A toast to PGGB, a heady brew of math and magic

[Zaphod Beeblebrox]

Quote

The section appears to be hastily written in response to yesterday's discussion....., it contains a lot of misspelled words. However the charts are interesting as it is exactly the kind of experiment I proposed.

I'm very curious how PGGB achieved a very long filter with a steep corner frequency with apparently less ringing than a short Gaussian filter.
The Gaussian filters occupy the minimum surface area in the time-frequency domain and there's usually no cheating math. I know PGGB says the algorithm doesn't do windowed sinc but still it seems implausible to me they have found a way to work around this mathematical tradeoff.

I have corrected some typos/grammar in the FAQ section. 

 

I am responding here to a question that @plupudding2 asked on the other forum as this is a very good question, one which I had initially added an explanation for, but felt it may be confusing to many.

 

As for as the graphs go what you see is what you get. It is true that a Gaussian filter will have a minimal footprint both in frequency and time, but the Gaussian filter is apodizing and needs to have a corner frequency earlier than the Nyquist (say 21kHz instead of 22.05kHz for CD rates), while PGGB will have a corner exactly at 22.05kHz. Any ringing you see in time domain is the result of the filter having to dissipate the energy beyond its cutoff frequency. An apodising filter has to deal with all the leftover energy starting from its cutoff frequency till the Nyquist that was left unfiltered by the ADC. While PGGB has to deal with whatever little energy that is left at the Nyquist after the ADC has already done what it can. So, it is not entirely surprising that you see less ringing with PGGB on a reasonably band-limited signal. 

 

 

[The Computer Audiophile]

I split several posts into a new topic called PGGB Performance Tuning, thanks to help from @austinpop. Please let me know if additional cleanup needs to happen. 

 

 

[austinpop]

1 hour ago, The Computer Audiophile said:

I split several posts into a new topic called PGGB Performance Tuning, thanks to help from @austinpop. Please let me know if additional cleanup needs to happen. 

 

Thanks for doing that. I thought it would be good to get the nitty gritty data analysis off on its own thread, so we could keep this thread devoted to PGGB function and sound quality.

 

Geeks who want to dive into the nitty-gritty performance details are welcome to join the performance tuning thread.

[Zaphod Beeblebrox]

Version 6.1.42: The DSD Edition (PGGB Plus)

I am happy to announce a faster PGGB. You can grab v6.1.42 here. We were able to complete the optimizations faster than we expected, we have completed testing of these optimizations for PGGB DSD conversions. The optimizations do not affect PCM conversions, they will perform as before. We see 2-3x speedup across all platforms for DSD conversions (both single and two stage) across all rates and modulators.

 

Release Notes:

Version 6.1.42: The DSD Edition (PGGB Plus)
    SQ Changes:    None, same as version 6.0.42
    Features:
        PGGB DSD is now 2-3x faster on all platforms*

 

*We are seeing 3x speeds when all of the processing fits in RAM and 2x when there is not a considerable slowdown due to NVME read/writes, YMMV applies. Please discuss/report performance in this thread. 

[kennyb123]

3 minutes ago, Zaphod Beeblebrox said:

PGGB DSD is now 2-3x faster on all platforms

Confirming this is what I saw on Mac.  2.6x faster with one track I compared.  
 

I should also mention that I’ve had to move away from using auto for workers.  My 16 core Xeon Mac Pro is faster with workers set to 16 using the hidden menu.  I have not yet figured out why.

[Mista Lova Lova]

2.8x faster for me (Windows 11) when tested on a track that's not bottlenecked by RAM (I currently "only" have 64GB available).

 

Thanks, @Zaphod Beeblebrox!

[GryphonGuy]

2.2 times faster with 2 x DSD64 tracks upsampled to DSD512.

Windows 10, limited to 24 workers with 192GB RAM

I chose Auto (9th order mod) but system log shows 7th order mod used for track 1 and 9th order mod used for track 2,

Track 1 RAM estimate was 92.2GB

Track 2 RAM estimate was 181.5GB

 

Processing time on each track in analysis spreadsheet adds up to 20 mins 22.8 secs but album debug spreadsheet says processing time was 29 mins 20.4008 secs.

 

Not sure where the almost 9 mins extra went as the time was closer to the stated 20 mins if you ask me.

[GoldenOne]

I've just posted a video demonstrating that not only can high performance oversampling indeed make an audible difference vs standard filters, but therefore DACs themselves even if 120dB SINAD can also still sound different and are in fact not necessarily 'Audibly Transparent'

 

[Zaphod Beeblebrox]

2 hours ago, GoldenOne said:

I've just posted a video demonstrating that not only can high performance oversampling indeed make an audible difference vs standard filters, but therefore DACs themselves even if 120dB SINAD can also still sound different and are in fact not necessarily 'Audibly Transparent'

 

Thanks for sharing this. You have gone through a lot of effort to make the ABX testing 'air-tight ', this would be a great foundation for future tests, kudos for going through this and congrats on your hearing!

