the consequence is often that they are 'demodulated into the baseband', which is audio-speak for making odd noises which weren't there in the first place. This problem was addressed in later versions of the SACD specification (the so-called Scarlet Book), which added the requirement for DSD DACs to incorporate a 50kHz analog low-pass post-filter in order to protect subsequent equipment. So maybe we don't need to worry about limiting ourselves to higher PCM rates.

In summary ..

So, to sum up, it seems that since those days, reduction in silicon geometry and resulting increases in speed and complexity have meant that DACs' interpolation filters have become far closer to the ideal, and can work at much higher sample rates, meaning that brick-wall responses are no longer necessary; whilst at the same time, faster multi-bit front and back ends have pushed back the limits of linearity and dynamic range - all of which are moving PCM quality ever onwards and upwards. For DSD, those same digital filters can sneakily decimate to PCM on the way to the back end DAC, whilst providing the stipulated 50kHz low-pass filter. Job done.

Actually, not quite. We think that it's nice to do the conversion process ourselves so that we aren't limited to the arbitrary response provided by a particular DAC device (which is often inscrutable for obvious reasons); thus we can achieve exactly the response we (Prism Sound) want. It also means that we can devote a suitable silicon budget to doing a really nice job of it. I might just add that we haven't adopted this approach lightly: Prism Sound and SADiE have been researching such processing and conversion technologies since the earliest days of DSD. Our Prism Sound ADA-8XR A/D D/A converter, which has long been the DSD conversion tool of choice in many of the world's top SACD mastering studios, uses exactly the same architecture as Callia.

The ADA-8XR won an important DSD converter shootout as the best and most transparent, in a recording studio in Switzerland at the height of the SONY/Philips push for SACD.

OK, now you can shoot me. But I hope you will go easy on me - I managed to trudge through all this geekery without ever once becoming judgmental. On the other hand, you could Google Prof. Stan Lipshitz, and find out why he thinks that "1-bit Sigma-Delta Conversion is Unsuitable for High Quality Applications". Or you could Google "DSD PCM myths" - and be sure to stop by Eelco Grimm's excellent analysis. I've just been scanning the "Direct Stream Digital" page on Wikipedia, which seems to address some of the same issues I've been discussing above, and is a good starting point for researching this interminable debate.