Chords New M -Scaler

[Andyman]

32 minutes ago, mansr said:

Food ingredients impart their flavour to the finished dish. That's all I said. Anything else is in your imagination.

 

For fuck’s sake Mans. You’re just digging a hole. You’ve got enough kudos here just to say “yeah - fair enough - bad analogy” (and it was). Get over it. Even I made a mistake once (allegedly!)

 

[jabbr]

On 7/26/2018 at 2:46 PM, ecwl said:

1) See if I have a personal preference for specific filters, maybe even over Chord's

2) See if I can replicate the Blu2/HMS sound using the sinc filter or other filters from HQPlayer

3) See the CPU/GPU load from HQPlayer running these 1 million tap filters

 

Remember that to compare PCM upsampling to DSD upsampling is entirely apples to oranges. PCM upsampling is child's play in the need for CPU resources ... I upsample PCM to 768k on a dual Celeron. 

[Andyman]

7 minutes ago, jabbr said:

Remember that to compare PCM upsampling to DSD upsampling is entirely apples to oranges. PCM upsampling is child's play in the need for CPU resources ... I upsample PCM to 768k on a dual Celeron. 

 

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

[jabbr]

22 minutes ago, Andyman said:

 

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

I didn’t say that. 

 

It does use a relatively low powered FPGA (Artix-7).

 

What it does is defined by its “software”. Programming the FPGA is very difficult and without knowing details of the implementation, I expect it’s not at all trivial. That’s what you are paying $$$ for.

 

What HQPlayer and XXHE both do are also software. The difference is that these both can be placed on different classes of hardware. 

 

With HQPlayer, for example, you can integrate DSD upsampling with digital crossover and room correction. 

 

The SGM2015 server for example might be limited in number of channels as well as which filters it is capable of using at DSD512 or even DSD1024, for example. HQPlayer has no such limitations.

[Andyman]

34 minutes ago, jabbr said:

I didn’t say that...

 

No you didn’t - but actually I think you did.

[asdf1000]

35 minutes ago, jabbr said:

It does use a relatively low powered FPGA (Artix-7).

 

Off topic here but I think Ted Smith's DirectStream DAC uses the cheaper and lower powered LX16 - which can up-sample to 20 x DSD rates... something HQPlayer can't do at the moment... 

 

[jabbr]

35 minutes ago, Em2016 said:

 

Off topic here but I think Ted Smith's DirectStream DAC uses the cheaper and lower powered LX16 - which can up-sample to 20 x DSD rates... something HQPlayer can't do at the moment... 

 

The Spartan-6 is in $20 range, the Artix-7 is $200 ish (I believe) 

 

I don’t know if HQPlayer has upsampling limits because there aren’t DSD2048 DACs etc. I’d let @Miska comment on that.

 

Also no free lunch: using a resource constrained device at a higher rate doesn’t allow more complex filters. That’s the whole point really.

[jabbr]

1 hour ago, Andyman said:

 

No you didn’t - but actually I think you did.

I said that PCM upsampling is child’s play for CPU *resources* ie ARM etc can do. No comment on complexity of software — moreover wasn’t addressing FPGA — ever tried to program one? Not trivial. Try to read what I wrote.

[asdf1000]

While the topic is M-Scaler and digital filters and I do find it really interesting... in the back of our minds (or the fore) we should always keep the late Charles Hansen's great post in mind:

 

https://www.computeraudiophile.com/forums/topic/35106-how-does-a-perfect-dac-analog-signal-look-different-than-a-cheap-dac/?page=7&tab=comments#comment-713189

 

The analogue and power supply sections of a DAC are the most important aspects - digital filters are down the list.

 

Rob Watt's DACs are quite state of the art in terms of analogue section too.

 

I've heard Dave with and without Blu2 M-Scaling driving closed back headphones directly (no room effects at play) and the difference to my ears wasn't as great as I'd hope for the added cost.

 

But Dave's transparency due to it's power supply and analogue sections is quite special. Even the Hugo2 (and Qutest for those that don't need a headphone amp) is quite amazing in this regard.