 

Your being able to hear till 21kHz (which is an exception than a norm I hope) provides a simple explanation. But I wonder if you have compared different upsampling filters that have a cut-off frequency beyond 21kHz?

 

The reason I ask is I cannot hear past 15kHz, and I am sure I am not in the minority for my age (middle-aged male), and yet be able to appreciate differences in DACs and oversampling filters. These differences can as easily be attributed to how the DACs' filters or filters used by software upsampling algorithms differ in their reconstruction of transients. These filters can still be flat to 20kHz and yet sound different. It is harder to tell upsampling filters apart by looking for tonal differences (at least when using different filters on the same DAC). Can you explain what you were hearing for and how you were able to tell them apart? 

 

 

 

[GoldenOne]

27 minutes ago, Zaphod Beeblebrox said:

Thanks for sharing this. You have gone through a lot of effort to make the ABX testing 'air-tight ', this would be a great foundation for future tests, kudos for going through this and congrats on your hearing!

 

Your being able to hear till 21kHz (which is an exception than a norm I hope) provides a simple explanation. But I wonder if you have compared different upsampling filters that have a cut-off frequency beyond 21kHz?

 

The reason I ask is I cannot hear past 15kHz, and I am sure I am not in the minority for my age (middle-aged male), and yet be able to appreciate differences in DACs and oversampling filters. These differences can as easily be attributed to how the DACs' filters or filters used by software upsampling algorithms differ in their reconstruction of transients. These filters can still be flat to 20kHz and yet sound different. It is harder to tell upsampling filters apart by looking for tonal differences (at least when using different filters on the same DAC). Can you explain what you were hearing for and how you were able to tell them apart? 

 

 

 

Generally the two things I was listening for there were:

 

1) Does one sound slightly closer in than the other. When swapping between the two, particularly just after one of the chimes where there's sound reverberating a bit, does the perceived stage collapse slightly when swapping to one or the other.

 

2) The sound of the transient element itself. There's a slightly sharper attack for me on the steeper/more extended filter

[Zaphod Beeblebrox]

2 minutes ago, GoldenOne said:

Generally the two things I was listening for there were:

 

1) Does one sound slightly closer in than the other. When swapping between the two, particularly just after one of the chimes where there's sound reverberating a bit, does the perceived stage collapse slightly when swapping to one or the other.

 

2) The sound of the transient element itself. There's a slightly sharper attack for me on the steeper/more extended filter

That is good to know, depth and quality of transients can be explained by transient timing, so that still keeps them open as another possible explination.

 

[EMINENT]

1 hour ago, Zaphod Beeblebrox said:

Thanks for sharing this. You have gone through a lot of effort to make the ABX testing 'air-tight ', this would be a great foundation for future tests, kudos for going through this and congrats on your hearing!

 

Your being able to hear till 21kHz (which is an exception than a norm I hope) provides a simple explanation. But I wonder if you have compared different upsampling filters that have a cut-off frequency beyond 21kHz?

 

The reason I ask is I cannot hear past 15kHz, and I am sure I am not in the minority for my age (middle-aged male), and yet be able to appreciate differences in DACs and oversampling filters. These differences can as easily be attributed to how the DACs' filters or filters used by software upsampling algorithms differ in their reconstruction of transients. These filters can still be flat to 20kHz and yet sound different. It is harder to tell upsampling filters apart by looking for tonal differences (at least when using different filters on the same DAC). Can you explain what you were hearing for and how you were able to tell them apart? 

 

 

 

Hallo. 

 

I theorize that while not being audible, ultrasonics can influence the perception. Also, while not being apparent at normal listening volume, at higher volumes it is possible other frequencies can fluctuate in volume as well. For example, I tested one of Goldens test files and while faint, I could perceive 21k but not 20 or 22. While unscientific and uncontrolled, this was done a few times by randomly cycling the playback of the 3 samples and when heard each time I opened my eyes it was 21k.

If you ever get a chance, play tone generator in these frequencies while playing music and you should hear attenuation as I do. This to me has an affect on the audible.

[Zaphod Beeblebrox]

  

56 minutes ago, mpaulson540 said:

I wish I preferred the sound of the 256DSD: 64fs x 4 - processing time with 6.1.42: 6 minutes, 18 seconds

The 256fs x 1 sounds more dynamic and yet also very liquid.

I can dedicate the trashcan to conversion so I might try a large batch (500 tracks) and see how that goes. 

Thanks for the feedback.

 

If I remember right your DAC is a Denafrips Ares ||, have you tried to do 256 x 2 or 512 x 1? 

 

 

[mpaulson540]

41 minutes ago, Zaphod Beeblebrox said:

  

Thanks for the feedback.

 

If I remember right your DAC is a Denafrips Ares ||, have you tried to do 256 x 2 or 512 x 1? 

