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

[rayon]

I'm so happy camper now. These 1024fs x 1 files sound amazing on May. Been able to process multiple albums already. Those performance updates really did change the game for me. Amazing job @Zaphod Beeblebrox and also special thanks to @austinpop for your support in the process!

 

The only downside I see in this is that I've started drooling p5800x after seeing those disk write cycles climbing up rapidly :)

[Sagittarius]

It seems to me that  PGGB development targeted taking a set of design choices to the extreme possible. Given that the results of this no-compromise approach have been established,  I am curious to know (given the file upsampling times being reported) if  the sound quality of filters which are still long but shorter than the entire audio file were compared during the development of PGGB to the sound quality of the filters which PGGB ended up using . An audio file contains sounds that are not continuous and usually not more than a few seconds long so I am wondering if it is essential to upsample/reconstruct the audio file in its entirety at once. There seems to be a consensus emerging that upsampling to DSD offers the highest sound quality from PGGB, which is also highly demanding of processing power and time. I think it could be worth it to explore whether upsampling / reconstructing parts of the audio file that are long enough (thinking out loud: maybe also overlapping to some degree ) can be made to approach or match the sound quality PGGB current approach offers by upsampling the entire file at once .  Re-processing an entire music library with each improvement in the filter design/algorithms will be quite challenging at the current file upsampling times and decreasing the computing load may also open the door to eventually getting the 128 bit  and 256 bit versions of PGGB to operate in real time, which would be ideal.  

[Zaphod Beeblebrox]

3 hours ago, Sagittarius said:

I am curious to know (given the file upsampling times being reported) if the sound quality of filters which are still long but shorter than the entire audio file were compared during the development of PGGB to the sound quality of the filters which PGGB ended up using .

Thank you for your interest. Yes, we have done this every time there was a significant change to PGGB. Here is some history:

 

When PGGB was first introduced, it was 64 bit only, it used windowed sinc filters and had options to choose the maximum length (taps in millions). While there may have been exceptions, typically the longest filter yielded the best quality. However, when one is memory constrained, there was a choice to go for shorter filters, which is better than just not processing. On some tracks (like long classical movements) that were artificially split into multiple tracks, we found that it was even beneficial to stitch them back into one track to make use of more taps. Of course this is not an automatic process, one needed to be familiar with the track and instruct PGGB to combine multiple tracks. PGGB continues to have this feature even now.

 

PGGB-AP introduced arbitrary precision computing and used precisions up to 256 bits for doing the same windowed sinc based filtering and also made use of better noise shaper to take best use of the higher precision.  This came with a significant computing burden. Here too we tested if the improved precision would allow us to use shorter filters. The answer here was the same as before, longer lengths still helped improve the final quality, so once again the choice was between quality and computational burden, you still had a choice to choose shorter filters.

 

PGGB 256 abandoned the using windowed sinc filters because even with the extreme lengths of the filter, there was still a compromise between time domain accuracy and frequency domain attenuation based on the window function. PGGB 256 used all the information in the track and abandoned the concept of taps. However, compared to PGGB-AP, PGGB 256 used highly optimized library and also had much better memory handling. So instead of offering a choice of filter lengths, the choice was to choose a precision that fit the hardware. I.e., 64bit vs 128bit vs 256. To put things in perspective, 64bit precision was fast enough to run in real-time so we made a free foobar plugin and it is more accurate than the original 64bit PGGB!, it continues to be available for free. 

 

The 64bit foobar plugin already has the option you are suggesting (though it does not support conversion to DSD). It allows you to split the track into smaller chunks to make the processing easier, so you get a choice between quality vs speed. It is easy for anyone to try and decide for themself.

 

3 hours ago, Sagittarius said:

There seems to be a consensus emerging that upsampling to DSD offers the highest sound quality from PGGB, which is also highly demanding of processing power and time.

 

PGGB DSD builds on PGGB 256, it uses the same technique but instead of stopping at PCM rates, takes it to the extreme to DSD rates. But here too there are choices. PGGB DSD is most demanding if one chooses to do a single stage upsampling. So, if you wished to do DSD512, you can do single stage directly to 512 which is the most memory and processor intensive. But that is not the only option, you could as easily do DSD512 = 16 x 32, which is 32 times less demanding or 512 = 128 x 4 which is 4 times less demanding, etc. However, one of the weaknesses of us audiophiles is we cannot settle for less!