 

[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. 

[esldude]

Can anyone give an explanation of what the benefits to accuracy of the signal reproduction are with upscaling?  I can only think of lowered aliasing.  

 

And no I don't consider, "just listen to it you'll hear" an explanation.  If it sounds different, something has changed about the signal we listen to....what has changed?

 

And I don't consider, "it has better transients because it is sampling faster" to be an explanation.  If you can explain why that is okay, but there aren't these super fast transients in music that people imagine there to be. 

[Ralf11]

gentle, smooth filters -- like buttah in orbit

[Confused]

17 hours ago, barrows said:

I think it is all important to step back a second, and take a breath...  Ahhhh.

 

Now it seems to me that comparisons between the HMS and, say, perhaps HQP are a bit odd in the first place:

 

Rob Watts is clearly an advocate for PCM, and only begrudgingly added a DSD mode to DAVE (not available on other Chord DACs, which convert DSD to PCM).  He clearly designs his DACs to perform best using his filters/oversampling and with PCM. Those using Chord DACs are finding good synergy with blu and likely will with HMS in the same way.

 

Miska (Jussi) is clearly an advocate for DSD, and designed both HQPlayer and his DSC-1 DAC to get the best out of oversampled DSD.  The DSC-1 DAC approach is now available (or very similar approaches) from various DAC makers, and those using that approach have found superb sound quality using HQP to oversample to DSD 256 or 512, with a DSC-1 style DAC.  Also many chip based DACs with AKM and ESS hardware will likely benefit from high rate DSD input (I personally love dSD 256 into the ESS 9038 here).

 

To me it seems clear that HQPlayer's big strength is in DSD oversampling, especially when combined with a discrete DSD DAC, not to say that it's PCM oversampling approaches are weak.  And expecting the Chord DAC to do really well with DSD, when Rob Watts himself shows little interest in DSD performance seems a bit odd.  The same could be said for expecting HQPlayer style oversampling/filtering to work as well as m scaler with Chord style DACs.

 

So, just perhaps: HQPlayer is "best" for DSD oversampling into a DSD style DAC, and m Scaler is best for oversampling into a Chord DAC.

 

Clearly we have advocates for both approaches, and hey, choices are good.

 

7

I understand what you are saying here and you make some valid points, however, I think there is a slightly different way of looking at the comparisons with HQPlayer.  As an example, there are many HQPlayer users that have DACs that work best with PCM, and they use HQPlayer upsampling PCM with the filters of their choice.  I am one such HQPlayer user.  So it seems to me that for someone with a DAC that just happens to perform well with upsampled PCM, then a comparison between the M-Scaler and HQPlayer would be perfectly valid.  In a similar way, someone who has a DAC that performs best fed upsampled DSD, could compare the performance of HQPlayer with something like the dCS Upsampler.  OK, the dCS tops out at DSD128 and is rather expensive, but it does sounds excellent and could form the basis of a comparison.  With an upsampled PCM loving DAC, you could even try the dCS Upsampler versus the M-Scaler versus HQPlayer.  So I think there are some valid comparisons to be had.  Of course, comparisons of what the M-Scaler can do with a Chord  DAC versus HQPlayer with a DSD512 capable DAC are pretty bogus, so I agree with you there.  I just think that in some situations there may be some very interesting comparisons to be had.  We have heard from some DAVE owners that upsampling PCM via HQPlayer does little for sound quality, OK fair enough.  But what about a DAC that does benefit from HQPlayer PCM upsampling, how would that sound with the M-Scaler?  In fact, it would be interesting if some DACs benefitted from both the M-Scaler and HQPlayer, whereas the DAVE only benefits from the M-Scaler.  It would at least be interesting in terms of establishing the technical reasons why, there might be something to learn here.

 

Personally, I run a Devialet that to me has the best sound quality when fed 24/192.  HQPlayer can support this perfectly, at very low cost, and using the M-Scaler to replicate this would be very expensive and only use a fraction of the M-Scaler's capabilities.  So a non-starter for me it would seem.  That said, I do have a good relationship with a dealer that stacks both Chord and Devialet products, so it might be fun to try it one day, and who knows how that might turn out!  It might be a complete waste of time, but it might just be good, for reasons we have not worked out yet.  It would be easy enough to try, so why not?

