Schiit Bifrost Multibit Upsampling Experiences?

[Joebah]

I am really looking forward to taking delivery of a Schiit Bifrost Multibit in the next couple of weeks (as soon as the next batch is scheduled to go out). I use Audirvana to play everything (including my iTunes library).

 

For 176.4kHz and 192kHz input sample rates, the Bifrost runs in non-oversampling (NOS) mode. What does this mean for input sample rates that are less than 176.4Hz? Most of my music files are 24/96 or 24/192, but I also have ripped all of my CDs in 16/44 lossless.

 

I know this is probably a personal preference, but for those of you with this DAC, do you do any upsampling in Audirvana for any of your music files? If so, what settings?

 

My apologies in advance if this is a stupid question with an obvious answer. I am still very much a newby.

 

Thanks!

[Feanor]

I am really looking forward to taking delivery of a Schiit Bifrost Multibit in the next couple of weeks (as soon as the next batch is scheduled to go out). I use Audirvana to play everything (including my iTunes library).

 

For 176.4kHz and 192kHz input sample rates, the Bifrost runs in non-oversampling (NOS) mode. What does this mean for input sample rates that are less than 176.4Hz? Most of my music files are 24/96 or 24/192, but I also have ripped all of my CDs in 16/44 lossless.

 

I know this is probably a personal preference, but for those of you with this DAC, do you do any upsampling in Audirvana for any of your music files? If so, what settings?

 

My apologies in advance if this is a stupid question with an obvious answer. I am still very much a newby.

 

Thanks!

 

Not a stupid question at all.

 

I might not have it entirely right but I believe 176.4 and 192 are the limits for oversampling on the Bifrost Multibit; (I think it's 352.8 and 384 on the Gungnir MB and Yggdrasil). Thus 44.1 is oversampled to 176.4.

 

All the Schiit Multibits employ oversampling to the mentioned limits and use a "close form" filter. As I understand, "closed form" ensures that original data points -- not interpolated points -- are sent on to the DAC after filtering high frequencies. Thus I think it might be a BAD IDEA to upsample before sending to these Multibit models becauses they would have no way know which were the original, 44.1 data points.

[Joebah]

Thanks for the explanation! This quote from the Bifrost FAQ is what got me wondering: "Have fun! Upsample lower-rate music with another algorithm, and compare to our “comboburrito” filter." So I guess that confirms your thought. Seems they are saying that lower-rate music can be upsampled, but may not need to because of the sound achieved with their filter.

[Feanor]

Thanks for the explanation! This quote from the Bifrost FAQ is what got me wondering: "Have fun! Upsample lower-rate music with another algorithm, and compare to our “comboburrito” filter." So I guess that confirms your thought. Seems they are saying that lower-rate music can be upsampled, but may not need to because of the sound achieved with their filter.

 

Yes, in fact not only "not needed", but not as good in the final outcome as the Bifrost MB's own oversampling and filtering.

[Sam Lord]

...I believe 176.4 and 192 are the limits for oversampling on the Bifrost Multibit; (I think it's 352.8 and 384 on the Gungnir MB and Yggdrasil). Thus 44.1 is oversampled to 176.4.

 

For Gumby and Bimby it's 24/192 max.

 

All the Schiit Multibits employ oversampling to the mentioned limits and use a "close form" filter. As I understand, "closed form" ensures that original data points -- not interpolated points -- are sent on to the DAC after filtering high frequencies. Thus I think it might be a BAD IDEA to upsample before sending to these Multibit models becauses they would have no way know which were the original, 44.1 data points.

 

It only means you lose closed-form *unless* your external filter is closed-form as HQP now offers (AFAIK uniquely for external software). I've heard that filter on HQP briefly but not properly (see below) and not long enough for an opinion.

 

Seems they [schiit] are saying that lower-rate music can be upsampled, but may not need to because of the sound achieved with their filter.

Yes, in fact not only "not needed", but not as good in the final outcome as the Bifrost MB's own oversampling and filtering.

 

The Schiit MB architecture and closed-form filters are getting lots of great reviews, but I haven't heard anyone compare the Schiit closed-form filter with HQP's yet. I'm burning in a Bifrost but won't have a good read for a few weeks. Like everybody says it sounds very shouty out of the box, but it has lots of potential. The soundfield reaches way out right to and beyong the listener's position from the start, a *very* good sign.

 

One caveat for anybody making comparisons using HQP is to A) eliminate attenuation on the output and B) turn off dithering when sending 16/44, or whatever native rate the recording has, to the MB DACs. Otherwise you're adding one dither to another, and destroying your native data points via attenuation. And be sure to *switch both back* before returning to HQP's filtering. Hopefully all this will leap onto Roon's interface, but it will be a wait.

