Which DACs bypass digital filtering?

[mansr]

44 minutes ago, jabbr said:

This is my understanding:

(I'm going to limit the discussion to DSD in this post and can discuss PCM separately if desired -- just ask)

 

The "sound" is contained in the digital recording. The goal of the reproduction system (DAC + Amp + Speakers) is to accurately product the "sound".  During the playback process, the "sound" is mixed with "noise". In a DSD (SDM) bitstream, the "sound" is directly contained in the "analogue" part of the bitstream, the "noise" is contained in the "digital" clock that is used to transport the stream from one place to another. The function of the DAC is to separate the analogue sound from the digital noise.

 

This is really really simple, so if you don't understand what I've written above, go back and reread, because understanding this is essential to understanding the process. The last sentence, in particular, accurately and specifically describes the function of the DAC.

 

In DSD/SDM the digital noise is contained in the carrier clock (BCLK) as well as its harmonics. The BCLK is necessary to interface the analogue signal with the digital system and the goal of the DAC is to remove all vestiges of the BCLK from the analogue signal without disturbing the signal itself. This where upsampling and filters come into play.

 

Let's say we allow everything to pass including the carrier BCLK -- we can't hear it right? Speakers can't reproduce it right? What's the big deal? That's where intermodulation distortion comes in: high frequency noise interacts with the electronics to produce measurable, audible and very harsh sounding distortion in the audible band.

 

One might consider a "brickwall" filter which would allow the analogue signal to pass and cut off everything above what we define as either 44 kHz or 96 kHz or whatever we define as the upper limit of the analogue signal we want.

 

Well it turns out that these "brickwall" filters also have distortion that extends below the cutoff frequency: the brickwall filters aren't perfect. So a much much better idea is to use a gentle filter at the corner frequency but in order to get the gentle filter to effectively filter out the digital noise we need to "noise shape" which is where the upsampling comes into place: the upsampling increases the frequency of the digital carrier clock (BCLK) thus increasing the frequency separation between the analogue signal and the digital noise and thus improving the ability of the gentle filter to remove the noise. Viola'

 

Now 99% of PCM starts out as SDM/DSD and ends up as SDM/DSD to the same argument applies with the added complexity of where, when and how to convert between SDM and PCM.

Sorry, but that makes no sense whatsoever. I suggest you study the maths involved properly before trying to explain things.

[mansr]

18 minutes ago, jabbr said:

Typically if a DAC accepts DSD256,512 its going to be very difficult to realtime filter that signal! so I'd assume there's no input filtering, but if there is, someones done a great engineering job  

You mean like ESS and AKM DACs do?

[mansr]

16 hours ago, Jud said:

18 hours ago, jabbr said:

I'm pretty sure that with the correct settings, the iFi iDSD Micro bypasses input filtering.

It applies its modulator, I think (whichever of the three available you select).  But if you've already used a software modulator, I don't know how much is left to do.

The DAC chip (TI DSD1793) upsamples internally to 384 kHz. If the input sample rate is lower, a switch offers three choices: sharp or slow rolloff internal filter (both linear phase, the "minimum phase" label is a lie), or bypass ("bit-perfect") where the input samples are repeated (zero-order hold) by the microcontroller to achieve the DAC's native rate. With DSD input, the filter switch selects which built-in analogue filter to use. There is no digital processing of DSD data.

[mansr]

6 minutes ago, jabbr said:

This is intended to be a simple natural English language explanation for people who speak English not a mathematical explanation for people who speak math.

It's still wrong in more ways than I can count.

[mansr]

52 minutes ago, jabbr said:

This could be difficult because you don't speak natural English as a primary language

My English is as good as anybody's.

52 minutes ago, jabbr said:

and I understand that you are having trouble counting the ways it is wrong so let's start with a simple sentence. This is an English language sentence. Do you understand? Do you agree? Do you need it spelled out?

I understand the words perfectly. They simply do not have a meaning which can be considered correct in any scientific or engineering sense. If I were to explain a DAC in simple terms, I'd go with something like this: The function of a DAC is to convert a digital representation of sound to an analogue form.

