Schiit (MultiBit) Bifrost (DAC digital filter techno-discussion)

Copied from the "Schiit (MultiBit) Bifrost" Thread (Schiit (MultiBit) Bifrost)

Octave code included.

+++++

+++++

Uploading only 32-bit wav

I will try to upload more files, but need Matlab instead of Octave and I'm going camping soon.

If you have access to a Bambi and have a means to use these filters in NOS mode, by all means post your results and impressions.

Can have technical discussions here. I'm hoping this thread will not turn into an "appreciation" HF like thread, nor a "depreciation" HC like thread. Just a thread for open and honest discussion (both technical and non-technical - I kind of hate the objective/subjective bullshit labels)

Worst case, if this thread does not pick up, at least it will serve as storage for filters folks can try, and modest source code that people can check out. Maybe.

 

Here is a new set using Matlab which is capable of a 64 bit .wav and a rejection of about 200 dB.

This is an 18001-tap filter using chebyshev windowing and requiring input and output samples to be the same (similar to the Burrito Combo). This is pretty much what Thune coded.

The file is unity gain, but does not use the full dynamic range. Can do the later if desired. Can do less taps if desired as well.

I'll add more files as I check them in the next posts.

Again, change the extension to .wav

Implementation is straight up convolution (assuming that's what JRiver needs). I one needs equivalent coefficients using cascaded 2nd order or whatever order, let me know.

Assumed interpolationis x4 for a final rate of 176.4 kHz (44.1 kHz original signal) or 192 kHz (48 kHz original signal)

File here: >Schiit (MultiBit) Bifrost (DAC digital filter techno-discussion)<

 

It seems it's possible to put Bifrost in NOS mode and use JRiver to do the Digital Filtering and bypass the Mega Burrito per Jason:

Schiit (MultiBit) Bifrost

This allows for testing of different Digital Filter approaches. Best guess is that this only applies to the multi-bit option.

I'm not sure if "images" would be a problem using the NOS mode approach (vs the built-in Mega Burrito Combo). But it can be verified.

 

Attached is a quick and dirty Parks-McClellan 64-bit set of coefficients.

Rejection = ~120 dB
Number of coefficients = 329
Corner = 0.907
Error difference between input and output samples = ~ -100 dB

Filter here: >Schiit (MultiBit) Bifrost (DAC digital filter techno-discussion)<

 

Attached is a quick and dirty LS 64-bit set of coefficients

Rejection = ~ 130 dB
Number of coefficients = 855
Corner = 0.96
Error difference between input and output samples = ~ -72 dB

(as in all cases, these are for 4x interpolation)

Filter here: >Schiit (MultiBit) Bifrost (DAC digital filter techno-discussion)<

 

Attached is a quick and dirty Windowed (very sharp) Kaiser 64-bit set of coefficients

Rejection = ~ 284 dB
Number of coefficients = 18001
Corner = ~1
Error difference between input and output sampels = ~ -288 dB

Filter Here: >Schiit (MultiBit) Bifrost (DAC digital filter techno-discussion)<

 

It seems to me that Windowed approaches (which are not approximations in the Parks-McClellan iterative sense, and maybe considered "closed-form" depending on how we define that term in this context), are closer to what Burrito Combo does.

Note the assumption is that the number of taps is not a problem (18001-taps is kind of highish). Another assumption is that one can play with 64-bit numbers (which is kind of highish too). I guess, processor power has evolved.

There is a large family of windows that can be applied with a large set of different parameters. Let me know if anyone has a specific flavor in mind.

Cheers!

(BTW, James McClellan is not from the WW1 era - maybe Chebyshev was - he is pretty much alive and maybe one can get ones man-boobs signed by him if one goes to Gerogia Tech. Thomas Parks - McClellans old instructor - is still around, but a bit older)

 

Wanted to point out a few things I've found out through trying a bunch of stuff.

LARGE ORDER FIR AND (THIS OR THAT) WINDOW VS. (THIS OR THAT) OPTIMIZATON

If the size of the filter is somewhat constrained (like a lot less than 18001-taps), then Parks-McClellan (and other random stuff) seems like a good choice.

If one goes the Mega Burrito Combo way and throws in a 9001-18001 tap filter, the optimizaton algorithms start to kind go unstable, and one might be better off using a window technique. At some point one is fairly close to the ideal sinc anyway.

POSSIBLE LIMITATIONS

As the size of the filter goes up, one starts having problems with bit-width precision. From what I've seen, an 18001 tap filter using windowing should be able to offer more than 120 dB of supression, unless the filter is limited by precision. The SHARC processors seem like 32-bit, but if they were crazy 64-bit I bet the same filter would be more powerful (and more demanding).

One also gets some latency which starts approaching a few fractions of a seconds. For some movies one might start using lip-synch or something like that depeniding on rate and other stuff.

BOTTOM LINE

Still, bottom line is, for 18001-tap filter and with 32-bit processing at one's disposal, might as well just go with the Windowing approach, like I assume Schiit did and look awesome while at it.

Again. Let me know if you guys want some other type of filter or window to test.

 

This is pure speculation, but audio purposed D/S (and segment-mixed) DACs may have a built-in DAC digital filter (or a set of them), and already have a serial I2S standard interface.

The MB DAC uses a parallel interface and may not have a built-in DAC filter (because it's not an audio purposed DAC chip). All of these functions would naturally have to be custom and require a separate processor such as the AD SHARC stuff.

 

Looking at the AKM4490 Data sheet (<linky) the DAC filter is part of the DAC chip. This is probably standard in most audio purposed DACs.

They offer 5 types of 32-bit digital filters:

32-bit 8x Digital Filter (Ripple: +/-0.005dB, Attenuation: 100dB)
- Short Delay Sharp Roll-off, GD=6.25/fs <-default
- Short Delay Slow Roll-off, GD=5.3/fs
- Sharp Roll-off
- Slow Roll-off
- Super Slow Roll-off (SSLOW bit)

There is also a de-emphasis filter which can be switched-off.

I don't think the other filters can be switched off, just selectable, which would rule out the usefulness of an external megaburritocombo. I also don't think one can program the megaburritocombo in the DAC chip.

 

Thanks for the correction! So actually, yes, the megaburritocombo might be possible with the 4490 then.

 

Yes, going multibit may make the burrito filter trickle down easier. But then Schiit products IMO tend to strive for providing options to it's customer base. Who knows. Maybe Schiit will pull out another daughter card configuration option.