HQ Player

[AudioDoctor]

Which part of the equation would a more powerful GPU help with?  I have a 5950x in my desktop computer and while working on some projects I have seen 2 cores boosting to 5.05ghz while all 16 cores/32 threads are being utilized. My CAPS has a 10900K in it now with a 2080 Ti GPU, how much of a difference would a 3090 make?

[Account Closed]

3 minutes ago, AudioDoctor said:

Which part of the equation would a more powerful GPU help with?  I have a 5950x in my desktop computer and while working on some projects I have seen 2 cores boosting to 5.05ghz while all 16 cores/32 threads are being utilized. My CAPS has a 10900K in it now with a 2080 Ti GPU, how much of a difference would a 3090 make?

There should be no filter that you could not run. Won’t help with going above DSD 256EC as that is a function of core speed. 8 GHz on two cores would be needed for DSD 512EC.

[AudioDoctor]

1 hour ago, bobflood said:

There should be no filter that you could not run. Won’t help with going above DSD 256EC as that is a function of core speed. 8 GHz on two cores would be needed for DSD 512EC.

 

With the 3090 or what I have now?

[SwissBear]

14 hours ago, Miska said:

 

None, everything completely smooth. NAA is UpBoard running my NAA image and DAC is Holo Audio Spring.

 

I was able to replicate this behavior and isolate the source of the cracklings in my installation.

The fact that we have an "off-the-shelf" solution for less than 1'000 USD including NAA, to run DSD 256 in stereo with the best filters/modulators available, seems to me a matter of celebration.

Thanks for this native ARM 64 code :-)

[Luca72c]

5 hours ago, k6davis said:

Really? I use the i7-9700K, which was released just over 2 years ago. It does DSD256 ADSM7EC x48 (poly-sinc-ext2, no convolution, 2ch) with any file format. If CPU performance had increased by 4x, wouldn't we be at DSD512 EC by now?

 

I counted 9-series Intel chips in this side of the (rough) 2 years border and surely they were a big performance improvement from previous 8-series generation, providing up to 8x simultanous 5ghz cores for the first time.

But no, a 4x total performance increase doesn't necessarily mean a 4x single thread performance increase, and unfortunately single thread performance is what counts most in HQP calculations (if it wasn't so, we would already have PCM -> DSD512 EC capable PCs). So in HQP performance we didn't see a 4x increase.

But i have to point out that M1 didn't throw in the game a significant single performance increase over x64 bests, either...

 

5 hours ago, k6davis said:

Since I already have DSD256 EC, I'm only interested in upgrading (eventually) to a chip that can go beyond that. The x86 chips are unable to split the modulator workload to multiple cores per channel, so additional cores don't help with my use case. Jussi has said that x86 will need 2x clock speed to go from DSD256 EC to DSD512 EC, so clock speeds like 5.5 Ghz won't help me. Intel & AMD will eventually get there, but I'm not aware of anything on their roadmap that will improve my simple HQP use case.

 

So you should not be interested in Apple M1 too, as it doesn't go beyond that anyway. If Apple, AMD or Intel next gen cpus will, we don't know. 5,5 ghz may not be enough to run PCM -> DSD512, but they're much closer than 3,2 ghz from an Apple M1 anyway... And adding more cores is exactly what Apple is going to do, as far as we know.

Surely x86/x64 architecture is an old school approach. But RISC architecture is from the 70s and ARM from the 80s, so they are not last minute tech too.

We have seen RISC architecture raise and (mainly) fall many times from then on, benefits and limits are well known and AFAIK there's no reason to think RISC chips can use multiple cores to perform modulators' calculations as the difference about CISC and RISC architectures is not about that. But maybe firmware/OS cores management and optimization could do, at least to some extent.

About ARM, please remember (as someone pointed out yet) that this is an architecture used in mobiles (Iphones and others), SOCs/SBCs (Raspberry PI and the likes), game consoles (Gameboy and others) and generally lower-than-PC performance and cost devices: so i can't understand all this hype about ARM architecture, except maybe about cost-per-performance or watt-per-performance improvements (and i think that's what have pushed cpu producers towards ARM). That should lead to an increase in the number of cores/decrease of total costs, not to an increase in single core performance (that is what counts here).

 

5 hours ago, k6davis said:

Apple Silicon may not be the solution either, but this is what's exciting to me. Jussi's test today showed that on the ARM chip, the modulator effectively used more than 1 core per channel. This is new. All 4 high performance cores in the chip were utilized. Will forthcoming Apple Silicon chips with 8 or 16 high performance cores offer any performance improvement? Is it possible that more than 2 cores per channel could be utilized? Perhaps not, but I'm looking forward to finding out. 

