Is there a plausible scientific explanation for pursuing extreme digital accuracy?
Just curious because I can’t find anything to explain why -200, -250, -300, -350 dB out of band noise suppression or 32 or 64, 128 or 256 bit accuracy in mathematical processing from extreme digital filters should result in better sound on analog devices that are limited to about 21 bit resolution and about -140 db S/N
Has anyone come across something scientifically plausible? Rob Watts believes in it and many here believe in extreme upsampling and filtering with HQ player? Roon offers upsampling also. There is also a group on Audiophile Style that believe in the benefits of offline 256 bit upsampling - PGGB. Granted these approaches can all achieve improved and sometimes extreme levels of in band digital amplitude linearity and inhuman levels of out-of-band digital noise suppression but what could be the mechanism that accounts for audible results coming out the real world analog stage?
Are there inherent problems of running FPGA digital filters close to analog output stages and therefore removing that upsampling processing step from the one box DAC is ridding the system of some unwanted correlated noise on the analog out ? If so how and why would extreme mathematical efforts sound better than what would normally be regarded as sufficient?
Obviously different filters will sound different and we see this all the time on countless DACs - minimum phase sounds different from linear phase, as does a slow roll off as do apodizing filters that window the filter function. So my question is NOT about different sound of different filters that should sound different but what is the scientific mechanism to explain how extreme mathematical digital accuracy and extreme out of band filtering results in something audible?
I am not disputing audibility claims - just looking for a good causal hypothesis or otherwise as to why “what should be inaudible”,from a common sense perspective, is actually audible?
Depends how you view “21 bit resolution” or 140 dB S/N. That is the case in 20 kHz bandwidth case yes. However, 20 kHz wide band doesn’t mask narrow discrete tones, masking bands are much more narrow and thus contain smaller portion of the noise.
In addition, if you want to retain for example 21 bit resolution, you need to do processing with much more resolution to avoid accumulation of rounding errors. I choose the needed calculation resolution based on the particular algorithm case.
And we are not talking about out of band noise suppression, but instead about out of band image suppression which is different. But something like -350 dB is pretty pointless.
Yes there is, bigger FPGA you put there, more noise it produces, more power it consumes and more expensive it gets. In addition FPGA price/performance/power ratio is not particularly good when we talk about complex DSP processing.
Most DACs use the smallest and cheapest FPGAs for this reason. This in turn limits the DSP algorithms even more.
One of the biggest problems in these hardware implementations is that due to resource constraints the oversampling factors performed with proper digital filters are very low. Typically either 8x or 16x. And rest of the oversampling is performed with some really ugly algorithm such as S/H, linear interpolation or very low order IIR (ESS and Chord). This leaves the digital reconstruction pretty inaccurate.
Also the delta-sigma modulators implemented such way cannot be very complex either, so those end up pretty simple, limiting the performance.
Typical computer has vastly more RAM and processing capacity, so it can do things properly without shortcuts. And at lower cost, typically even without any extra cost if you already use computer for playback anyway. When done correctly, this also allows removing unnecessary components from the DAC and putting more money where it really matters. Which is the actual D/A conversion and analog stages. For example my Holo Audio Spring DACs are excellent example of such, just bit-perfect conversion and nothing else.
Am I hearing that 24 or 32 bit extra precision makes sense for calculations?
Totally agree that a powerful PC is the cheapest option for customized and extreme filtering that avoids the DAC chip limitations.
This suggests that DAC’s like Holo Spring/May and T&A D200 are a better choice for those looking to use a PC for upsampling and conversion.
I understand that some DAC designers already bypass the DAC chip (like Benchmark) limitations so there is agreement there even within the engineering & DAC design community.
Any thoughts as to what is a sufficient level of out-of-band suppression and pass band linearity given the analog performance of analog circuits and speaker/transducer performance.
Benchmark just puts their “2x” digital filter in front of the ESS digital filters and modulators. So not really bypassing anything.
It is only one of the many aspects, so I rather not try to simplify things into one or two numbers. And thinking about digital filters anyway gets you just 50% of the picture, modulators being the other 50%.