Chords New M -Scaler

[Jud]

1 hour ago, Miska said:

 

Order is just as useless number to describe a modulator as is number of taps to describe a filter.

 

Higher order modulator makes the increasing noise slope steeper, which in turn puts more demand on the analog filter.

 

So just blindly increasing number of taps or modulator order is not necessarily best solution...

 

Is there any significance to even vs. odd order modulators? 

[Jud]

On 7/23/2018 at 11:20 AM, ecwl said:

I have listened to DAVE with and without Blu2 M-Scaler for the past 14 months. I can tell you the sonic difference is not subtle. I completely agree with you that conceptually it seems hard to justify 1 million taps vs 164,000 taps. But it’s definitely something you should try to listen to some day.

 

You can do the same with Audirvana Plus on the Mac with the iZotope upsampling software just by moving a slider in a preference window.  In fact I think iZotope allows up to 2 million taps and as few as 10,000.  

[Jud]

4 minutes ago, firedog said:

Isn't this "filling in the gaps" a misconception? So I don't think that's the rationale behind the design. 

 

That's right. The filling-in-the-gaps metaphor is "analog intuition," which often doesn't work well with digital audio.  ?

 

However, it doesn't mean upsampling can't be beneficial (though by another means than gap-filling).  That's presumably why it's been a standard step in processing digital audio since before there were such things as commercially available separate DACs.

[Jud]

On 7/23/2018 at 7:41 AM, mansr said:

Rob Watts is technically correct in that a longer filter gives a more accurate reconstruction. However, I'm certain that nobody can hear the difference between 100k and 1M taps. Conferring such great importance on the filter length falls, in my opinion, in the same category as talking about skin effect at audio frequencies: real phenomena of limited or no relevance to audio applications.

 

Separate and apart from what filter lengths might be audibly distinct from each other -

 

Would it be roughly correct to say:

 

(1) the number of taps is the number of times (or length of time) the filter acts on the signal; 

 

(2) as a consequence of #1, a filter with a greater number of taps will cut the signal further (more steeply) than one with a lower number of taps - at least to the point where the signal disappears into the noise? 

[Jud]

20 minutes ago, rayl1234 said:

 

"Filling in the gap" equals low pass filtering.  So I would not say it is a "misconception."

 

Sure, you can do it entirely in analog... and eventually, that gets done....  (i.e.even at 705.6kHz or whatever, it is still not the "infinite" sampling rate that is analog)..... 

 

The philosophy is, you can get better system performance if you do "more" of your work in digital and leave "less" work for analog.....

 

This is no different from the philosophy behind upsampling with HQP, or even in most oversampling DAC chips, etc, though Chord, by selecting what is essentially a windowed sinc function, is embracing the further philosophy that reconstructing towards "100 correctness" is better than towards psychoacoustic goals a la MQA, for example.  Whether you agree or disagree with the digital trumps analog premise is a different matter.  (e.g. if you are an NOS believer, you would do it all in analog and maintain that is the best sounding.)

 

If you had infinite digital compute power and infinite time, you can go to an arbitrary level of "smallness" in gaps with 100% mathematical correctness on the assumption that the original digitally sampled signal was bandwidth limited (below the Nyquist rate).  i.e. if I want the "gaps" to be small enough for a 1GHz sampling rate, I can compute that.  By 100% correctness, what I mean is, you take a bandwidth limited signal (below half 44.1kHz sampling rate) and sample it at say 44.1kHz and 705.6kHz.  If I apply the reconstruction to the 44.1kHz, I can get the exact same values as your 705.6kHz (modulo any small sampling errors in ADC).  This is true for any sampling rate you choose, 1GHz, 1 gadzillion Hz.... 

 

That came from the work of Nyquist, Shannon, Whittaker....   This reconstruction is Shannon-Whtitaker interpolation.  (Rob Watts and most others call it Whittaker-Shannon -- but I am biased due to historically having one of Shannon's disciples as an advisor).  I believe it was formally proven by Shannon shortly after WW2.

 

The Chord game is to get as close to that as possible and as high of an upsampling rate as possible with today's FPGA offerings, which is a length 1.016MM filter interpolating up to 16FS.

 

But none of this has to do with the "smallness" of the "gaps," since as the people you mentioned proved, once you have got more than 2 samples at the highest frequency of interest, there are no more gaps. 

 

Once we move from math to filtering in the real world, we have to contend with considerations like aliasing (minimized by steeper filters) and ringing (made worse by steeper filters).  My impression is that oversampling allows the filter to cut enough to help with aliasing while not causing as much of a problem with ringing as there might be without oversampling.

[Jud]

31 minutes ago, ecwl said:

Except I think how iZotope defines 2 million taps is different than Chord. Chord's 1 million taps used for 16fs upsampling actually takes in say 1.41s of music (62500 digital samples) to compute each 16fs digital sample. It is my understanding that most software algorithm taps is actually a "virtual" tap length which means that if they are upsampling to 16fs, they are running 50 tap length filters to go from 1fs to 2fs to 4fs to 8fs to 16fs so 50x50x50x50 = 6.25 million. But that means they are only taking 0.00056s of music (25 digital samples) to eventually compute the final 16fs digital sample. I think Rob Watts has mentioned once that if he were to use this kind of calculation, then M-scaler should be thought of as having 125000x125000x125000x125000 taps = whatever...

