[JaZZ]

I know nothing. I'm still extrapolating from the 1 million taps experience, which was a surprise after all. So I wouldn't be surprised if doubling the tap count would surprise us again.

 

[miketlse]

  I know nothing. I'm still extrapolating from the 1 million taps experience, which was a surprise after all. So I wouldn't be surprised if doubling the tap count would surprise us again.

 
That would certainly surprise Rob, and probably trigger the next wave of his research.

 

[Jawed]

During his university days, Rob Watts developed his hypothesis that a 16 bit data input, would need 1 mill taps to convert this back into an analogue waveform, that the human ear/brain could not distinguish from the original.

A low-pass "perfect sinc" filter with 1 million taps will produce a different result depending upon whether you feed it 2x, 4x, 8x or 16x (etc.) sample rate data. The frequency response will vary, as will the rejection of frequencies above Nyquist. The coefficients have to vary depending upon the ratio of cut-off frequency to the upsample rate and the count of original samples that are used to compute each new sample varies with the upsampling rate, for a given tap length.

I'm unclear on what sample rate Rob was thinking of when he did his calculation, originally.

My understanding is that for a given tap length, e.g. the 164,000 taps in DAVE, Rob heard differences depending upon the upsample rate. Eventually he rejected 8x upsampling for the first stage WTA filter in DAVE (it sounded too soft) and chose 16x instead.

Also, M Scaler uses the WTA algorithm, which, as far as I can tell, means it is not the pure sinc function running with such high precision due to sufficient taps, that there is no residual error in the resulting analogue waveform, which is what Rob apparently hypothesised originally (he hadn't developed WTA back then, of course). In this case I think it's reasonable to assume that the WTA coefficients are not the same as the coefficients produced by sinc. That appears to imply that the sound of M Scaler is not the same as it would be if sinc coefficients were used instead.

The reason to use WTA is "transient accuracy", implying high-frequency, high amplitude signals are not softened by the low-pass filter (DAVE's transient speed is amazing, after all!). In theory M Scaler with WTA coefficients was chosen because it sounds better than the coefficients derived purely from sinc. If that's the case, then it would seem it already contradicts the idea that 1 million taps is enough, since the original idea was 1 million taps based solely upon sinc.

So, these two reasons combined are why I doubt a million taps at 16x upsampling is the limit for audible improvements.

Now playing: Shelly Manne [At the Blackhawk] - Poinciana

 

[romaz]

Also does anyone know if the future Chord digital amp will only work with Dave, and if so what function will Dave perform in that set-up, if the user also has a Blu2? I guess no one knows yet just trying to understand if there is any point to wait for the Digital amp if i go for Hugo 2. Thank you

Rob's digital amp will have only a DX digital input and so you will need a device with DX digital output to use it.  As of right now, this means only DAVE.

 

[romaz]

  I know nothing. I'm still extrapolating from the 1 million taps experience, which was a surprise after all. So I wouldn't be surprised if doubling the tap count would surprise us again.

Rob had stated that the purpose of Davina is to find out how many biscuits are left in the barrel, meaning how many more taps are necessary before you can no longer hear any differences.  Depending on what he learns with Davina, he told me he is prepared to push to 16M taps if necessary.

 

[romaz]

I'm not sure if these will serve the purpose with Blu2. Download the manual for the original Blu. It has word clock input and output for pro applications and dual BNC outputs for use with DACs. Blu also does not come with 75 ohm cables so I assume that you will have to have a pair on hand on when setting up the Blu2.

I agree, however, that you will not need to have the most expensive BNC cables to get Blu2 to work with Dave. This does not discount the possibility that higher end cables may potentially give you a detectable performance boost. I suspect that Rob may eventually weigh in on this question and that Roy will eventually post on his adventures with different BNC cables. Speaking for myself, unless HFC comes out with a Reveal BNC to BNC cable, I will probably get a pair of Atlas Mavros Ultra cables and call it a day.

