GoldenSound's Official Headphones.com Community Forum Thread

[NickZ]

@donjoe0

Because if Nyquist is right, whatever information is captured in the original samples at the original rate is sufficient for perfect reconstruction

This is common misunderstanding.

Here are a couple of quotes from a paper that describes Nyquest and sampling (Tim Wescott paper about Nyquest).

It is a common misconception that the Nyquist-Shannon sampling theorem could be used to provide a simple, straight forward way to determine the correct minimum sample rate
for a system…

The difficulty with the Nyquist-Shannon sampling theorem is that it is based on the notion that the signal to be sampled must be perfectly band limited. This property of the theorem
is unfortunate because no real world signal is truly and perfectly band limited…

What this means is that no system that samples data from the real world can do so
perfectly—unless you’re willing to wait an infinite amount of time for your results. If
no system can sample data perfectly, however, why do we bother with sampled time sys-
tems? The answer, of course, is that while you can never be perfect, with a bit of work you
can design sampled time systems that are good enough…

This means there is no such thing as a “properly implemented dac” that people use to “prove” that all dacs sound the same.

The only thing a real-world dac can do is approximate Nyquest, and there is no single correct way of doing that.

The more effort that is put into the approximation, the closer the output is to the original signal.

[pennstac]

I think @Lou_Ford has been a proponent of SWARM/DEBRA style subs in the past.

Your thoughts on that modality?

[SSN757]

Interesting. To me, it’s hard to discern DAC differences with playback, in large part because of the headphone amp and the actual headphone being used to evaluate. Some headphones are more sensitive to hardware than others. When I switched out my DAC/Amp combination to the Topping D900/A900 combination, it seemed like the sound from all the headphones in the collection underwent a sonic upgrade, some more so than others. How much of that was the DAC vs. the Headphone amp, can’t say. The ability to upscale the PCM file from a max of 192 Khz using a TI 1792A DAC to 705/768 Khz DSD DAC also changed the playback sound.

I agree that blind testing is a good tool for engineers to help develop their wares. It just seems that there are limitations with blind testing that need to be taken into account.

[GoldenSound]

@donjoe0

I mean normally it does, when it’s software oversampling and the low-pass is handled by the DAC, but I guess you’re saying PGGB itself doesn’t, as a matter of design choice.

It’s not a PGGB design choice, it’s just how Nyquist reconstruction works. You can only reconstruct up to half the sample rate of the original source material. For 44.1khz, that’s 22.05khz.
Upsampling to a higher rate is part of this reconstruction, but upsampling to 768khz doesn’t move your nyquist frequency to 384khz. When you look at the output of any oversampling DAC or tool, you’ll see the cutoff at half the original sampling rate, regardless of what you’re upsampling to. Information above that cannot be constructed as it was never in the original file to begin with.

@donjoe0

Because if Nyquist is right, whatever information is captured in the original samples at the original rate is sufficient for perfect reconstruction

It is sufficient, but what people miss out is that Nyquist theory does NOT state that it is perfect and that’s just it. It states it’s sufficient to perfectly reconstruct IF you perfectly band limit.
The problem is ‘perfect’ band limiting (leaving everything under 22.05khz untouched and fully attenuating everything above that) requires infinite compute power, which we don’t have.

As a result, DACs do what they can with the compute power available. Most DAC chips use filters with 128-1024 filter coefficients, which is mostly ok but doesn’t give you enough room to actually fully attenuate by 22.05khz, hence why any AKM/ESS DAC running stock filters either rolls off early and attenuates treble in the audible band or doesn’t properly attenuate by 22.05khz.

Whereas products that dedicate more compute power to being able to run higher performance filters can do that properly.
And then if you go for stuff like PGGB you can get the mathematically best possible reconstruction given the number of samples in the file, including over 800dB of stopband attenuation if desired. (Cause hey why not)

[SSN757]

Excellent explanation. I do hear a difference when I use PGGB-RT for PCM files. It does seem to be more noticeable with headphones.

So, for DSD DAC’s (Topping D900), what’s going on with the filtering there compared to a ESS/AKM DAC?