 

I assume that when you pre-oversample a signal feeding a Bimby, it will dither the 24 bit data nicely when converting to 16 bits. That would be another comparison to make, HQP dither options (there are many) vs. whatever Schiit does.

[Feanor]

 

 

For Gumby and Bimby it's 24/192 max.

 

That being the case Schiit Audio's copy would be a bit deceptive ...

 

The Schiit websites says about Bifrost Multibit ...

 

"How the heck can you do Bifrost Multibit so affordable? Gungnir Multibit and Yggdrasil cost a lot more!

Because it’s a much simpler DAC than either of them, using less parts, a more basic power supply, a single dual-channel DAC, and a 4X digital filter, rather than an 8X digital filter. However, it still offers amazing value, from its advanced bitperfect clock management system to USB Gen 2 input to upgradable, modular design."

 

And about Gungnir Multibit ...

 

"Gungnir Multibit: Above and Beyond

Choose Gungnir Multibit, and you get the same proprietary closed-form DSP-based digital filter as Yggdrasil, coupled to four precision Analog Devices AD5781BRUZ digital to analog converters for true hardware balancing and 19 effective bits of resolution. ..."

 

and ...

 

"So what’s the real answer? How is Gungnir Multibit half the price of Yggdrasil?

Easy. It’s a simpler DAC. The power supply has no choke-input, shunt-regulated stage like in Yggdrasil, the chassis and construction more streamlined, we’re using more surface-mount parts in the analog section, and (most importantly), we’re using 18-bit AD5781BRUZ D/A converters, rather than the mind-blowingly expensive AD5791s we use in Yggdrasil. However, the completely insane, closed-form, DSP-based, 18,000-tap filter remains in Gungnir Uber."

 

So maybe this is deceptive copy?

[Tuco1965]

The big thing here is that with the 2 non oversampled rates you are not getting the benefit of filter. The magic of the Schiit dacs is the filter. With the Bimby you can upsample Redbook to the non oversamped rate to compare that with the filter. To me there is quite a difference in sound between the two. The filter gives the most natural sound to me. YMMV

 

Only Bimby is 24/192 max

[Joebah]

The big thing here is that with the 2 non oversampled rates you are not getting the benefit of filter. The magic of the Schiit dacs is the filter. With the Bimby you can upsample Redbook to the non oversamped rate to compare that with the filter. To me there is quite a difference in sound between the two. The filter gives the most natural sound to me. YMMV

 

Only Bimby is 24/192 max

 

The vast majority of the music I listen to is 24/96, but the number of 24/192 files is increasing fairly rapidly.

 

I guess I'll just have to wait to try it and see if its working for me across the board.

 

I am hoping that I like it more than the relatively lowly V-Dac I have now.

 

Without getting into a debate about the value of 24/192 files, would I be better off converting them to 24/96 for use with the Bifrost Multibit to take advantage of the filter?

[Feanor]

The vast majority of the music I listen to is 24/96, but the number of 24/192 files is increasing fairly rapidly.

 

I guess I'll just have to wait to try it and see if its working for me across the board.

 

I am hoping that I like it more than the relatively lowly V-Dac I have now.

 

Without getting into a debate about the value of 24/192 files, would I be better off converting them to 24/96 for use with the Bifrost Multibit to take advantage of the filter?

 

NO. he "closed form" filter advantage is only relevant where files are oversampled by the Multibit itself, however there is no advantage to down-converting then having the Multibit up-convert to get back (imperfectly) to the original rate. Always retain the original rate to get the best result from the Schiit Multibits.

 

The vast majority of my own collection is 16/44.1. With the Bifrost Multibit I have notice that the greatest improvement over the Uber is with Redbook 16/44.1 resolution. That is, 24/88.2 and up sound great on the Multibit, best I've personally heard, but the improvement is greatest in case of Redbook.

[Tuco1965]

Yes Schiit multibits are all about getting the most from Redbook material.

[Sam Lord]

That being the case Schiit Audio's copy would be a bit deceptive ...

The Schiit websites says about Bifrost Multibit ...

 

"How the heck can you do Bifrost Multibit so affordable? Gungnir Multibit and Yggdrasil cost a lot more!

Because it’s a much simpler DAC than either of them, using less parts, a more basic power supply, a single dual-channel DAC, and a 4X digital filter, rather than an 8X digital filter..."

 

So maybe this is deceptive copy?

 

Odd. I wouldn't say deceptive, maybe it's a typo. The only dispute is what (4x or 8x) signal the Gumby can accept. Here's the copy on its inputs:

"Input Capability: up to 24/192 for all inputs"

[Sam Lord]

And I have another question for the Schiit gentlemen or anyone who can answer. What about intersample peaks? Let's say your feeding your Bimby a proper diet of ripped 16/44 (redbook) audio. With its 16 bit converters the bit depth remains, but the 4x oversampling will still cause intersample overs of up to, I dunno, 3-6dB? So is the level reduced in compensation? Surely it must be, but I haven't seen any documentation on it.