[jabbr]

49 minutes ago, Confused said:

As an example, there are many HQPlayer users that have DACs that work best with PCM, and they use HQPlayer upsampling PCM with the filters of their choice.  I am one such HQPlayer user.  So it seems to me that for someone with a DAC that just happens to perform well with upsampled PCM, then a comparison between the M-Scaler and HQPlayer would be perfectly valid.  

 

Heres a problem: are you asserting the “sound” comes from the software or hardware or both? Think about it. 

 

HQPlayer is not limited to particular hardware. When I first used HQPlayer with my then newly installed NVidia GTX1080 GPU it sounded horrible! Then I moved my DAC to an NAA (over fiberoptic Ethernet) and voila it’s great. That big ugly workstation is spewing all sorts of noise. Not near my audio anymore. That’s just one variable — another is that various HQPlayer modulators & filters work differently with different DACs — has to do with noise shaping and properties of analog low pass filters. How does M-scalar handle DAC other than Chord in this respect?

 

So look at it differently but just because some bloke somewhere got a better sound with x, y or z doesn’t tell us that A is better than x. 

[jabbr]

9 hours ago, esldude said:

Can anyone give an explanation of what the benefits to accuracy of the signal reproduction are with upscaling?  I can only think of lowered aliasing.  

“Pushing” digital noise into a region where an analog filter can remove while least (or ideally not) affecting the audio.

 

In an ideal world, the audio could be perfectly separated from ultrasonic noise which would be well above 100 kHz — and 100 kHz could be a gentle corner slope. 

I

Maybe thats just a longer longer explanation of the benefit of lowered aliasing?

[Confused]

17 minutes ago, jabbr said:

 

Heres a problem: are you asserting the “sound” comes from the software or hardware or both? Think about it. 

 

HQPlayer is not limited to particular hardware. When I first used HQPlayer with my then newly installed NVidia GTX1080 GPU it sounded horrible! Then I moved my DAC to an NAA (over fiberoptic Ethernet) and voila it’s great. That big ugly workstation is spewing all sorts of noise. Not near my audio anymore. That’s just one variable — another is that various HQPlayer modulators & filters work differently with different DACs — has to do with noise shaping and properties of analog low pass filters. How does M-scalar handle DAC other than Chord in this respect?

 

So look at it differently but just because some bloke somewhere got a better sound with x, y or z doesn’t tell us that A is better than x.  

I was not trying to assert anything, quite the opposite, I am completely open-minded about this one and interested in what might work or might not.  I agree that there are many variables in play, so any comparisons will be somewhat problematic, or at least system specific, but they might tell us something.

[jabbr]

2 minutes ago, Confused said:

I was not trying to assert anything, quite the opposite, I am completely open-minded about this one and interested in what might work or might not.  I agree that there are many variables in play, so any comparisons will be somewhat problematic, or at least system specific, but they might tell us something.

 

Right, the problem is that there are multiple audio source formats, multiple software/hardware upsamplers and multiple DACs. Assuming we stick to great equipment they will all sound good, so need long term listening comparisons, and by the time that’s done there will be a new generation of DACs CPUs and FPGAs ... ?

[ecwl]

Comparing HQPlayer to M-Scaler created more questions in my mind than before.

First off, M-Scaler clearly sounds superior to HQPlayer sinc filter or poly-sinc-xtr filter doing 1fs to 16fs conversion, sent to Chord DAVE for playback. In fact, I also prefer the sound of Mojo’s 28000 taps 16fs filter to HQPlayer sinc or poly-sinc-xtr filter although by a much smaller margin.

My standard listening tracks for these types of comparisons are Alkan piano concerto, Brubeck’s Unsquared Dance and Hamilton the musical’s It’s Quiet Uptown.

