[Yuri Korzunov]

   
Sure, 20dB difference, but when the worst performer already has a noise level 150-160dB below full-scale, I think it's about time to readjust your priorities.

Worst result is noise -121.7 dB, best -147.2 (in both articles). It's for DSD decoding/encoding.
 
For PCM 24/96 we have -146.3 dB.
 
Also pro software work at levels of noise about -160 ... -180 dB.
 
It give advantages for keepeng weak fragments of music pieces (-60 ... -80 dB) and allow do multi-stage processings without significant noise level accumulations during music production.

As example, in my software used 64-bit float point (C double precision) calculations only. It's slowest math, however allow use long IIR filters with minimal loss for rounding errors, that accumulated for each operation. Also floating point format allow avoid internal overloading. After all processing we can correct it via soft cliping or decreasing of volume.
 
As is known during oversampling possible overloading even without any gain.

 

[limpidglitch]

It's 150dB or below within band, that's all I need to know.

 

[Yuri Korzunov]

   
Sure, 20dB difference, but when the worst performer already has a noise level 150-160dB below full-scale, I think it's about time to readjust your priorities.

 
 

  It's 150dB or below within band, that's all I need to know.

 
Sorry, I don't understand. Need more info.
 
1. What is term performer here? It's conversion software?
 
2. Where 150 dB in the articles? Or what it is value "150 dB"?

 

[stv014]

  Worst result is noise -121.7 dB, best -147.2 (in both articles). It's for DSD decoding/encoding.

 
Much of that difference is due to the noise shaping algorithms used when encoding, and how much of the high frequency noise is lowpass filtered when decoding.
 

Originally Posted by Yuri Korzunov /img/forum/go_quote.gif
 
Also pro software work at levels of noise about -160 ... -180 dB.
 
It give advantages for keepeng weak fragments of music pieces (-60 ... -80 dB) and allow do multi-stage processings without significant noise level accumulations during music production.

 
Multi-stage processing is of limited relevance here, as there is not much point repeatedly converting back and forth between PCM and DSD. And in my opinion the ideal case is skipping the use of DSD entirely. PCM uses the same overall bit rate more efficiently (DSD is just 1-bit PCM at an extreme sample rate, but for the same data size as with DSD64, one could use for example 176400 Hz/16-bit instead, which still gives 88.2 kHz bandwidth for those bat-eared audiophiles, and even with noise shaping to achieve 20-bit quality in the audio band, less high frequency noise than DSD64), and is more suitable for processing.
 

Originally Posted by Yuri Korzunov /img/forum/go_quote.gif
 
As example, in my software used 64-bit float point (C double precision) calculations only. It's slowest math, however allow use long IIR filters with minimal loss for rounding errors, that accumulated for each operation. Also floating point format allow avoid internal overloading.

 
The use of floating point arithmetic in DSP software is not exactly a new or special feature, and modern CPUs can also perform operations on these data types very fast. It was only in the 1990's that integer/fixed point arithmetic provided huge performance benefits in PC software.

 

[stv014]

  Sample rate conversion is enought complicate processing

 
It is a mathematically well defined operation (which was also discussed in this thread), the only trade-offs are in filter design and code optimization. The difficulty is therefore mainly down to how important the computational efficiency (CPU usage at a given quality) is.

 

  1. What is term performer here? It's conversion software?

 
"worst performer" = the worst performing conversion software in the test
 

2. Where 150 dB in the articles? Or what it is value "150 dB"?

 
I guess it is from the FFT plots. Obviously the overall level of noise over the entire spectrum is higher than that, but most likely still inaudible as CD quality PCM typically has the noise floor at about -130 dBFS, depending on the window size/type/averaging used for the analysis.

 

[Yuri Korzunov]

   
It is a mathematically well defined operation (which was also discussed in this thread), the only trade-offs are in filter design and code optimization. The difficulty is therefore mainly down to how important the computational efficiency (CPU usage at a given quality) is.

 
Mathematically - yes. Practically - no. Filters is very complex. Exists different algoritms of calculations of coefficients, realisation and, of course, creative approach and "know how" for better result.
 
 
 
Quote:

  I guess it is from the FFT plots. Obviously the overall level of noise over the entire spectrum is higher than that, but most likely still inaudible as CD quality PCM typically has the noise floor at about -130 dBFS, depending on the window size/type/averaging used for the analysis.

 
Here better see complex features like signal noise ratio.
Integrated spectrum for estimation of conversion algorithms is not enought informative. Due for different frequrency we can see different spectrum.
On integrated spectrum all it merged.
 
For estimation conversion algorithms quality I primary use sweep sine 0 dB from 0 to 1/2 sample rate on time-frequency diagram.
In tests by infinitywav.ca used more soft test - sweep sine with level -6 dB.
I allow estimate:
 
1) Noise floor / non-linear distortions (conversion/filtration) artefacts in full frequency range. Fully control quality of filters.
 
2) Momentally detect overload and unevenness frequency response.