[donjoe0]

@GoldenSound

It is sufficient, but what people miss out is that Nyquist theory does NOT state that it is perfect and that’s just it. It states it’s sufficient to perfectly reconstruct IF you perfectly band limit.

Ohh OK, wow, that’s news to me, indeed I usually don’t see this underscored when sampling/reconstruction and the Nyquist frequency are discussed. You usually find the theorem stated as “the signal can be perfectly reconstructed if it’s sampled at a rate higher than 2x the maximum frequency it contains” - the notion of band-limitation is there, it’s just not called that, in those words. It’s band-limited because it has a maximum frequency (B). So every spectral point above B is at exactly 0 amplitude, and then we can perfectly reconstruct the signal from a periodic sampling taken at a rate higher than 2B. Damn.

And even this instance of “can reconstruct” is only the purely mathematical/theoretical one, because with the way time to frequency conversion works, longer recordings tending toward infinity are needed the closer to Fs/2 you’re trying to get with the reconstruction. Also didn’t realize this until I dug deeper today: Fs/2 itself is already not recoverable, and the closer you get to it the more reconstruction errors you get regardless of ADC or DAC or PGGB quality! It’s just because a musical note containing our coveted… I don’t know, 22049.9 Hz component would need to last longer than the artist’s full discography put together (or somesuch impractical value ) in order to be perfectly reconstructed from a 44.1k digital version.

The theory allows it, it’s just that we will never find such a thing in realistic recordings (and would we even like to listen to them if we did? lol), and there’s nothing any device or piece of software can do about it… is my understanding thus far. This would explain why they included that odd-seeming (when you first find out about it) extra 2 kHz on top of the 20 kHz supported by Red Book with the 44.1k sampling: to keep the Fs comfortably far away from 2B, so that all of B can be reconstructed without audible errors (for human listeners who do actually hear absolutely nothing past 20k, which back then I guess they assumed was everyone).

[kandamrgam]

Hey GoldenSound.. Thank you for having this platform. This is excellent. I have a question regarding great measuring R2R DACs like Holo Cyan 2, Holo May etc. I was wondering if an R2R DAC measures that well like a typical DS DAC from Topping/SMSL, then is there a difference sonically between R2R and DS DAC at all? If they both measure equally well, shouldn’t they sound the same too?

Side note question, I was wondering if you ever felt the R2R magic (i.e. a difference when compared to DS DACs) ?

[donjoe0]

@donjoe0

the closer you get to it the more reconstruction errors you get regardless of ADC or DAC or PGGB quality

No, wait, the infinite-signal requirement is just the reciprocal of the strictly-band-limited signal requirement, i.e. the infinitely steep filter requirement. High-taps filtering like what PGGB does is exactly what addresses the problem and improves reconstruction closer to 22050 than is possible with more simplistic filters, based on the exact same music samples - it effectively reduces the number of samples needed to accurately reconstruct the amplitudes of stuff like 22049 Hz to practically available amounts. Then the only degradation we can’t fix is from whatever errors the ADC baked in there.

Took me a bunch of extra steps but I think I got there.

[kandamrgam]

Thank you for this, eye opening! I do not know how to figure out if a headphone class-D amp is safe or not.

I always wanted to buy E1DA PowerDAC v2.1. Just to try out the class-D / full digital approach. But will it be safe or not, I dont know. Will it be even possible to figure out from the specs, I dont know either. This is what I could find from the official site (https://e1dashz.wixsite.com/index/pdv2),

Output power: > 320 mW @ 32 Ohm @ 1 kHz @ THD = 1%

Output power: > 580 mW @ 16 Ohm @ 1 kHz @ THD = 1%

Will this be enough to conclude 32 ohm and 16 ohm headphones are fine to use? I messaged Ivan directly, but no response.

[SSN757]

Over on another website, there is a debate raging about upscaling CD files vs. standard CD file playback. My experience with PGGB-RT has demonstrated that the overall sound signature seems to sound slightly better overall, whilst the naysayers claim that upscaling does not provide any improvement in sound.

What say you?