 

I'll send them an email.

[Feanor]

And I have another question for the Schiit gentlemen or anyone who can answer. What about intersample peaks? Let's say your feeding your Bimby a proper diet of ripped 16/44 (redbook) audio. With its 16 bit converters the bit depth remains, but the 4x oversampling will still cause intersample overs of up to, I dunno, 3-6dB? So is the level reduced in compensation? Surely it must be, but I haven't seen any documentation on it.

 

I'll send them an email.

 

I'll look forward to hearing whatever response you get from them.

 

I don't know but it seems to me the issue (or reality?) of "intersample overs" is a matter of how the oversampling process interpolates the input data points. As I understand there are different methods of doing this that yield somewhat different results.

 

I'm no expert, but I gather that the simplest method is to (figuratively) draw a straight line between the input data points and take the oversample points from along this line. A much more complex method would be "quadratic spline" interpolation which can simulate something very close to real hi-rez input ... see

. However this sophisticated interpolation method requires a lot of computing power and it's doubtful that any existing DSP chips can do it. Maybe software such as Bughead Emperor can do it running on a Core i7 computer with adequate RAM.

[Feanor]

I'll look forward to hearing whatever response you get from them.

 

I don't know but it seems to me the issue (or reality?) of "intersample overs" is a matter of how the oversampling process interpolates the input data points. As I understand there are different methods of doing this that yield somewhat different results.

 

I'm no expert, but I gather that the simplest method is to (figuratively) draw a straight line between the input data points and take the oversample points from along this line. A much more complex method would be "quadratic spline" interpolation which can simulate something very close to real hi-rez input ... see

. However this sophisticated interpolation method requires a lot of computing power and it's doubtful that any existing DSP chips can do it. Maybe software such as Bughead Emperor can do it running on a Core i7 computer with adequate RAM.

 

Anyway what is reasonably certain is that Schiit's "closed form" filter will treat incoming data, at whatever resolution, as input and pass it on to the DAC, rather than interpolation points that it, itself, might generate.

[Miska]

I'm no expert, but I gather that the simplest method is to (figuratively) draw a straight line between the input data points and take the oversample points from along this line. A much more complex method would be "quadratic spline" interpolation which can simulate something very close to real hi-rez input ... see

. However this sophisticated interpolation method requires a lot of computing power and it's doubtful that any existing DSP chips can do it.

 

Splines are very light weight way of doing interpolation. I have two such in HQPlayer and CPU load while upsampling RedBook to 705.6 kHz rate takes about 10% of one core. So about 2.5% of a quad-core CPU.

[Miska]

For 176.4kHz and 192kHz input sample rates, the Bifrost runs in non-oversampling (NOS) mode. What does this mean for input sample rates that are less than 176.4Hz? Most of my music files are 24/96 or 24/192, but I also have ripped all of my CDs in 16/44 lossless.

 

For input rates below 176.4/192 it runs the data through it's own digital filter to convert it up to 176.4/192 before passing it on to the DAC chip. For the 176.4/192 input rates data goes straight to the DAC bypassing the digital filter and allowing use of full external digital filters to output at those rates for content that has rates lower than 176.4/192...

[Feanor]

Splines are very light weight way of doing interpolation. I have two such in HQPlayer and CPU load while upsampling RedBook to 705.6 kHz rate takes about 10% of one core. So about 2.5% of a quad-core CPU.

 

That's good to know.

[Mir]

Someone compared multibit Bifrost's internal filter with external closedform filter in HQPlayer:

 

 

At the end of the day, though, and I'm getting into crazy audiophile territory, I simply don't like how the HQPlayer software sounds in general. JRiver Media Center sounds more natural. HQPlayer sounds more digital, cool, and hard. Same reason I don't use Foobar2000, actually. So, not only do I prefer the native Schiit filter, I prefer the sound of a different audio player regardless of filters. But it's still cool to play around with HQPlayer and worth trying yourself. I have yet to find an audio player with so many upsampling filter options!

 

Schiit Gungnir and Bifrost Multibit Comparison (And HQPlayer Closed-Form Filter Comparison) | Super Best Audio Friends

[misterspense]

I'll look forward to hearing whatever response you get from them.

 

I don't know but it seems to me the issue (or reality?) of "intersample overs" is a matter of how the oversampling process interpolates the input data points. As I understand there are different methods of doing this that yield somewhat different results.