 

So I think it is definitely possible to get a high-powered GPU and CPU to do the Chord type computation as sync filter and coefficients are standard math things. The issue is that to compute each digital filtered sample, you would need to store in memory the 62500 original digital samples, do the 62500 multiplications and add up the sum, and then do it again and again using 16 different sets of coefficients a million times to get to 1 million samples from the original 62500. FPGA can be programmed to have 100's of cores and store this in memory and do these computations in a power efficient manner in real-time. To do this with CPU & GPU on your PC, you would have to sort out how to manage the memory and then you would have to use your non-dedicated CPU/GPU cores to do the multiplications and additions one at a time so it would consume a lot of power. Your question is similar to asking why there are dedicated graphics chips or dedicated neural network chips. It is not because our general CPU cannot do the same computations as the GPU or the neural network chips. It's that they are not designed to do parallel processing so would take longer and be much less energy efficient. This is all fine if you want to take every song you own and upsample it to 705.4kHz flac ahead of time. But it wouldn't work if you just want to stream Tidal.

 

At least that's my understanding of the issues. Although people who actually write these software upsampling programs can correct me if I'm wrong.

 

If Chord actually has hardware that will do tasks billions of times larger than can be done now in CPUs or GPUs, and/or do them billions of times faster, I suspect the UK military, Intel, Apple, Microsoft, etc., would all long since have been knocking at his door. 

 

I could be wrong, of course.  I would be curious to know what @Miska has to say, if he cares to.

[Jud]

25 minutes ago, mansr said:

Whatever gave you that notion? A tap in a FIR or IIR filter is a well defined term. It can only mean one thing.

 

https://dsp.stackexchange.com/questions/8685/filter-order-vs-number-of-taps-vs-number-of-coefficients

[Jud]

8 minutes ago, mansr said:

Oh, and those "transient" waveforms he shows are not band limited as he claims them to be.

 

Yes, bandwidth-unlimited waveforms at a specific frequency are ever popular for that sort of thing.

 

 

[Jud]

Deleted - though it's a quote from this thread, I think it is probably off the M-Scaler topic. 

[Jud]

2 hours ago, Andyman said:

 

Are you suggesting therefore that what the M Scaler does is trivial?

 

Please don't try to pick fights where they don't exist.  He was talking about the CPU resources required, not the sonic result.

[Jud]

On 7/26/2018 at 11:07 AM, mansr said:

That difference only matters if the noise is coupled through to the DAC output at a level that is a) audible and b) correlated to the power consumption. If the computer manages to keep the noise contained, or the DAC can keep it out, your argument goes nowhere.

 

Ah, the "ounce of cure equals a pound of prevention" argument. ? 

 

Kidding of course, though conceptually it's appealing not to have the noise in the first place instead of having it and then seeking to keep it from entering the DAC. 

[Jud]

1 hour ago, ecwl said:

I think Chord DAVE and Mojo have been reviewed by multiple magazines with multiple measurements that you can simply google online. I am not sure what @jabbr is trying to say here. Are we comparing Chord to HQP+DSC1? If so, I think the onus is for HQP+DSC1 to be reviewed by neutral third parties and publish their measurements, not the other way around.

 

@Miska does post plenty of measurements of HQPlayer with various DACs.  The DSC1 is not something he is selling, but an Open Hardware design that anyone can build and/or modify as they like.  @mansr has posted measurements of Audirvana Plus using the SoX upsampling and its delta-sigma modulator (which he contributed to the open source SoX software).  And I believe he has his own DAC design available for people to build or modify as they like (again, not something he is selling).

 

As @jabbr says, these measurements, like the measurements for the M-Scaler, all seem to show the filters and modulators generally do a good job.

 

Meanwhile, anyone who likes the sound or the idea of the M-Scaler and wants to buy one, good for you.

[Jud]

10 minutes ago, mansr said:

To be clear, Miska is the one with a DAC design.

 

Thanks. My impression is you've built one and provided at least some information about it, though I haven't seen much of what you may have said. I thought you may have provided enough information for someone else to build a duplicate or modification, but I stand corrected.

[Jud]

8 hours ago, jabbr said:

I’m very familiar. It’s a terrific design particularly for learning about how a discrete DSD DAC works. He doesn’t concentrate on the analog output stage too much aside from very well designed active filters. 

 

Consider it a reference design rather than a commericial implementation. 

 

Talking about the DSC-1, or something Mansr has discussed?

[Jud]

7 hours ago, jabbr said:

I haven't seen @mansr's work. Is there a link available?

 

There is no "DSC1" commercial implementation.

There are DSC1 derivatives commercially available (at least as boards/kits) but best I can tell use transformer output stages which I'm not interested in. There are also posted schematics which discuss balancing, etc. So the DSC1 is not a specific commercially available product. I don't know the details of the Estelon implementation but understand that this includes hqplayer's naa integrated with a discrete DSD DAC.