 

 

Yes, I do have an Ares II, 512 x 1 was even better than 256 x 1 but my license only lets me batch create 256 x 1.

I also created an epic 1024 x 1 version on the original PGGB DSD version - 9.5 hours!!! Even better!

Have to live and listen within my means and remaining earthly time. 

256 x 2 only took 12 minutes but I still preferred 256 x 1 (at 31 minutes).

Hopefully the trashcan MacPro doesn't blow up during the next week. 

[Zaphod Beeblebrox]

58 minutes ago, mpaulson540 said:

Yes, I do have an Ares II, 512 x 1 was even better than 256 x 1 but my license only lets me batch create 256 x 1.

I also created an epic 1024 x 1 version on the original PGGB DSD version - 9.5 hours!!! Even better!

Have to live and listen within my means and remaining earthly time. 

256 x 2 only took 12 minutes but I still preferred 256 x 1 (at 31 minutes).

Hopefully the trashcan MacPro doesn't blow up during the next week. 

Perfectly understand, thanks for the details.

[Zaphod Beeblebrox]

20 hours ago, EMINENT said:

at higher volumes it is possible other frequencies can fluctuate in volume as well. For example, I tested one of Goldens test files and while faint, I could perceive 21k but not 20 or 22. While unscientific and uncontrolled, this was done a few times by randomly cycling the playback of the 3 samples and when heard each time I opened my eyes it was 21k.

While I do not contest what you heard, can you rule out distortion as being a cause? With high volume it is a real possibility. This is still very high compared to the energy in typical recorded content at 21kHz.

[Apollo]

I am interested to test PGGB but I have by no means a computer that would be capable of best quality upsampling.

 

Is there any way to get 1 (or 2) songs upsampled in best quality DSD512 and DSD1024 , and have the resulting files shared in the cloud for a limited time for me to download?

 

It does not make sense to make a hugh investment upfront, without knowing if the results are amazing enough to choose the PGGB path.

[seeteeyou]

@dmance already shared several free samples in DSD512

 

https://audiowise-canada.myshopify.com/collections/pggb

[Adyc]

On 4/27/2024 at 8:36 AM, jelt2359 said:

More impressions from anyone?

Not in my case. I prefer PCM 16FS over DSD 256 on DAVE. PCM 16FS are much more majestic and dynamics.

[kennyb123]

41 minutes ago, Adyc said:

Not in my case. I prefer PCM 16FS over DSD 256 on DAVE. PCM 16FS are much more majestic and dynamics.

I think one needs to go to DSD512 in 1 stage to beat 16FS.  This was the case with my DAC and I’m aware others who came to the same conclusion.  Unfortunately with DAVE, one needs to run Windows to get to DSD512 as I understand it.

[jelt2359]

Yep. I preferred DSD 512x1 to DSD 256x1 on the Dave. The latter sounded veiled, soft, more DSD-like (which I didn’t like). 
 

DSD 512x1 is a different story entirely. Frankly, it doesn’t sound like DSD (in a good way!)

[kennyb123]

1 hour ago, jelt2359 said:

DSD 512x1 is a different story entirely. Frankly, it doesn’t sound like DSD (in a good way!)

It’s been my impression that PGGB moves us to the point where we can no longer tell which format we are hearing (but only when one gets to at least 16FS or DSD512x1).  

[Zaphod Beeblebrox]

Version 6.1.44: The DSD Edition (PGGB Plus)

 

I have released an experimental feature for this update. You can grab v6.1.44 here. This is to address bottlenecks due to read speeds when using a single NVME drive for paging. NVME reads happen during the modulation stage where PGGB reads reconstructed data from page files (when they do not fit in RAM). The experimental feature tries to reduce the number of reads by processing longer buffers.

 

There are tradeoffs to this approach, so it is best to check with and without this option. A good indicator that the optimization may work for you is when you have a single NVME drive and 1fS -> DSD conversions seem to take significantly longer compared to 2fS -> DSD conversions.

 

Release Notes:

Version 6.1.44: The DSD Edition (PGGB Plus)
    SQ Changes:    None, same as version 6.0.42
    Features:
        Adds experimental tuning to optimize NVME reads.

 

[Mista Lova Lova]

1 hour ago, Zaphod Beeblebrox said:

There are tradeoffs to this approach, so it is best to check with and without this option.

 

Are there any sound quality trade-offs or would they be only in relation to how the processing is done (e.g. depending on what OS is used)?

[Zaphod Beeblebrox]

1 hour ago, Zaphod Beeblebrox said:

 

Release Notes:

Version 6.1.44: The DSD Edition (PGGB Plus)
    SQ Changes:    None, same as version 6.0.42

 

4 minutes ago, Mista Lova Lova said:

Are there any sound quality trade-offs or would they be only in relation to how the processing is done (e.g. depending on what OS is used)?

No, if that was the case, I would always indicate in release notes.  Tradeoff is more data is kept in memory for processing and there are longer reads.