[GoldenSound]

I’ve not tested the powerDAC myself and am not super familiar with its design, but Ivan knows what he’s doing. I’d be shocked if it didn’t perform properly

[GoldenSound]

In the ‘Anatomy of a DAC’ video today I mentioned that there was one thing I was going to touch on that needed to be cut for time.

The video is here:

What does everything inside a DAC actually do? - DAC internals demystified

But what I wanted to discuss was:

FPGA DACs don’t actually exist.

What do I mean by this? Well, a lot of people commonly refer to various products as ‘FPGA DACs’ but this description is actually not particularly ideal.

An FPGA is a programmable logic device, and they’re found in all sorts of products. Sometimes more basic FPGAs and CLPDs are found even in quite cheap DACs that are still using off the shelf DAC chips, doing things like driving a display, handling control of the DAC or changing settings like DSP volume control etc. Here’s one inside the SMSL SU-9 pro for instance:

But when people say “FPGA DAC”, they’re referring to stuff that isn’t using off the shelf chips right? Things like dCS, Chord, Meitner etc.

Well, let’s have a look inside a couple of those DACs. It is indeed true these products aren’t using off the shelf DAC chips, they’re using proprietary designs, but the FPGA itself is NOT the DAC.
Inside the dCS LINA we do indeed see a large FPGA (1), but we also see a bunch of stuff above it (2):

dCS uses their “ring DAC” circuit for the actual conversion. The FPGA is doing DSP, control, clocking management and other aspects of running the DAC, but the actual circuit converting/outputting the analog signal is the ring DAC array above, NOT the FPGA.

If we look inside the Chord DAVE, we see something similar:

The FPGA here is doing high performance DSP (oversampling and modulation), but it isn’t actually outputting the analog signal, the “pulse array” is. Once again, the FPGA is not actually the DAC.

Whenever you see a product referred to as an “FPGA DAC”, all it means is there is an FPGA inside, it does not mean the FPGA is the DAC.
In Chord DACs, the FPGA controls the pulse array
In dCS DACs, the FPGA controls the ring DAC
In Holo DACs, the FPGA controls the R2R Ladder
In Audiobyte DACs, the FPGA controls the 1-bit converting circuit
There are also plenty of DACs using FPGAs with off the shelf chips

Hopefully this illustrates that ‘FPGA DACs’ aren’t technically speaking a thing, as the DAC itself is a different circuit entirely, and I also think that only talking about the FPGA ignores a lot of the clever designs and work many manufacturers have put in to their various proprietary conversion circuits themselves.

Additionally, even if we do want to keep the term “FPGA DAC” for anything using an FPGA, we quickly find that a huge number of DACs on the market suddenly fall under that umbrella. It’s often harder to find one NOT using an FPGA of some sort. Not only stuff like the SMSL SU-9 above, but also my Motu M2 interface, DACs that people wouldn’t typically describe as “FPGA DACs” but are still using plenty of FPGAs like R2R DACs, or even ones like the RME ADI-2:

FPGAs aren’t DACs, they are programmable logic devices used for a wide variety of purposes, but I’m not personally aware of any device that is actually using an FPGA directly as a digital to analog converter

[NickZ]

I also find it surprising how little of a dacs circuitry is involved in the actual digital-to-analog conversion.

Most of it is doing processing before and after the d/a, and things like usb and the power supply.

Makes you wonder how much of the actual dac chip / r2r you are hearing, and how much is the remaining circuitry.

[Enop]

I’m very interested in hearing you expand on your views about the pros and cons of external master clocks, which you alluded to briefly in your excellent and informative video about what’s inside a dac.

[shinbojan]

Which component in the signal chain should handle PEQ? I was a bit surprised that pure streamers don’t include parametric EQ. Is that an intentional design choice?

Should PEQ be implemented at the DAC stage as it requires more processing power than streamers typically provide?

I’ve tried using PEQ on Volumio on rpi and it sounded way worse than PEACE.

[SSN757]

My setup allows for PEQ from Foobar2000. It also allows for convolution EQ, which is what I use.

The Topping A900 DAC has an EQ feature. It seems either prior to or at the DAC works well.