HFC Reveal Digital RCAs will soon be available with BNC terminations available upon request.  The Reveal will have a performance level that is better than HFC's CT-1 and just below CT-1E (but closer to CT-1E than CT-1) and will cost less than CT-1.  The CT-1s and CT-1Es will be phased out in favor of the Reveal and will be priced less than the CT-1.  Anticipated price will be $700 USD for 1m ($1,400 for a set of 2).  In HFC's internal testing, their new entry level Reveal line is outperforming Nordost Odins and HFC is willing to guarantee it or your money back.  I would expect that they will easily outperform the Atlas Mavros Ultra.  
 
As with all digital coax cables, to avoid the negative impact of reflections, the ideal length of these cables will be 1.25m to 1.5m.  This is one of those situations where shorter does not mean better.

 

[romaz]

As a very happy Hugo owner (that I use at home only), I would still love to have galvanically isolated USB for the Hugo 2 as requested by so many.
 

This probably will never happen since Hugo2 is meant to be portable and galvanic isolation would eat up battery life.

 

[romaz]

Here's more information regarding Blu Mk2 that may be helpful for some to know.  
 
Blu Mk2 will have both a USB and BNC/SPDIF input.  I was afraid USB would be replacing BNC/SPDIF and this will not be the case.  While I'm not sure I would use the BNC/SPDIF input, I believe it's good to have options and the more digital inputs, the better.
 
Neither the USB nor the BNC/SPDIF input will be galvanically isolated.  Rob has stated that this is because the DX outputs will already be galvanically isolated and there's no need to have them at the inputs also.  What this means is that only when paired with DAVE will Blu Mk2 (or M-scaler) offer galvanic isolation since only DAVE has DX connections.  Perhaps, another reason to buy the DAVE instead of Hugo2 (or any other DAC) for those interested in an M-scaler.

 

[EVOLVIST]

A low-pass "perfect sinc" filter with 1 million taps will produce a different result depending upon whether you feed it 2x, 4x, 8x or 16x (etc.) sample rate data. The frequency response will vary, as will the rejection of frequencies above Nyquist. The coefficients have to vary depending upon the ratio of cut-off frequency to the upsample rate and the count of original samples that are used to compute each new sample varies with the upsampling rate, for a given tap length.

I'm unclear on what sample rate Rob was thinking of when he did his calculation, originally.

My understanding is that for a given tap length, e.g. the 164,000 taps in DAVE, Rob heard differences depending upon the upsample rate. Eventually he rejected 8x upsampling for the first stage WTA filter in DAVE (it sounded too soft) and chose 16x instead.

Also, M Scaler uses the WTA algorithm, which, as far as I can tell, means it is not the pure sinc function running with such high precision due to sufficient taps, that there is no residual error in the resulting analogue waveform, which is what Rob apparently hypothesised originally (he hadn't developed WTA back then, of course). In this case I think it's reasonable to assume that the WTA coefficients are not the same as the coefficients produced by sinc. That appears to imply that the sound of M Scaler is not the same as it would be if sinc coefficients were used instead.

The reason to use WTA is "transient accuracy", implying high-frequency, high amplitude signals are not softened by the low-pass filter (DAVE's transient speed is amazing, after all!). In theory M Scaler with WTA coefficients was chosen because it sounds better than the coefficients derived purely from sinc. If that's the case, then it would seem it already contradicts the idea that 1 million taps is enough, since the original idea was 1 million taps based solely upon sinc.

So, these two reasons combined are why I doubt a million taps at 16x upsampling is the limit for audible improvements.

Now playing: Shelly Manne [At the Blackhawk] - Poinciana

Man, that is a really great post! Only, I think that all I understood was A, E, I, O, U and sometimes Y.

Could you dumb that down a little bit for this dummy?

 

[Jawed]

When you design a low pass filter, the upsampling rate (8x or 16x, say) affects the design, and the sound of the filter. 8x upsampling wasn't high enough for 164,000 taps in DAVE. Perhaps 16x isn't enough for 1 million?