 

I'm no expert, but I gather that the simplest method is to (figuratively) draw a straight line between the input data points and take the oversample points from along this line. A much more complex method would be "quadratic spline" interpolation which can simulate something very close to real hi-rez input ... see

. However this sophisticated interpolation method requires a lot of computing power and it's doubtful that any existing DSP chips can do it. Maybe software such as Bughead Emperor can do it running on a Core i7 computer with adequate RAM.

 

 

 

Linear interpolation, I hope this is not used in any hifi equipment. Or I would be very disappointed that if linear interpolation is used that audiophiles don't hear the huge quantisation noise.

Linear interpolation would do more harm than good.

[Sam Lord]

Anyway what is reasonably certain is that Schiit's "closed form" filter will treat incoming data, at whatever resolution, as input and pass it on to the DAC, rather than interpolation points that it, itself, might generate.

 

Not exactly. For the Bimby the bit depth is unchanged (non-"upsampled") but 44.1k, 48k, 88.2k, and 96k sample rates are *oversampled*. The original data points remain but other points are created between the originals.

[Sam Lord]

Someone compared multibit Bifrost's internal filter with external closedform filter in HQPlayer:...

 

I have a recently-upgraded Bimby which I used for about 1.5 years before getting an Oppo105D. My Bimby has weeks more to burn in, I'm only a week into it. So *far* it's a bit short of my Oppo105D fed DSD128 by HQPlayer. Noisier and shoutier. I need to troubleshoot HQP (4x too high a CPU load), so I'm taking this time to burn in the Bimby. But I expect the Schiit to shine eventually and I'll report then. For testing I'll equalize the situation too by running both through my Bel Canto preamp. Then I'll compare preamp vs. 105D variable out and single-ended. And I'll see what Schiit says about intersample peaks, no response yet.

[Miska]

Someone compared multibit Bifrost's internal filter with external closedform filter in HQPlayer:

 

 

 

 

Schiit Gungnir and Bifrost Multibit Comparison (And HQPlayer Closed-Form Filter Comparison) | Super Best Audio Friends

 

This is utter bullshit:

The HQPlayer closed-form filter sounds like an imitation, which is exactly what you'd expect from someone that's not directly involved with Schiit and can only try to make a closed-form filter based on what Schiit has said publicly.

 

HQPlayer doesn't try to imitate what Schiit does. I'm not really interested on what and how they do it. My closed-form filter is my take on the subject that doesn't have anything to do with Schiit's....

 

For example compared to those Schiit quotes:

However, the completely insane, closed-form, DSP-based, 18,000-tap filter remains in Gungnir Uber.

 

For RedBook to 4x Bifrost MB output with HQPlayer you get 65k-tap filter and with 8x output rates you get 128k-tap filter when using closed-form. But nothing I'd yet call "insane"...

 

 

I think this tells the whole story:

I simply don't like how the HQPlayer software sounds in general

 

He doesn't like HQPlayer in general and that's fine. But this doesn't look like objective evaluation between the two implementations... Just one more subjective opinion.

[Sam Lord]

He doesn't like HQPlayer in general and that's fine. But this doesn't look like objective evaluation between the two implementations... Just one more subjective opinion.

 

I don't know whether he intended to listen or report fairly, but I and the great majority of those trying HQP have been extremely happy with its sound. I haven't made the HQP+Roon leap yet, looking forward to it.

 

Miska since you're here, what digital output attenuation do you recommend for your closed-form filters? It seems to me that to safely handle intersample peaks yet retain *proportionally* exact samples in the output, you would need to have a simple fraction relating input and output. That is, Vout/Voriginal = *exactly* 0.500000.... which is about -6.02dB. Is that right?

[Miska]

Miska since you're here, what digital output attenuation do you recommend for your closed-form filters? It seems to me that to safely handle intersample peaks yet retain *proportionally* exact samples in the output, you would need to have a simple fraction relating input and output. That is, Vout/Voriginal = *exactly* 0.500000.... which is about -6.02dB. Is that right?

 

Just normal digital volume with dithering. Doing volume control only on two's exponents would not be very handy.

 

In any case, you need dither to correctly reduce precision of the calculated new samples to match DACs resolution so at the same time you can also apply digital volume. Actual error of the digital volume in HQPlayer is below -320 dB.

[Sam Lord]

Just normal digital volume with dithering. Doing volume control only on two's exponents would not be very handy.

Thanks, that makes sense. But I was referring to the single fixed attenuation setting (in Preferences), not user volume control. Digital VC loses any relation to closed-form filtering, maybe one more reason Schiit doesn't employ it, but that doesn't bother me.

 

In any case, you need dither to correctly reduce precision of the calculated new samples to match DACs resolution so at the same time you can also apply digital volume. Actual error of the digital volume in HQPlayer is below -320 dB.

 

Understood, thanks. But -320dB concerns me. Could you tweak it to -323, given the known bump at -322? Try sticking a maple puck under the CPU.