 

Take this hypothetical example:

Even though DSD512 = 64 x 8 may sound better than16fS PCM to someone, if they find DSD512 = 128 x 4 to sound even better, they will keep going till they hit the single stage upsampling. That is why most reports you may find is about the single stage upsampling and I am sure you can ask them, why not just do 128 x 4?  and I am quite sure the answer would be not because it is bad, but because 512 x 1 is better! 

 

Though PGGB has earned itself the name for needing extreme hardware, it is not any more extreme than choosing to invest in custom cooled PC setups with multi-core processors and dedicated GPU to run the highest DSD rates with the best modulators that will work in real-time when there is always the choice to choose to run what will comfortably run in their existing hardware. PGGB has always offered choices of quality vs performance to help you choose the best quality that will work on your PC/Mac.  PGGB DSD is no different it offers choices to fit one's hardware, it is up to everyone to decide how much they want to push and if the quality is worth it invest in the right hardware.

 

3 hours ago, Sagittarius said:

I think it could be worth it to explore whether upsampling / reconstructing parts of the audio file that are long enough (thinking out loud: maybe also overlapping to some degree ) can be made to approach or match the sound quality PGGB current approach offers by upsampling the entire file at once. 

 

Yes, as I mentioned before, this is an experiment we do every time there has been a significant improvement to PGGB (part wishful thinking on our side). With PGGB DSD too, we found that chopping the file, processing in chunks and then stitching them back degraded the sound quality, it relieves the memory pressure yes and helps with processing. But we found, just doing two stage DSD was less of a compromise in this regard and was faster. I.e., instead of chopping a file into two, you could do DSD512 = 256 x 2 and still get as good or better results and it will take less time too! I know it is good to have choices so one can find out for themselves, so PGGB-IT! has the option to chop a track, process them and pieces them and as you suggested (right thinking by the way), it already makes sure there is an overlap so that when they are stitched after processing, they are seamless.

 

3 hours ago, Sagittarius said:

An audio file contains sounds that are not continuous and usually not more than a few seconds long so I am wondering if it is essential to upsample/reconstruct the audio file in its entirety at once.

We have discussed this a few times in this thread, in depth knowledge on how a track was mastered is very rare, but on most occasions what we found is just using the whole track sounds best. Though I don't quite agree that they are only few seconds worth of continuous recording, it could be so for some specific genre but not for all music. Even if it were true, there is no way to know where exactly to split the tracks, so even a short filter will be reconstructing across unrelated segment.

 

3 hours ago, Sagittarius said:

Re-processing an entire music library with each improvement in the filter design/algorithms will be quite challenging at the current file upsampling times and decreasing the computing load may also open the door to eventually getting the 128 bit  and 256 bit versions of PGGB to operate in real time, which would be ideal.  

Some are already doing this, i.e., choosing 2 stage processing to process their entire library first and then doing a one stage processing while they are listening to what has already been processed. This is part of the fun, listening to the library but expreiencing it differently every time.

 

PGGB is constantly evolving, PGGB DSD was released only about a week back and if you saw the last update, it is already 2-3x faster! and we are continuing to make improvements to memory and paging, so it is only going to get faster and require even less resources. I am sure there will come a point where PGGB DSD is possible in real-time. A good example of that is how 64bit PGGB PCM needed significant hardware but now it can run in real-time without needing a GPU. With the optimization we did recently it is now possible to run PGGB PCM with up to 32 decimal digits of precision (very close to 128bit precision) in real-time which we hope to again release as free foobar plugin as our way of giving back to this community. We do not have an ETA yet, but hopefully before fall this year. 

 

[kennyb123]

3 hours ago, Sagittarius said:

There seems to be a consensus emerging that upsampling to DSD offers the highest sound quality from PGGB

ZB’s response covered the points exceptionally well, as usual.  I just wanted to mention one other thing.  I believe the results still depend on the DAC.  If DAC can only get to DSD256 then 16FS might still deliver the highest sound quality from PGGB.  How much processing the DAC does also factors into this.  