When you've chosen the upsampling rate and the number of taps, the choice of algorithm (e.g. WTA or sinc) affects the sound. The original theory is that sinc sounds best when you have enough taps. If you have less than enough taps, then you use WTA (and there are other choices, which Rob has rejected as inferior) to apply a correction for the error. M Scaler uses WTA, implying that it sounds different from sinc. This might imply that sinc is not accurate at 1 million taps.

So, both of those ideas make me suspect Rob's going to be busy exploring more than 1 million taps.

Now playing: Calvin Party - Repetition No. 2

 

[esimms86]

HFC Reveal Digital RCAs will soon be available with BNC terminations available upon request.  The Reveal will have a performance level that is better than HFC's CT-1 and just below CT-1E (but closer to CT-1E than CT-1) and will cost less than CT-1.  The CT-1s and CT-1Es will be phased out in favor of the Reveal and will be priced less than the CT-1.  Anticipated price will be $700 USD for 1m ($1,400 for a set of 2).  In HFC's internal testing, their new entry level Reveal line is outperforming Nordost Odins and HFC is willing to guarantee it or your money back.  I would expect that they will easily outperform the Atlas Mavros Ultra.  

As with all digital coax cables, to avoid the negative impact of reflections, the ideal length of these cables will be 1.25m to 1.5m.  This is one of those situations where shorter does not mean better.

Thanks Roy. I guess now I know which cables I'm going to buy. Will probably get a Reveal power cable to go with it. The Atlas Mavros at approximately 520 USD per pair are roughly a third the price of the Reveal BNCs. Of course, when you figure in the fact that my entire system is connected with HFC cables there's system synergy that makes for an even larger improvement.

 

[romaz]

Thanks Roy. I guess now I know which cables I'm going to buy. Will probably get a Reveal power cable to go with it. The Atlas Mavros at approximately 520 USD per pair are roughly a third the price of the Reveal BNCs. Of course, when you figure in the fact that my entire system is connected with HFC cables there's system synergy that makes for an even larger improvement.

You and me both! 

 

[EVOLVIST]

I wish Wireworld made BNC cables. I'm so stuck on their PCs, having a Silver 7 BNC would be choice.

 

[iDesign]

  Yes life has been extremely busy - last week was finishing production code for Blu 2 (tested and delivered to Chord) and finishing Davina PCB (just front panel left to do).
 
I have just arrived in Japan with John, picking up another award for Dave and next week I plan to do a full posting about Hugo 2 on my blog showing all the slides in full plus adding some more comments.
 
But just to clarify - the Hugo 256 FS filter is identical to Dave's - and with the M scaler and Dave the HF filter option is best off as I have innately improved the stop-band rejection - so noise isn't needed to be filtered away because it's already gone. That was why it was odd that 44.1 sounded better overall with the filter on, as my expectation was for no change as rejection of out of band noise was already 130 dB, and in the critical areas (around multiples of the sample rate) with the most noise present, was about 150 dB rejection.
 
Rob

I'm looking forward to reading your full posting about Hugo 2. In your post I'm wondering if you can please share details about the Hugo 2's charging design as it relates to lithium ion battery longevity for owners who wish to keep the Hugo 2 continuously charged in a 2-channel or desktop application? I realize Hugo 2 is now charged via USB micro but I hope Chord still includes a USA wall charger so owners don't need source their own.
 
On a DAVE related note, it took me several hours tonight to find the Vaughan Williams, Academy of St Martin-in-the-Fields, Neville Marriner, Agro (1986), CD in my library because it wasn't in the right spot. After listening to the album, I am continually amazed by how your DAC designs seem to excite the entire surface of the headphone's driver, giving the impression of music filling the x, y, and z axes. And because DAVE is so musical, the listening experience is spectacularly immersive-- few other pieces of audio kit I have ever purchased can do this. Fantastic stuff, Rob. 