 

Jussi already answered about that, as you can see below:

 

13 hours ago, Miska said:

Adding more cores won't help for modulators. Modulators can run either on one or two cores per channel.

If they cut at all on the clock speeds or single core performance to trade for more cores, HQPlayer performance will suffer. So the high performance core clocks need to be same or more, with possibly additional cores for things to improve.

Overall, this is very complex topic without straightforward answer. We can already see this from Intel vs AMD. Cache sizes and bus architecture (latencies), etc matter a lot.

 

That speaks about "needed" 8 ghz clock speed too: that was an approximate assessment, it's not this simple, you have to take into account other factors too, like IPC, cache management, latencies, etc... that can matter a lot. Surely M1 brings very good engineering and optimization in, but at present that doesn't seem to be enough to make a sufficient performance jump.

I guess the cross-core improvements in M1 Jussi talked about were due to cache/UMA management, they're very welcome but that doesn't automatically mean that modulators calculations can be performed by multiple cores instead of only 1 or 2 cores per channel. 2 cores x 2 channels = 4 cores, so nothing to be amazed here.
But as Jussi says, this is a very complex topic without straightforward answer, so we'll see and maybe i'll be surprised next months. That could be by Apple or by AMD/Intel either... So let's hope, but don't pin yourself down on Apple too much.

 

About prices, sure M1 MBs are a neat and welcome detour in Apple high prices policies, but in broad PC world that is not a matter to celebrate so much: today i can already build a desktop (non-Mac) PC that can perform PCM -> DSD256 with the best filters/modulators available for about 1000 USD or even less, too, and add in multichannel capabilities that M1 doesn't have. Surely ARM architecture is very promising rough cost wise, but about cost/performance we still have to see how much really this Silicon architecture can better what general PC market is offering.

 

 

[k6davis]

1 hour ago, Luca72c said:

5,5 ghz may not be enough to run PCM -> DSD512, but they're much closer than 3,2 ghz from an Apple M1 anyway... And adding more cores is exactly what Apple is going to do, as far as we know.

 

If I understand you, you're saying that the x86 chips are significantly more powerful that the ARM chips. That is true. But most of the power of the x86 chips can't be used for HQP DSP. In the M1, you get (2) 3.2 Ghz cores working together per channel, which is (apparently) why it's capable of DSD256 EC. A 3.2 Ghz x86 machine would not be capable of DSD256 EC. So, for running HQP, the M1 already outperforms an x86 machine with the same clock speed.  

 

My basic point is that the improvements that Intel & AMD have made to their chips over the last couple of years have not been beneficial for my use case in HQP. The same can be said about the chips they are promising for 2021. 

 

On the other hand, Apple has exceeded expectations (in HQP and lots of other use cases) with this introductory Apple Silicon product, which will (naturally) go on to be the lowest performing "M" chip they ever make. The next "M" chip is scheduled to be released in a few months and several still more powerful "M" chips are on the roadmap for 2021 and beyond.

 

If, and it's a very big "if", Apple Silicon can effectively scale its multicore performance in HQP DSP beyond 2 cores per channel, that alone will make it the most effective HQP chip. Even if it can't scale, Apple may be able to create a desktop-oriented 5 Ghz "M" chip (which could theoretically use 2 cores per channel) before Intel/AMD are able to make an 8 Ghz or 10 Ghz chip (using 1 core per channel). 

 

It's all speculative. I'll happily open my wallet for whichever company can deliver the results I'm looking for. In the meantime, I'll continue to enjoy the great DSD256 EC performance I'm getting from my Intel chip. 

[Luca72c]

1 hour ago, k6davis said:

If I understand you, you're saying that the x86 chips are significantly more powerful that the ARM chips. That is true. But most of the power of the x86 chips can't be used for HQP DSP. In the M1, you get (2) 3.2 Ghz cores working together per channel, which is (apparently) why it's capable of DSD256 EC. A 3.2 Ghz x86 machine would not be capable of DSD256 EC. So, for running HQP, the M1 already outperforms an x86 machine with the same clock speed. 

 

That probably is the effect of cache and UMA management M1 introduces and, as i said, is absolutely welcome.

But things are not as easy as you say, because:

- Newer 4 ghz x64 cpus are capable of DSD256 EC too, so not very far. Newer i5 (and probably Ryzen 5) cpus are capable to do so still retaining some spare power for convolution, and even some newer i3 machines can do DSD256 EC AFAIK. Outperforming an x64 machine at same frequency is absolutely normal if you increase IPC and do cache management optimizations and that's exactly what AMD and Intel are doing in next cpus too. That's not necessarily related to being x64 or ARM, or to other main architecture differencies. To prove that, AMD Zen3 chips clearly outperform Intel chips with same or higher frequencies even in single thread loads, but they're still x64 chips (we'll soon see if this applies to HQP too).