[Sagittarius]

Thank you very much ZB for your comprehensive reply. You actually also replied to some additional questions which I had in mind. 

[Sagittarius]

Yes, he definitely did. The point you mentioned is noted. I currently have a DSD 512 capable DAC (T+A DAC 8 DSD). Unfortunately, manufacturers keep chasing buzzwords and trying to get attention with shiny spec sheet figures for marketing reasons rather than commonsense. So we now have DSD 1024 capable DACs getting more common and hopefully they will stop at that. If I remember correctly,  T+A designer mentioned in one post that implementing that DSD rate in their DACs was not easy. Faster clocks also tend to have worse jitter. However, several posters noted that operating their DAC at its maximum DSD rate gave better sound quality (my personal guess is that this probably varies from one DAC model to another depending on its design). I think this improvement in sound quality was first noticed when the DAC 8 DSD became popular and its owners found that feeding it upsampled files with increasingly higher upsampling ratios gave progressively smaller improvements in sound quality till they hit DSD 512 which was the maximum rate the DAC was capable of and then there was a noticeable jump in sound quality. The best explanation I have read for this was the one given by Miska, who suggested that operating the DAC at the native rate of its clock without clock dividers could possibly be the reason. So it is probably a case of lower jitter more than a case of  upsampled file reconstruction getting audibly better all the way to DSD 1024. But as ZB mentioned, we find it difficult to settle for less when we hear what is better. So, regardless of the technical reasons, we will want to upsample to the highest data rate that the DAC we have or want to buy can handle. 

 

17 hours ago, kennyb123 said:

ZB’s response covered the points exceptionally well, as usual.  I just wanted to mention one other thing.  I believe the results still depend on the DAC.  If DAC can only get to DSD256 then 16FS might still deliver the highest sound quality from PGGB.  How much processing the DAC does also factors into this.  

 

[Zaphod Beeblebrox]

17 minutes ago, Sagittarius said:

Yes, he definitely did. The point you mentioned is noted. I currently have a DSD 512 capable DAC (T+A DAC 8 DSD). Unfortunately, manufacturers keep chasing buzzwords and trying to get attention with shiny spec sheet figures for marketing reasons rather than commonsense. So we now have DSD 1024 capable DACs getting more common and hopefully they will stop at that. If I remember correctly,  T+A designer mentioned in one post that implementing that DSD rate in their DACs was not easy. Faster clocks also tend to have worse jitter. However, several posters noted that operating a DAC at its maximum DSD rate gave better sound quality in several cases. I think this was first noticed when the DAC 8 DSD became popular and its owners found out that feeding it upsampled files with increasingly higher upsampling ratios gave progressively smaller improvements in sound quality till they hit DSD 512 which was the maximum rate the DAC was capable of. Then there was a noticeable jump in sound quality. The best explanation I have read for this was the one given by Miska, who suggested that operating the DAC at the native rate of its clock without clock dividers could possibly be the reason. So it is probably a case of lower jitter more than a case of  upsampled file reconstruction getting audibly better all the way to DSD 1024. But as ZB mentioned, we find it difficult to settle for less when we hear what is better. So, regardless of the technical reasons, we will want to upsample to the highest data rate that the DAC we have or like can handle. 

 

 

At least on T+A DAC  200, I would say the jury is still out on DSD1024.

 

One has to jump through hoops to get DSD1024 working, not many players support it and if they did the OS may not and then the DAC's driver or firmware needs to be updated etc.

 

On the Holo May, based on the feedback I have received, there seems to be a clear preference to using DSD1024. On T+A 200, the feedback I have received has been mixed, anywhere between DSD1024 sounding slightly worse than DSD512 to DSD1024 is clearly better. That does not make one a better DAC than the other, it just shows where the sweet spot for the DAC is.

 

These are still early days with very small sample sizes, as many have not had the opportunity to listen to DSD1024 on either of these DACs, hopefully there will be some form of consensus over time.

 

But I generally agree, there is no point in pushing for higher numbers unless there is an audible advantage.