With M-Scaler, the piano sounds more life like and warmer, sinc filter makes the sound too exciting and the piano sounds harsh whereas poly-sinc-xtr filter makes the piano sound smoother but when an exciting strike of the keyboard occurs, the sound seems soft. It’s Quiet Uptown is almost like a beautiful lullaby and I’ve listened to it on a high-end ESS DAC where the music sounds ultra exciting when the vocals are supposed to be smooth and warm. The sinc filter sounds similar to my memory of the ESS DAC while the poly-sinc-filter is better and smoother but still sounded off. Unsquare Dance involves a lot of percussion instruments and hand clapping and the same theme runs here. With M-Scaler, these percussive sounds and hand claps sound ultra realistic. Poly-sinc-xtr filter makes the sound soft and the claps more indistinct. Whereas sinc filter makes the sound sharper and almost as dynamic as M-Scaler but the instruments and hand claps just don’t sound as realistic.

The new questions I found involve the following:

1. I don’t see a setting for tap length in HQPlayer. Moreover, to do 1 million taps at 16fs, you’ll need to cache 0.71s of music but HQPlayer has very little audio delay. Is it possible that HQPlayer is only capable of million taps at dsd512 but not at 16fs? I know it’s not because HQPlayer processed the file so quickly that there’s no delay because the CPU usage throughout the whole song was consistent.

2. Rob Watts says DAVE FPGA is capable of 0.1 Tera instructions per second so presumably M-Scaler can do at least 1 Tera instructions per second. Now Intel CPUs can do more complex calculations but they’re only capable of maybe 0.1 Tera floating point operations per second. My i7-3770K is capable of probably less. Yet HQPlayer max CPU usage is 5%. Is this because of shorter tap lengths or algorithmic optimization using floating point computations?

3. Ultimately, it comes down to what people have been debating about in this thread. Do filters matter? If they do, do they only matter based on their anti-aliasing abilities? Or is there more to them? Because for those who feel strongly that longer filter tap length doesn’t matter, is it because their preferred filters don’t benefit from longer tap lengths?

Rob Watts has said in the past that he has designed a few filters with the same antialiasing properties but the filter that is closest to the ideal sinc filter always sounds better. It sounds like my ears agree with him. Maybe others are optimizing filters for other characteristics such as antialiasing only and ringing and computation speed and as a result, they’re highly sceptical that tap lengths matter because the filter they’re optimizing deviates significantly from the ideal sinc function, regardless of tap lengths?

 

[Miska]

2 hours ago, ecwl said:

Poly-sinc-xtr filter makes the sound soft and the claps more indistinct. Whereas sinc filter makes the sound sharper and almost as dynamic as M-Scaler but the instruments and hand claps just don’t sound as realistic.

 

So you liked filter with less taps more than one with more "poly-sinc-xtr" has about 5 times more taps than "poly-sinc". You could try "poly-sinc-short-mp" which I use mostly, which has about half of the taps of "poly-sinc".

 

Just for fun, you could also try "closed-form" which is quite long like "poly-sinc-xtr", but completely different approach.

 

2 hours ago, ecwl said:

1. I don’t see a setting for tap length in HQPlayer. Moreover, to do 1 million taps at 16fs, you’ll need to cache 0.71s of music but HQPlayer has very little audio delay. Is it possible that HQPlayer is only capable of million taps at dsd512 but not at 16fs? I know it’s not because HQPlayer processed the file so quickly that there’s no delay because the CPU usage throughout the whole song was consistent.

 

HQPlayer doesn't pre-process tracks, all processing happens during playback. Startup delay is purely due to filter initialization. I don't have settings for number of taps because it would be stupid and useless setting. Each filter is designed to certain set of specifications I've come up with after extensive work. Filter designs involve many many more things than just number of taps. Such as shape of the roll-off curve, etc.

 

2 hours ago, ecwl said:

2. Rob Watts says DAVE FPGA is capable of 0.1 Tera instructions per second so presumably M-Scaler can do at least 1 Tera instructions per second. Now Intel CPUs can do more complex calculations but they’re only capable of maybe 0.1 Tera floating point operations per second. My i7-3770K is capable of probably less. Yet HQPlayer max CPU usage is 5%. Is this because of shorter tap lengths or algorithmic optimization using floating point computations?