[mansr]

1 hour ago, Jud said:

Are there DACs (the piece of equipment) incorporating these chips that accept DSD256/512 rates as input?

 

 I believe I remember ESS chips producing output at 40+ MHz rates (wasn't aware of AKM doing this), but I'd never heard of DACs with those chips accepting rates that high.  I thought ESS in particular would only allow 384KHz input max.

The AK4490EQ (among others) accepts up to 768 kHz PCM and DSD256. Its internal rate isn't as high as that of ESS, but it does process DSD inputs unless put in bypass mode. This is the chip used in the TEAC UD-503.

[mansr]

1 hour ago, Jud said:

There are rare DACs that do.  I've listened to one.

They are rare for a reason. Put bluntly, they suck. Worse, they can blow up your tweeters if used with a high bandwidth amp.

[mansr]

15 minutes ago, jabbr said:

In the simplest engineering terms I can think of, and assuming a single DSD channel switching between 0 and 5v:

 

The simplest (DSD) DAC is nothing more than a low pass filter.

That I can agree with.

15 minutes ago, jabbr said:

the corollary being:

 

The function of the low pass filter is to remove the high frequency digital noise from the analogue signal.

This, however, makes no sense. As long as the DSD stream is seen as a sequence of bits, both signal and noise are digital. When seen as a varying voltage, signal and noise are both analogue. Since the bulk of the noise has been separated in frequency from the signal of interest, it is possible for a low-pass filter to remove the former while retaining the latter.

 

[mansr]

9 hours ago, Jud said:

If oversampling is applied, filtering is necessary to avoid aliasing and consequent harmonic and intermodulation distortion.

Filtering is always necessary. The point of oversampling is enabling the most critical filtering to be performed digitally where it can be far more accurate than in the analogue domain.

[mansr]

15 minutes ago, Jud said:

Looking at the UD-503 owner's manual (page 18), it appears that DSD input is handled as the micro-iDSD does.  There are two analog filter options for DSD input.

Yes, but does it put the chip in DSD bypass mode?

[mansr]

14 minutes ago, jabbr said:

Ah, ok. I am using the (admittedly nontechnical) term "separate" to mean "remove" as in "remove the noise while retaining the signal". 

That choice of vocabulary was not what I objected to. Your first offence was characterising the noise as wholly digital and the signal as wholly analogue. From there it only got worse.

[mansr]

4 minutes ago, Jud said:

Can't filtering at high sample rates be analog, and digital at Redbook rates?

 

I thought the idea of oversampling was to allow more headroom to avoid the necessity of a "brickwall" filter.

A sampled signal is an a representation of the original up to half the sample rate (if the original was not band-limited, the representation is inaccurate). To properly reconstruct the continuous-time signal, the images of the base band must be filtered out.

 

With a Redbook input, the images start at 22.05 kHz. In order to preserve frequencies below 20 kHz, a rather sharp (brickwall) filter is required. This is difficult to realise as an analogue circuit. Oversampling still requires the same brickwall filter, but now it can be done digitally which is much easier. A 2x oversampled signal with digital filtering has content  only up half its Nyquist frequency. With our Redbook input, the analogue reconstruction images now start at 44.1 kHz. Moreover, since only half the digital bandwidth is used, no actual images are present for another 22.05 kHz, allowing subsequent filters a full 44.1 kHz transition band. This permits a lower order, more easily implemented analogue filter. Still higher digital oversampling extends this shift from analogue to digital filtering.

 

Whenever the source is Redbook, proper reconstruction requires a brickwall filter somewhere in the chain. The choice we're given is between analogue and digital, and if accuracy is the goal, digital always wins.

[mansr]

1 hour ago, Jud said:

I don't know.  However, there isn't anything in the manual to suggest DSD input rates are handled by something other than the analog filter options.

That means only that those are user-settable options. To find out what's really going on, one would need to inspect the commands sent to the DAC chip with a logic analyser. I've done this with some iFi devices, which is how I know what they're doing.