 

[Zaphod Beeblebrox]

• Myths and Half Truths

  • Don't test outside of spec!

  • Three wrong reasons

    1. Rings like a bell (maybe for a bat)

    2. Blurs Transients (show us your math)

    3. Oversampling filters affect outcome only above audible band (why so naive?)

  • Did we pretend Gibbs phenomenon did not exist and defy physics?

 

For those who are interested, I have added examples with real music track to explain what PGGB does and does not do.

[ajm]

I have been trying out the effect of DSD512 (single pass) with Chord DAVE. I do not possess a behemoth PC for conversions as I had never envisaged performing these conversions having been rather happy with PCM. As my conversion machine is obviously underpowered in comparison with the keenest users, I am somewhat limited in my assessment - to convert my library with this system would take about a year.

On the basis of what I have tested and bearing in mind that I am primarily a classical listener and exclusively use speakers rather than headphones, I have found that with PCM base files, I do not prefer the effect of PGGB conversions to DSD512. With DSD base files, I find that those of highest resolution in my collection (DSD256) do seem to be preferable when converted to DSD512. The difference is quite striking to my ear in cases such as Honek's Pittsburgh Beethoven 5&8 and the Podger Vivaldi Quattro Stagioni. In these cases, I do prefer the DSD512 to the PGGB 705/768 PCM versions. To perform these comparisons, I have had to use a windows 11 NUC that is certainly not optimised for audio as an endpoint for HQ Player NAA with USB direct to DAVE. My usual path for PCM is HQPe to SRC-DX in a low powered audiolinux i7 NUC with cleanish power from an Uptone JS-2.

As a sidenote, I find that notwithstanding the advice from ZB, it is easier to control the conversion PC (using windows RD) when running Process Lasso and I have not found this causes problems.

 

[Zaphod Beeblebrox]

47 minutes ago, ajm said:

I have been trying out the effect of DSD512 (single pass) with Chord DAVE. I do not possess a behemoth PC for conversions as I had never envisaged performing these conversions having been rather happy with PCM. As my conversion machine is obviously underpowered in comparison with the keenest users, I am somewhat limited in my assessment - to convert my library with this system would take about a year.

On the basis of what I have tested and bearing in mind that I am primarily a classical listener and exclusively use speakers rather than headphones, I have found that with PCM base files, I do not prefer the effect of PGGB conversions to DSD512. With DSD base files, I find that those of highest resolution in my collection (DSD256) do seem to be preferable when converted to DSD512. The difference is quite striking to my ear in cases such as Honek's Pittsburgh Beethoven 5&8 and the Podger Vivaldi Quattro Stagioni. In these cases, I do prefer the DSD512 to the PGGB 705/768 PCM versions. To perform these comparisons, I have had to use a windows 11 NUC that is certainly not optimised for audio as an endpoint for HQ Player NAA with USB direct to DAVE. My usual path for PCM is HQPe to SRC-DX in a low powered audiolinux i7 NUC with cleanish power from an Uptone JS-2.

As a sidenote, I find that notwithstanding the advice from ZB, it is easier to control the conversion PC (using windows RD) when running Process Lasso and I have not found this causes problems.

 

Thanks, I assume you were using DSD+ and making sure the files were played as Direct SDM with no processing. 

[jelt2359]

6 hours ago, ajm said:

I have been trying out the effect of DSD512 (single pass) with Chord DAVE. I do not possess a behemoth PC for conversions as I had never envisaged performing these conversions having been rather happy with PCM. As my conversion machine is obviously underpowered in comparison with the keenest users, I am somewhat limited in my assessment - to convert my library with this system would take about a year.

On the basis of what I have tested and bearing in mind that I am primarily a classical listener and exclusively use speakers rather than headphones, I have found that with PCM base files, I do not prefer the effect of PGGB conversions to DSD512. With DSD base files, I find that those of highest resolution in my collection (DSD256) do seem to be preferable when converted to DSD512. The difference is quite striking to my ear in cases such as Honek's Pittsburgh Beethoven 5&8 and the Podger Vivaldi Quattro Stagioni. In these cases, I do prefer the DSD512 to the PGGB 705/768 PCM versions. To perform these comparisons, I have had to use a windows 11 NUC that is certainly not optimised for audio as an endpoint for HQ Player NAA with USB direct to DAVE. My usual path for PCM is HQPe to SRC-DX in a low powered audiolinux i7 NUC with cleanish power from an Uptone JS-2.