 

Those "tera instructions per second" are completely useless figures.

 

Chord devices don't even use floating point, they use integer math which is much simpler. And yes, HQPlayer has a lot of algorithmic optimizations.

[Miska]

2 hours ago, ecwl said:

In fact, I also prefer the sound of Mojo’s 28000 taps 16fs filter to HQPlayer sinc or poly-sinc-xtr filter although by a much smaller margin.

 

I also have Mojo, but didn't find any way to make it perform technically properly or sound tolerable...

 

[Miska]

18 hours ago, esldude said:

Can anyone give an explanation of what the benefits to accuracy of the signal reproduction are with upscaling?  I can only think of lowered aliasing. 

 

By removing images in higher frequencies, IOW, improving accuracy. All images are distortion/error on the reproduction.

 

For example poly-sinc filters on HQPlayer have stop-band attenuation of >192 dB (more than resolution of 32-bit PCM) and poly-sinc-xtr has attenuation of >240 dB (more than resolution of 40-bit PCM). While Chord filters have 100 - 120 dB attenuation.

 

For example typical DAC chip with 8x digital filter and 0 - 22.05 kHz sweep at 44.1k input:

You can see images at multiples of the digital filter output rate (352.8 kHz). First image is just about -55 dB down. This if fully correlated with the source signal, so any intermodulation products would also create discrete correlated tones.

 

Same DAC, same source file with upsampling to 22.5792 MHz DSD:

No images and any DSD noise is random uncorrelated with the source data, so any intermodulation products would be random noise, sounding a bit like radio or tape background hiss.

 

[esldude]

1 hour ago, Miska said:

 

By removing images in higher frequencies, IOW, improving accuracy. All images are distortion/error on the reproduction.

 

For example poly-sinc filters on HQPlayer have stop-band attenuation of >192 dB (more than resolution of 32-bit PCM) and poly-sinc-xtr has attenuation of >240 dB (more than resolution of 40-bit PCM). While Chord filters have 100 - 120 dB attenuation.

 

For example typical DAC chip with 8x digital filter and 0 - 22.05 kHz sweep at 44.1k input:

You can see images at multiples of the digital filter output rate (352.8 kHz). First image is just about -55 dB down. This if fully correlated with the source signal, so any intermodulation products would also create discrete correlated tones.

 

Same DAC, same source file with upsampling to 22.5792 MHz DSD:

No images and any DSD noise is random uncorrelated with the source data, so any intermodulation products would be random noise, sounding a bit like radio or tape background hiss.

 

Which is reduced aliasing on the ADC side, and reduced imaging on the DAC side.  

I do understand that is a form of distortion and nearly all DACs display some of it.  Especially with the prevalence of half band filtering in DAC chips.   Better designs still have that at levels difficult to believe are audible.  Yet it is the only thing higher sample rates seem to provide. 

 

 

 

 

[esldude]

Here is an odd example on the other end.  On the ADC side with aliasing. 

 

Below is a spectrogram of a sweep where the Focusrite Forte is the ADC sampling at 192 khz.  The DAC is running at 48 khz for this 20 khz sweep. 

 

The background goes to light gray at -110 db.  Everything above that level shows on the spectrogram.  The noise above 70 khz is from noise shaping in the output of the DAC. You see some 2nd, 3rd and 4th harmonics of the tone which I have labeled.  These are in the DAC.  Around -100 dbFS or less.  There is an idle tone in the Forte at just short of 31 khz.  It lies at -106 dbFS.  Inputs alias around this point though that shouldn't happen.  From other tests this happens at lower sample rates as well.  Even 44 khz you have aliasing around 31 khz.  A max signal aliases at -99 dbFS.  The Forte doesn't have imaging at any appreciable level used as a DAC.  

 

 

[barrows]

Higher sample rates also allow for a very simple, discrete component, D/A conversion stage.  Such as Chord's Pulse Array, or Jussi's DSC-1 approach.

Many people feel there are sonic advantages to these simple, discrete component, conversion stages.