[mansr]

20 minutes ago, Jud said:

Must an interpolation ("upsampling") filter be digital?

Yes. Although strictly speaking, sampling is a time-domain quantisation where the sample values are arbitrary, all practical storage systems use a digital representation of sample values.

20 minutes ago, Jud said:

Can a final reconstruction filter be digital?

No. A/D conversion by definition includes an analogue stage. It can be simple or complex, but there is always something.

[mansr]

24 minutes ago, Jud said:

So the OP is looking for DACs that permit a PCM or DSD bitstream under one or more circumstances to be sent from input to the final analog reconstruction filter without being acted on by a digital interpolation filter.

I believe so. Why this aspect should be important is something I do not understand. What matters is the accuracy if the output. How it is achieved should be irrelevant.

[mansr]

2 minutes ago, Jud said:

I don't know the OP's reason(s).  To me, the goal of accuracy and a DAC that works in the way the OP is asking about can be related.  If we hypothesize that we have interpolation filtering and/or SDM in software producing more accurate results than can be obtained in the DAC's internal processing, then bypassing the internal processing would make sense.

What matters is the accuracy of the software plus hardware available. If the best performing solution involves some hardware processing, that should not be seen as a weakness.

[mansr]

6 hours ago, semente said:

Could you elaborate a bit on the hardware processing for PCM and DSD? 

Perhaps it's best to study a few examples.

 

Below is a block diagram of the TI DSD1793 chip used in iFi DACs. The PCM1795 in the TEAC UD-501 is similar.

 

 

This shows a standard PCM path with 8x upsampling followed by sigma-delta modulation. A filter bypass mode allows direct input to the modulator at 384 kHz. The datasheet reveals that this is in fact a hybrid design where the modulator, a 3rd order 5-level design, operates on the low 18 bits only. The output is combined with the high 6 bits to form a 66-level code which forms the input to the actual D/A conversion stage.

 

The DSD path is separate and does not involve any digital processing. The D/A stage supposedly consists of a shift register style FIR filter similar to Miska's design but with different weights on each position. The datasheet is vague but supports this idea. Four different filter choices are available.

 

Now look at AKM's top range. Several variants with similar design are available. They show up in devices from TEAC (UD-503), Linn, ESOTERIC, Marantz, and others. This diagram is from the datasheet of the AK4497:

Again, a fairly typical PCM path. The main difference compared to the DSD1793 is the addition of a digital attenuator (DATT). A filter bypass mode is available. I can't find any information on the sigma-delta modulator, but I would assume it is a multi-level design. The SCF (switched capacitor filter) blocks convert the modulator output to analogue.

 

DSD handling is quite different from the TI chip. In the default mode of operation, DSD input is low-pass filtered before going through the same digital volume control and sigma-delta modulator as PCM. As far as I can tell, the modulator is operated at the same rate as the DSD input, i.e. no resampling is performed. A bypass mode allows sending DSD data directly to the SCF without going through the modulator. Oddly, some versions of this block diagram show the digital low-pass filter always being active while others (the figure above) indicate that this too is skipped in the bypass mode. I don't know which is correct.

 

Finally, the ESS Sabre series found in various products ranging from the Audioquest Dragonfly to the Benchmark DAC3.

This is different from what most manufacturers do. PCM input is upsampled using a two-stage FIR filter to a maximum of 1.536 MHz. These filters are fully programmable and can also be bypassed entirely. The FIR filter is followed by volume control and an IIR filter. After this comes "THD Compensation" which I have no idea what it does. Next the data goes through an ASRC which upsamples further to a rate supposedly in the vicinity of 40 MHz (even a datasheet I'm not supposed to have doesn't say). Finally, there are the usual sigma-delta (which I assume is what hides behind the Hyperstream label) and D/A stages.

 

Like in the AKM chip, DSD input is digitally low-pass filtered and subjected to the same processing as PCM. Apparently unique to ESS is that even DSD is upsampled further, and there is no option to disable this.