As a sidenote, I find that notwithstanding the advice from ZB, it is easier to control the conversion PC (using windows RD) when running Process Lasso and I have not found this causes problems.

 

9th order too I assume. And remember to check dsd+
 

What we are discovering is that the non-decimating Dave is actually an ideal DSD Dac as long as the upsampling is done outside of the Dave. Same situation to what Rob is doing with his Mscaler and Quad scaler. Quite remarkable.
 

Background blackness and noise floor are probably most obvious with headphones.
 

Dynamics more so with speakers.
 

Also I do not listen to classical at all, so it’s not surprising to me that you and I may have different preferences. 

 

Thank you for the feedback!

[LowOrbit]

The question of which DSD output rate gives "best" results with the T+A Dac200 is one I am currently exploring. I am doing this in a casual manner and hoping to form an opinion over time as otherwise the mental effort of trying to willfully compare and contrast based on specific "testing" tends to muddy the listening experience and force the conscious brain into a critical mode which isn't necessarily the best way to evaluate the rather subtle differences between DSD512 & DSD1024.

 

I am hoping that playing a variety of music over an extended period without forcing a conscious evaluation will yield a more robust view to form. Both processes yield excellent results and consciously I am trying not to think about which I am listening to, so converting a wide variety of music to both formats and enjoying the music may lead me to prefer one over the other. I don't want technical factors or logistical factors (time to process files at the various rates, storage consumption rates etc) to drive a bias.

 

So, early in the game I am happy with either DSD512 x 1 or DSD1024 x 1 (but also fine with DSD1024 (512 x 2)) and could happily live with either option if that were the only option.

 

 

[Mista Lova Lova]

Just highlighting the fact that @Zaphod Beeblebrox has asked @Miska a question which has not been answered.

[The Computer Audiophile]

Clean up done, posts moved to the PGGB and HQPlayer thread. 

 

 

[jelt2359]

One thing I appreciate about ZB is- like some of my favourite designers (eg. Papa Pass)- that he doesn't simply listen with their brains, but with their ears, too. He has taken in a lot of feedback about what sounds better or worse (eg. -600db is audible, surprise surprise), rather than assuming what should be correct. (PS, this doesn't mean he designs purely based on subjective feedback, but that he uses that to validate his theories- a page right out of the scientific method).

 

Unfortunately not everyone will have the chance to try PGGB because it's a unique product suited mainly for those who play primarily their own offline files + ideally a powerful machine (eg. gaming machines, AI machines, rendering machines, upsampling machines...)

 

But for those who do meet those two criteria, I encourage you to listen with your ears as well. It's free.

 

As another poster here noted, PGGB is a sampling approach taken to its most absolute extreme end. You may not like it, but if you do, you probably won't listen to anything else ever again.

[Zaphod Beeblebrox]

@The Computer AudiophileSorry for bothering you, please move both of the previous posts to the other thread. 

[Zaphod Beeblebrox]

1 hour ago, jelt2359 said:

Unfortunately not everyone will have the chance to try PGGB because it's a unique product suited mainly for those who play primarily their own offline files + ideally a powerful machine (eg. gaming machines, AI machines, rendering machines, upsampling machines...)

A minor correction, PGGB can run on simple laptops too, it just depends on what you want it to do. The Foobar plugin will run just fine on most PCs with 8 - 16GB of RAM. 

[ajm]

10 hours ago, jelt2359 said:

9th order too I assume. And remember to check dsd+
 

What we are discovering is that the non-decimating Dave is actually an ideal DSD Dac as long as the upsampling is done outside of the Dave. Same situation to what Rob is doing with his Mscaler and Quad scaler. Quite remarkable.
 

Background blackness and noise floor are probably most obvious with headphones.
 

Dynamics more so with speakers.
 