 

[stv014]

  Simpler DSD ADC/DAC allow get same result with lower cost. And many (most? all?) consumer PCM ADC/DAC based on DSD.

 
The theoretical cost increase with PCM is not really an issue, since there are already good enough PCM converter chips for <= $10. PCM support also needs to be included anyway, as it would be difficult to sell a DSD-only DAC to a reasonably large market. Additionally, even though modern oversampling PCM DACs use delta-sigma modulation internally, it is now usually multi-bit, unlike DSD. This also allows for simpler and cheaper analog filters than what is required for 1-bit DSD, because of the lower level of high frequency shaped noise that needs to be removed.
 

Currently 64-bit floating point remains slowest, comparing with 32-bit floating point and integer.

 
It may be slower, but the difference is generally within a factor of 2, especially if vector instructions are not used (4xfloats with SSE vs. 2xdoubles with SSE2), and memory bandwidth is not an issue like it may be with large FFTs. With simple non-SIMD FPU math, I have seen comparable performance with floats and doubles. In any case, audio software with 64-bit internal precision existed for more than a decade.

 

[Yuri Korzunov]

   
The theoretical cost increase with PCM is not really an issue, since there are already good enough PCM converter chips for <= $10. PCM support also needs to be included anyway, as it would be difficult to sell a DSD-only DAC to a reasonably large market. Additionally, even though modern oversampling PCM DACs use delta-sigma modulation internally, it is now usually multi-bit, unlike DSD. This also allows for simpler and cheaper analog filters than what is required for 1-bit DSD, because of the lower level of high frequency shaped noise that needs to be removed.

 
I suppose devices less $10 haven't native PCM DAC. There builtin DSD/PCM modulator/demodulator.
 
Native PCM DAC more faster than DSD DAC. I suppose I now native PCM DAC used in narrow pro applications only. As example in wide band telecommunication.
 

  It may be slower, but the difference is generally within a factor of 2, especially if vector instructions are not used (4xfloats with SSE vs. 2xdoubles with SSE2), and memory bandwidth is not an issue like it may be with large FFTs. With simple non-SIMD FPU math, I have seen comparable performance with floats and doubles. In any case, audio software with 64-bit internal precision existed for more than a decade.

 
I many time ago perform comparing test by calculation precision performance. However if difference within 2 times - it's huge difference in software performance. If you can improve 20% it's already good

 

[stv014]

Mathematically - yes. Practically - no. Filters is very complex. Exists different algoritms of calculations of coefficients, realisation and, of course, creative approach and "know how" for better result.

 
Well, having a look at samplerateconverter.com, I guess for someone who sells an "audiophile" sample rate converter for up to $250, that kind of opinion is not unexpected when there are free alternatives. Never mind, then.

 

[Yuri Korzunov]

   
Well, having a look at samplerateconverter.com, I guess for someone who sells an "audiophile" sample rate converter for up to $250, that kind of opinion is not unexpected when there are free alternatives. Never mind, then.

 
I suppose, selling software is not only distributing unchanged digital copies, isn't it?
 
Also it is permanent researching, development and user supporting.
 
Alternatives is good things. It base of freedom of choice. And stimulus for improving of commercial product.

 

[Rodrigues]

Alguém conhece alguma gravadora que vende versões DSD e PCM da mesma gravação? Quero fazer um teste e preciso de algum material de origem

 

[sander99]

@Rodrigues: If you want to do a comparison yourself then you should convert a file yourself. That is the only way to make sure that you are not comparing 2 different masters.

 

[Jose in Jersey]

Not only does NativeDSD.com offer both formats of the same file sometimes, they actually do recordings in pure DSD/DXD. And multichannel stuff as well. Absolutely phenomenal material offered for those curious about DSD.

 

[gregorio]

Not only does NativeDSD.com offer both formats of the same file sometimes

There’s no way to be sure of that. Certainly on some SACDs and other DSD sites the PCM versions, particularly 16/44 versions, have added compression, even though they’re listed as the same. The only way to be sure, as @sander99 suggested, is to do the conversion yourself.

they actually do recordings in pure DSD/DXD.

That doesn’t actually mean anything, DXD is PCM so I can’t see what is pure about a DSD/DXD recording. “Pure DSD” sometimes refers to a DSD recording that hasn’t been converted to PCM for mixing or mastering, sometimes that means it will be of relatively low fidelity/resolution because it is just copied from an analogue tape master, other times it means the recording was mixed on an analogue system in real time and then recorded to DSD without further mixing or mastering and again the result is likely to be of lower quality/fidelity than doing it traditionally.

G

 

[bigshot]

Pentatone is the only label I know of where you can be confident that the redbook layer is the same mastering as the SACD layer. They do "pure DSD" which is essentially direct to disc. That works fine for some kinds of music (the disc I did an ABX on was chamber music), but you aren't going to find a pure DSD Pink Floyd album.

I've ran into a Rolling Stones SACD where one of the songs was a completely different mix on the two layers... with different instrumentation!