 

 

Reason I asked is because I was talking about something Mansr had discussed doing, whereas I figured you were talking about the DSC-1, for which schematics, etc., have been published.

[Jud]

6 hours ago, mav52 said:

Lets stay on topic.   We are drifting like a bad clock

 

My apologies.

[Jud]

35 minutes ago, ecwl said:

From a jitter, clock perspective, since HMS is meant to be paired with Chord’s newest DACs which are extremely jitter immune, that’s why Rob Watts is okay with not having USB out and sticking with dual S/PDIF. 

 

"Extremely jitter immune" is words.  What @adamdea is saying is that technically S/PDIF can't have jitter as low as an async USB interface.  The reason is that the clock is local to the DAC with async USB, while with S/PDIF the DAC clock always has to try to synchronize itself with another clock.

 

This difference of course may not be remotely audible, but that's not what was being said.

[Jud]

4 hours ago, rayl1234 said:

 

I read it differently.  I read it as the particular set of coefficients that achieve a certain error bound but are computable on the particular FPGA's capacity -- computable in the sense that unlike math, you don't have infinite precision...  The term "better than 16 bits" is thrown around, but unfortunately, that is one of the areas where I've not gotten a clear answer.  Is the implication that:

 

1. Each coefficient is w/in 2^-16 of the coefficient of the mathematical 1MM windowed sinc?  (This would say very little interesting as certain coefficients accumulate much greater errors than others, obviously)

2. Compared to using a mathematical 1MM windowed sinc, the results computed will match the mathematical computation when quantized to 16 bits?

3. For all bandwidth limited signals, the result of applying the filter will result in the equivalent digital values as a 705.6kHz sampling of the analog signal. (This would be the strongest defition, but this has not been confirmed much less had a proof presented).

 

However, I think my original thesis still stands -- although technically "feasible" in software, I haven't seen a truly packaged solution even for offline conversion, much less for streaming. (I will say, not to bring up a big debate on the topic, that I have my reasons to be a bit skeptical of "5%" CPU bec my admittedly unoptimized offline test code when I tried it last fall was closer to 60% of a 4 core CPU... Of course, on a GPU with 2000+ DSP cores, it would be easy, but would probably consume around 70+ watts of power.)  Thus, if you like what WTA does for you (and I do), all other chatter is just noise until this situation changes.

 

I can't see how CPU usage for offline conversion is relevant to anything.

[Jud]

If anyone wants to see whether using what I will technically call a "shitload" of taps will make a difference, you can try Audirvana Plus with the iZotope SRC, which does anywhere from 10,000 to 2,000,000 taps. Or you can get iZotope itself on trial and do the conversion offline if you like.

[Jud]

3 hours ago, Whitigir said:

Which iZotope? And does it use WAT filter ? Or Linear phase

 

Either version of iZotope.

 

IZotope does allow you to configure the filter to be linear phase, yes.  That's a pretty routine thing.

 

Haven't read the thread closely enough to figure out what a WAT filter is (with the name similarity, I suppose it's the filter Chord uses), so I can't say how one might configure the Chord filter to be similar.

 

The main point was that if you wanted to see what a particular aspect of the Chord filter - use of lots of taps - might do for you, this would be an opportunity to try that particular aspect for free. 

[Jud]

Is another reason not to produce a version tied to hardware that the pure software version may be easier to copy, or nothing much to choose between the two?

[Jud]

6 hours ago, ecwl said:

Just like nobody who charges for their software upsampler so far has been willing to open source their code so that we can view them.

 

The SoX upsampling and modulators available in Audirvana Plus are AFAIK those available open source in SoX. 

[Jud]

2 hours ago, ecwl said:

Great. And I think @mansr is a SoX maintainer so I was hoping that he would provide a recommended setting that he thinks would approximate or beat the Hugo M-Scaler/Chord Blu2/Mojo. I'm downloading the Audirvana plus trial version now as I type this... Would also be interesting to see if iZotope 64-bit SRC is also open source.

 

iZotope is not open source, though there is a free trial.

 

Do you hear a difference among 10,000 taps, 100,000 taps, and 2 million taps with iZotope in Audirvana Plus?

[Jud]

20 minutes ago, hmartin said:

How about testing the Mojo using the optical input running on battery? The mojo is very sensitive to noise on the USB input, as many other USB DACs....

 

I have read in this one thread that Chord's hardware is less susceptible to jitter than other manufacturers' equipment, and that at least one Chord hardware product is very sensitive to noise on the USB input. This seems contradictory to me. Is it just this one product that is sensitive and not the M-Scaler or any product likely to be used with it? If I remember correctly, the "jitter immune" thing comes from a Rob Watts quote, though I haven't seen measurements.  Where does the information about noise sensitivity of the Mojo come from?

[Jud]

@ecwl, I see you have great enthusiasm for the new off topic button.  

 

I am not sure how the potential noise/jitter susceptibility of the M-Scaler and Chord DACs that will be used with it are off topic in this M-Scaler thread, and hope you will let me know why you think so.