Also I do not listen to classical at all, so it’s not surprising to me that you and I may have different preferences. 

 

Thank you for the feedback!

I confirm use of 9th order and DSD+ in my testing.

I think there are many variables which make a minefield out of comparing experiences between different systems.

One respect in which my use case is pretty unusual, I believe, is using DAVE to drive speakers direct which I began after reading Romaz experiences long ago and I remain quite happy with this using B&W 805D3's with DB3 sub. I originally began this when I was expecting to proceed to the projected and then putatively imminent Chord digital amps but was so surprised at how good this arrangement was that I have not felt obliged to seek any alternative in the last few years. Another factor is the improvements in Audiolinux over the years which certainly advanced in quality of reproduction (to my ear) with the progress to 2000Hz kernels. Of course, one cannot readily compare what one now hears with all the latest tweaks and upgrades with what was available some years ago and just to compare using optimised playback systems for PCM and DSD is not straightforward as I gather even Taiko Extreme users diverge from the Taiko preferred USB driver for DSD512/ 1024.

[The Computer Audiophile]

8 minutes ago, Zaphod Beeblebrox said:

@The Computer AudiophileSorry for bothering you, please move both of the previous posts to the other thread. 

Not a bother at all. My life is made easy when I’m told exactly what to do / what people want done.

 

If I have to read through a thread, attempt to understand the context of each persons’ relationship with the others and how it relates to specific posts, it’ll take me a month to take action. 
 

Back to the topic at hand. 

[Zaphod Beeblebrox]

3 minutes ago, Crwilli57 said:

I am lazy.  I have tried but failed  to learn the answer to perhaps a fundamental question re: PGGB.

 

Do I run it on my entire library and thus create a new library or does it need a PC to run it in real time?

 

I use an Antipodes Kala K50 with ~4 TB of internal media.   
 

Does PGGN rewrite that media into a higher sample rate?

 

Sorry for perhaps an obvious question.

I am responding to you on the appropriate thread.

Yes, it is offline resampling, so you will need all your files that you want to be processed to be available on your hard disk (not streamed). 

 

Then process them in batches and create a new library (PCM or DSD).

 

You should be able to run them with your Antipodes server, so you do not need an additional PC to run in real time, but just make sure Antipodes can support the rate you want to play.

 

PGGB upsamples or downsamples to the rate your DAC will support. Please tell us more about your DAC or email me instead.

 

 

 

[Crwilli57]

3 minutes ago, Zaphod Beeblebrox said:

I am responding to you on the appropriate thread.

Yes, it is offline resampling, so you will need all your files that you want to be processed to be available on your hard disk (not streamed). 

 

Then process them in batches and create a new library (PCM or DSD).

 

You should be able to run them with your Antipodes server, so you do not need an additional PC to run in real time, but just make sure Antipodes can support the rate you want to play.

 

PGGB upsamples or downsamples to the rate your DAC will support. Please tell us more about your DAC or email me instead.

 

 

 

Thank you!  I will try a trial over this weekend.

[kennyb123]

15 minutes ago, Zaphod Beeblebrox said:

You should be able to run them with your Antipodes server, so you do not need an additional PC to run in real time, but just make sure Antipodes can support the rate you want to play.

@Crwilli57Squeeze on your K50 can handle up to 768K PCM and DSD512 as of AMS 5.0. 

[Crwilli57]

Hmmm, that will take my 4tbs of media to WELL over 10tbs.  I need to think about that before launching.    Will start slowly.

[Crwilli57]

36 minutes ago, kennyb123 said:

@Crwilli57Squeeze on your K50 can handle up to 768K PCM and DSD512 as of AMS 5.0. 

My AMS 5.0 does not load.  Shows ‘update available’ but never completes.  Waiting for Mark et al to fix whatever the issue is.  In the meantime, I think am limited to DSD128 via USB.

[austinpop]

3 minutes ago, Crwilli57 said:

My AMS 5.0 does not load.  Shows ‘update available’ but never completes.  Waiting for Mark et al to fix whatever the issue is.  In the meantime, I think am limited to DSD128 via USB.

 

What DAC do you have?