 

From these examples we can see that each manufacturer has chosen a different approach. The TI design favours simplicity while ESS relies on heavy processing. AKM falls somewhere in the middle. All achieve excellent performance figures. In a final product, the surrounding electronics, notably clocking and analogue output drivers, matter far more than the DAC chip itself.

[mansr]

22 minutes ago, jabbr said:

PCM is conceptually different from SDM (DSD) in that the analogue signal itself has been converted to a digital signal (numbers). DSD stream does not directly contain numbers -- the analogue signal remains present being at the low frequencies while the digital carrier is present at the higher frequencies.

This is a common misconception. I really ought to do a proper write-up explaining how these things actually work.

22 minutes ago, jabbr said:

In PCM, the aliasing noise has its largest component at the sampling frequency and this is what needs to be filtered out. Thus the brickwall filter. The "problem" with analogue filters is that they typically have effects not only on the frequencies above the corner frequency but also at frequencies lower than the corner and so with a corner frequency of 44.1 kHz, the output filter clearly has effects on the analogue signal itself. Not good. By upsampling PCM, this is the same as "noise shaping" DSD in that the aliasing frequency is pushed higher and similarly the analogue output filter may then have less effect on the desired signal.

Aliasing is not noise and has nothing to do with noise shaping. Also, your filter frequencies are off by a factor 2.

[mansr]

3 minutes ago, jabbr said:

Not a misconception at all. Are you able to defend yourself in English?

Which part of my English do you have trouble understanding? Do I need to dumb down my vocabulary for you?

[mansr]

9 minutes ago, jabbr said:

Try not to argue by authority -- I don't consider that you've copied/reimplemented what Miska did years ago anything that grants you authority in my book.

 

"Aliasing is not noise" -- I never stated that. I have defined the term "noise" explicitly. Never said that aliasing has anything to do with noise-shaping.

 

Perhaps you are having a problem reading & understanding? 

Got any more insults while you're at it?

[mansr]

1 hour ago, mansr said:

Below is a block diagram of the TI DSD1793 chip used in iFi DACs. The PCM1795 in the TEAC UD-501 is similar.

Seems like the image vanished. I'm sure it was there earlier since it's in my browser cache. Here it is again:

Perhaps @The Computer Audiophile can put it back where it belongs.

[mansr]

15 minutes ago, semente said:

What are the advantages of further upsampling DSD?

Digital filters are always more flexible than analogue ones. Especially for DSD64 where the noise is quite close to the audio band, a digital low-pass filter followed by remodulation at a higher rate can give an end result superior to using an analogue filter directly.

15 minutes ago, semente said:

And potential disadvantages?

It's always possible for a poor implementation to do more harm than good. If the implementation is of reasonable quality, I see no downsides.

15 minutes ago, semente said:

Is it better to do it with hardware or in a computer by software?

There is no difference in principle between hardware and software digital filters. A hardware filter is simply a network of multipliers and adders hardwired to perform one function. A computer uses exactly the same kind of arithmetic blocks to do its calculations, only here they are controlled by a program so the function is not fixed. That said, cheap DAC chips might use lower precision arithmetic in order to save silicon space (and power). Good DACs have 32-bit filters whereas computers can easily use 64-bit precision, although the benefit of that is marginal at best. The main advantage of software is the flexibility it offers.

[mansr]

12 minutes ago, Jud said:

Not accurate.  Various people have done sigma delta modulators at various times.  I believe I know the source of at least some of the basic ideas mansr used, and it wasn't Miska.

How could it be when his source code is secret? Besides, Miska didn't invent SDM.

[mansr]

30 minutes ago, Jud said:

Yes, though rather than source code, I was thinking of academic articles providing some ideas you could then implement in your code.

I don't recall seeing any academic articles by Miska.

[mansr]

6 minutes ago, jabbr said:

Right, so they way I personally apportion credit would be to the first to publish as well as the first to popularize. Some certainly for an open source implementation because people can learn from this. I've never claimed credit myself for reimplementation of algorithms -- of course I come from the days where code was assumed to be open source -- I'd certainly give credit for a visibly elegant implementation.

Where have I demanded credit for anything? You were the one who brought up my software along with vague accusations.