[sfw0]

Leaving room here to be corrected, my current understanding is that it is virtually no “work” to convert a pulse width modulated (PWM, also known as DSD or 1-bit) digital signal to analog. Per Paul McGowan (PS Audio), basically all that is necessary is to run the PWM signal through an analog low-pass “reconstruction” filter; the resultant output is analog. Perhaps then the term “FPGA DAC” could be reserved for DACs that use an FPGA to convert PCM to PWM (the algorithm known as Delta-Sigma) that then goes through a simple analog reconstruction filter (and on to the final analog buffering output circuitry shared by all DAC types).

[GoldenSound]

This is true (though worth noting DSD is a pulse-density-modulation format, not a pulse-width-modulation format. The two are similar in principle though.

I don’t think it’d be fair to call anything that converts to 1-bit PDM/DSD or anything running PWM an FPGA DAC, as technically then you’d encompass all ESS DACs for instance which are PWM.

My personal view is that the term “FPGA DAC” kinda just shouldn’t be used. I personally describe things just by what the DAC architecture actually is to avoid confusion. This can generally be separated into “off the shelf” and “proprietary/discrete design”. But you could break it down further then into things like R2R, 1-bit, pulse array, 5-bit switched-resistor etc

[Polygonhell]

@GoldenSound

But you could break it down further then into things like R2R, 1-bit, pulse array, 5-bit switched-resistor etc

The problem is that it doesn’t really convey anything.

There really are only two ways to do D to A conversion, you have something like a resistor ladder that varies output voltage based on switches, or you turn something on and off so that the average of that switched voltage is the desired output when run through a filter. Everything is some combination of those.

The devil though is in the details if you have say a 5bit direct conversion your then switching, you have choices how you use those bits either as directly decoded higher level bits, or noise in the low bits, and it’s not obvious what that means, so two DAC’s with 5bit direct conversion and PDM can do it very differently.

I can’t imagine there are very many pure 1 bit converters anymore, so either your some variant of PDM/PWM or your directly converting. I do agree that the term FPGA DAC designation is beyond misleading, there is nothing magic there, but it should differentiate using the FPGA to control some output circuit directly (vs an off the shelf DS DAC chip) from something that uses an FPGA for just digital processing, or providing USB logic.

As I’ve said before I’m not even sure the conversion mechanism is all that interesting.

[GoldenSound]

@Polygonhell

The problem is that it doesn’t really convey anything.

I think it conveys more than just whether a product has an FPGA in it or not though.
Even if wanting to just differentiate between native PCM/R2R and non-R2R stuff I’d just call it R2R and Delta-Sigma rather than confusing things with FPGAs, there are FPGAs in R2R DACs and plenty of delta-sigma DACs that don’t use FPGAs

@Polygonhell

There really are only two ways to do D to A conversion, you have something like a resistor ladder that varies output voltage based on switches, or you turn something on and off so that the average of that switched voltage is the desired output when run through a filter. Everything is some combination of those.

There are many other ways, most delta sigma DACs are not 1-bit, and don’t convert/modulate to a PDM or PWM signal.
Stuff like the AKM chips for example are high speed 5-7 bit DACs. They can be considered to operate very similarly to R2R in that they can output up to 7 bit (for the latest generation) as-is, but achieve their >7bit performance by modulating to 11Mhz or 22Mhz 7bit.

Most delta-sigma DACs are not 1-bit, as this puts far more difficult requirements on the DSP and modulator design. You need a lot more compute power and/or operational speed to get the same actual dynamic range in a 1-bit DAC as you do in a 5-bit delta sigma DAC.

Stuff like the dCS DACs for instance are often referred to as “FPGA DACs” but are also not 1-bit. They’re delta sigma, but the actual ring DAC hardware itself is 5-bit

@Polygonhell

I do agree that the term FPGA DAC designation is beyond misleading, there is nothing magic there, but it should differentiate using the FPGA to control some output circuit directly (vs an off the shelf DS DAC chip) from something that uses an FPGA for just digital processing, or providing USB logic.

I agree, which is why I think the term “proprietary DAC” or “Discrete DAC/converter” is far more accurate as that plainly describes what something is/isn’t, rather than leaving the argument about whether something may or may not be an “FPGA DAC” regardless of the fact it does or does not have an FPGA or what it’s doing