- We don't know if next Silicon chips' power could be used for HQP DSP as well, based on M1 actual performances.

- We don't know (but it is highly dubious in my opinion) if MORE than 2 cores can be used together per channel using Silicon chips and EC modulators. If this can't be done, to have DSD512 EC on Silicon chips we would need >6 ghz frequency, that is not on the horizon so far.

- We don't know if Apple is interested in going near 5 ghz in next Silicon releases nor if this CAN be done, considering architecture differences. Increasing frequencies would mean increasing heat generation, power consumption, chips binning,  and so on. ARM is not the obvious choice to have highest frequencies and it seems that ARM choice is not much focused on that, as M1 frequencies (well lower than some x64 notebook chips) clearly show.

 

1 hour ago, k6davis said:

On the other hand, Apple has exceeded expectations (in HQP and lots of other use cases) with this introductory Apple Silicon product, which will (naturally) go on to be the lowest performing "M" chip they ever make. The next "M" chip is scheduled to be released in a few months and several still more powerful "M" chips are on the roadmap for 2021 and beyond.

 

Maybe it exceeded your expectations, but surely not mine. It has exactly the kind of performance that were to be expected, i.e. comparable performances to the best Intel/AMD newer notebook chips, as Jussi pointed out too.

I agree that higher perfomances are to be expected in next Silicon chips, but we'll have to see if this performance increase will go in HQP needs direction (single thread performance increase and/or modulators' calculations spread on multiple cores) or not so much (core count increase + optimizations). Actual infos are not very promising to be honest, but we'll see... Surely i'm not holding my breath.

But i think it's unreasonable and maybe disturbing to make this thread much longer by speculating on things we simply cannot know now, so i suggest to calm down fancies and dreams and drop this tennis-like discussion about Apple Slicon family outlook, to focus most on what we have NOW.

[k6davis]

12 minutes ago, Luca72c said:

But i think it's unreasonable and maybe disturbing to make this thread much longer by speculating on things we simply cannot know now, so i suggest to calm down fancies and dreams and drop this tennis-like discussion about Apple Slicon family outlook.

 

You find a discussion about CPU tech disturbing? That's odd.

 

I'll be sure to send you a P.M. before my next post to makes sure you're okay with it first.

[Luca72c]

14 minutes ago, k6davis said:

 

You find a discussion about CPU tech disturbing? That's odd.

 

I'll be sure to send you a P.M. before my next post to makes sure you're okay with it first.

 

Not much disturbing for me, otherwise i wouldn't have posted my thoughts. But i fear other readers could be not so much interested in long fancy speculations about future chips instead of more useful discussions about what is NOW available, HQPlayer applications use in primis and CURRENTLY AVAILABLE chips performance too. They have to be respected, of course (me too...). 

So that was my suggestion, you're free to take it into consideration or not.

 

P.S. I'm able to do stupid sarcasm too, but i don't do that as i think that's not the best way to talk to an unknown distant person...

[StreamFidelity]

@k6davis A suggestion 😉

 

 

[Zauurx]

Another ... a simple i5, 2 cores ... Redbook limited to DSD 128 with EC .. DSD256 without EC.

Roon + HQP (+ HF convolution) under Win10 LTSC with the right optimizations
My RME (256 max) is happy.

 

 

[Account Closed]

7 hours ago, AudioDoctor said:

 

With the 3090 or what I have now?

I do not have any personal experience as I don’t use a GPU but other people here have reported that the 2080 you have can’t run some filters but the 3090 should run all of them.

[ted_b]

12 hours ago, bobflood said:

DSD 1024 me thinks. Seriously, it’s pretty cool that we now have a $700 solution for DSD 256/ASDM7EC .

I'm beginning to think about the M1 as a stand alone 2 channel server (I also do lots of multichannel and would maybe use a heavy Windows server) .  Do I understand correctly, though, that Mac core audio is limited to DSD256, so a non-EC modulated DSD512 is not possible with the M1?  I have not heard EC modulators yet, so this may, in practice, be a moot point if 256 with EC is way superior to anything non-EC at 512 anyway.

[The Computer Audiophile]

Just fired up my wife’s M1 MacBook Pro to test HQP 4.8. DSD256 using ASDM7EC was a no-go. 

[Plato65]

This is very interesting! So what is the M1 Mac Mini not able to do? Convolution, filters? Is 16GB RAM required or is 8GB enough?