 

[Rob Watts]

When you design a low pass filter, the upsampling rate (8x or 16x, say) affects the design, and the sound of the filter. 8x upsampling wasn't high enough for 164,000 taps in DAVE. Perhaps 16x isn't enough for 1 million?

When you've chosen the upsampling rate and the number of taps, the choice of algorithm (e.g. WTA or sinc) affects the sound. The original theory is that sinc sounds best when you have enough taps. If you have less than enough taps, then you use WTA (and there are other choices, which Rob has rejected as inferior) to apply a correction for the error. M Scaler uses WTA, implying that it sounds different from sinc. This might imply that sinc is not accurate at 1 million taps.

So, both of those ideas make me suspect Rob's going to be busy exploring more than 1 million taps.

Now playing: Calvin Party - Repetition No. 2

Absolutely - when we talk about a finite tap length we are already well away from theory, as that implies an infinite number of taps, infinitely over-sampled - and infinite squared is a big number!
 
And the decision to split the job over two WTA filters (WTA 1 getting you to 16FS, WTA 2 getting you to 256FS) plus a non WTA digital filter (3rd order getting you to 2048FS) was based upon listening tests to see which arrangement sounds the best. And you are trying to reduce the peak error (or the peak difference from the original un-sampled signal in the ADC at any one moment in time). So peak error is what I talk about when I say that the M scaler, guarantees that the peak error at 16FS is better than 16 bits accuracy as the coefficients are identical to an ideal sinc function to the 16 bit level.
 
Now this does not mean that if one uses a 16 bit source, it is good enough - and that was never my original idea - it just occurred to me that if you wanted 16 bit accuracy at all times from one sample to the next then you needed something like a million taps. Now to give you an example - the peak error is an instantaneous distortion or difference from the original un-sampled continuous waveform in the ADC. So nobody would say if I have a 16 bit source than I need an amp that is 16 bit accurate in distortion (0.0015% THD) and so if I used an amp that had ten times lower distortion than the ideal 16 bits I would not be able to hear the difference; that suggestion would be absurd, as the distortion would be audible whatever the source (assuming reasonable recording quality). But I am not about to disappear for a year and design a 16M tap filter and then find out that 1M is enough - Davina will tell me exactly how much loss in sound quality we get in going from 768>48>768 using an M scaler; and I will publish these test recordings, so all can hear the losses that the decimation interpolation actually gives us.
 
Another point - people have posted about HD recordings, and why they don't make a huge difference to SQ, when WTA filter clearly does. It's possible to explain that. Now a conventional interpolation filter (and MQA) is about 6 to 8 bits accurate for the interpolated samples with transients; now when you use a 44.1k and replace that with a 88.2k recording then the delta from one sample to the next is halved as you now have twice the samples; this in turn, for a given conventional interpolation filter will halve the peak error. So instead of being 6-8 bits it will now be 7-9 bit accurate; doubling the sample rate is only halving the peak error; so a 192k recording will only be 8-10 bits accurate with conventional filters. But with the WTA M scaler we are guaranteed better than 16 bit accuracy for all signals for every 768 kHz sample; this is why HD recordings do not guarantee great sound, as they are all a long way away from 16 bit performance.
 
One other point - this is a very complex subject - the number of samples the filter process's is important, and there are factors I do not talk about as it is way too complex. Now it's good to talk about tap length, and for sure increasing the tap length gives you better sound - but this is not the whole story - the oversampling rate, the algorithm used, the coefficient quantization, and other factors are also very important. And there have been some odd things about the M scaler that I still have not fully digested yet - why does depth get so much better with it - this is something I have not experienced before with increasing tap length. Why is the M scaler so much better and much bigger than I expected? One poster (@DaveRedRef-III) PM me about this and said its like a camera getting into focus - and that last adjustment finally everything snaps into place, and the small changes to focus has a much bigger impact than before. I like that analogy. It also implies that 16 bit at 768k is close to ideal - and that is where Davina will tell me for sure.
 
Rob