 

[SwissBear]

33 minutes ago, The Computer Audiophile said:

Just fired up my wife’s M1 MacBook Pro to test HQP 4.8. DSD256 using ASDM7EC was a no-go. 

Did you check this box ?

 

See: 

 

[SwissBear]

55 minutes ago, ted_b said:

I'm beginning to think about the M1 as a stand alone 2 channel server (I also do lots of multichannel and would maybe use a heavy Windows server) .  Do I understand correctly, though, that Mac core audio is limited to DSD256, so a non-EC modulated DSD512 is not possible with the M1?  I have not heard EC modulators yet, so this may, in practice, be a moot point if 256 with EC is way superior to anything non-EC at 512 anyway.

Mac's core audio seems limited to DoP (DSD over PCM). To enjoy full DSD, you would need an NAA. 

Miska mentioned he uses one of these: https://up-board.org/up/specifications/ with his NAA distro if I'm not mistaken. I am using an old Mac Mini with a Linux distro on it and the NAA software of Miska.

[ted_b]

20 minutes ago, SwissBear said:

Mac's core audio seems limited to DoP (DSD over PCM). To enjoy full DSD, you would need an NAA. 

Miska mentioned he uses one of these: https://up-board.org/up/specifications/ with his NAA distro if I'm not mistaken. I am using an old Mac Mini with a Linux distro on it and the NAA software of Miska.

I use an NAA but not sure how that would get around the server’s OS limitations for upsampling. NAA doesn’t do that work

[Kirbyt]

1 hour ago, The Computer Audiophile said:

Just fired up my wife’s M1 MacBook Pro to test HQP 4.8. DSD256 using ASDM7EC was a no-go. 

Are you able to do asdm5ec?

[SwissBear]

2 hours ago, ted_b said:

I use an NAA but not sure how that would get around the server’s OS limitations for upsampling. NAA doesn’t do that work

The limitation is not about the OS preventing upsampling. It's about using the Core Audio layers of the OS limiting the way DSD are handled through USB ports. When you use an NAA, you are driving it through the network (TCP/IP), and therefore not through the Core Audio layers of the OS. At least this is my understanding.

AFAIC, it works very well up to DSD 512 with Roon and DSD 256 with HQPlayer and EC modulators.

[bogi]

3 hours ago, ted_b said:

if 256 with EC is way superior to anything non-EC at 512 anyway

 

That seem to be the case. Even 128 IMO. Others may have different meaning.

[bogi]

2 hours ago, SwissBear said:

Mac's core audio seems limited to DoP (DSD over PCM). To enjoy full DSD, you would need an NAA.

 

Yes, DoP has DSD256 limit, for DSD512 the 'native' mode is needed.

[ted_b]

17 minutes ago, bogi said:

 

Yes, DoP has DSD256 limit, for DSD512 the 'native' mode is needed.

I fully understood that, but are you guys saying Native mode, even from a Mac server, can be obtained by using an NAA?  How?  The NAA is just a fifo buffer.

[bogi]

1 hour ago, ted_b said:

I fully understood that, but are you guys saying Native mode, even from a Mac server, can be obtained by using an NAA?  How?  The NAA is just a fifo buffer.

Simple ... NAA is running other OS than Mac.

NAA is FIFO buffer, but it uses it's own audio interface(s). So only the environent of NAA OS is relevant for these audio interfaces. The audio interface(s) NAA has are seen by HQPlayer through NAA protocol. HQPlayer has no problem to deliver DSD512 to any NAA - no matter which OS NAA is running - if NAA OS is able to play DSD512 to attached device and if your HQPlayer computer is able to play or upsample to DSD512. Since HQPlayer is playing into NAA through LAN, no audio interface restrictions apply to HQPLayer computer in this case. All audio interface stuff is then relevant only for NAA.

[ted_b]

2 hours ago, bogi said:

- if NAA OS is able to play DSD512 to attached device and if your HQPlayer computer is able to play or upsample to DSD512. Since HQPlayer is playing into NAA through LAN, no audio interface restrictions apply to HQPLayer computer in this case. All audio interface stuff is then relevant only for NAA.

Wow, I have been using HQPlayer and NAA for many years (as most of you know; it's all over my reviews, etc) but never realized that Mac can upsample to DSD512 if it is connected to an NAA (say Linux or Windows ASIO).  Learn something new every day...thx.   😀  So, even though I am still currently using HQplayer 3 in my secondary system (my main system is still boxed up waiting for room) and therefore upsample to DSD512 with non EC modulators for best sq, I could easily do this on a Mac Mini if my NAA is Linux or Windows (ASIO driver) into my Holo?  Thx  👍