Best CPU for hqplayer

[jabbr]

11 hours ago, k6davis said:

 

People on this forum, including the HQP developer, have reported that they're getting DSD 512 EC with the 12th gen Intel chip alone - and with the 11th gen Intel chip combined with a powerful GPU. 

 

Apparently, the logical assumption that twice as much clock speed would be needed to go from DSD 256 to DSD 512 was incorrect. Based on what's been said here, improvements in other areas of the design of the chip have us there today. 

 

Ok, this is a very new development!

[LoryWiv]

5 hours ago, Miska said:

I've spent a week now working on adding specific support for Intel's 12th Gen CPUs. The whole core allocation needed rework. It should benefit also machines with two or more sockets/dies.

That's GREAT to hear @Miska. I just ordered a PC build based on an i9-12900 / Nvidia RTX 3080 and am really looking forward to better success using high resource modulators and filters. Thank you for your continued efforts to keep HQ Player at the top of the pack!

[Luca72c]

6 hours ago, Miska said:

 

It has got better, because the overhead for communicating between CPU cores has gone down on x64 so modulators can split to two cores instead of one in a way that it brings enough benefits. This was the case first on Apple's M1, now Intel is trying to catch up on that area as well. Now I need to do more testing if I can squeeze similar performance out of AMD too.

 

I've spent a week now working on adding specific support for Intel's 12th Gen CPUs. The whole core allocation needed rework. It should benefit also machines with two or more sockets/dies.

 

 

Great! This kind of work can be long and complex but it is very welcome! 👍

[pis99]

19 hours ago, Miska said:

 

You run out of GPU RAM? Or it fails some other way?

 

I am not sure what's wrong but I can not even play 5ECv2 Sinc-L at 48x512. The GPU looks fine as well as CPU. But the network seems to be at question because at times after tried out Since-L, my HQP no longer sees NAA. I had to reboot several times to get it back to normal. No problem at all when play 7ECv2/ext3 at 48x512. 

[pis99]

9 minutes ago, pis99 said:

I am not sure what's wrong but I can not even play 5ECv2 Sinc-L at 48x512. The GPU looks fine as well as CPU. But the network seems to be at question because at times after tried out Since-L, my HQP no longer sees NAA. I had to reboot several times to get it back to normal. No problem at all when play 7ECv2/ext3 at 48x512. 

I also found setting I can play before the resale of ECv2 is no longer doable at current version. For example, I can do 7EC/Sinc-L at 48x256 on previous version but not with the latest HQP. 

[Luca72c]

3 hours ago, pis99 said:

I also found setting I can play before the resale of ECv2 is no longer doable at current version. For example, I can do 7EC/Sinc-L at 48x256 on previous version but not with the latest HQP. 

Yes, this is the problem with hqp 4.14 on windows. Jussi is trying to do what he can to optimize the new compiler for our platform, but that's not easy

[pis99]

9 minutes ago, Luca72c said:

Yes, this is the problem with hqp 4.14 on windows. Jussi is trying to do what he can to optimize the new compiler for our platform, but that's not easy

Thank you! I am very happy with what I hear right now. Just what to share findings in hope to help Jussi solving problems. Great job, no complaint!

[Miska]

Now I have results with AMD Ryzen 7 5800X too. poly-sinc-gauss-xla + ASDM7ECv2 from RedBook.

 

Multicore DSP fully enabled with my new core allocation code (for next release).

 

to DSD256:

 

 

to DSD512:

 

DSD1024 is too much to ask from this machine, but it is not so far, since peak load is 63% meaning that DSD1024 would be about 125%. Highest core clocks being at 4.5 GHz in DSD512 case. With CUDA offload, turbos could maybe reach 5 GHz which could be enough (or if the bottleneck is RAM, then DDR5 could help enough).

 

[k6davis]

7 hours ago, Miska said:

DSD1024 is too much to ask from this machine, but it is not so far, since peak load is 63% meaning that DSD1024 would be about 125%. Highest core clocks being at 4.5 GHz in DSD512 case. With CUDA offload, turbos could maybe reach 5 GHz which could be enough (or if the bottleneck is RAM, then DDR5 could help enough).

 

It is great progress that you're able to split the workload to more than 1 core per channel with the new generation of chips.

 

You mentioned that your testing was done with Redbook. Are you having any success, with any chip, with 24-192 => DSD 512 ECv2 x 48?

[pis99]

Folks, can I get help on a NAA question?

 

My Window desktop HQP always showed " NAA QoS not available" when started. My question is how do I solved this issue? I use the recommended up-gws01 hardware with NAA image. Music plays pretty good with this warning but I do not know anything about the up squared NAA. Any help will be appreciated.

 

 

[stefano_mbp]

35 minutes ago, pis99 said:

NAA QoS not available

… you can read about it here

 

[chipvn]

@Miska With your new core allocation code, will it be any benefit of 5800x over 5900x? Understand that the 5800x is based on one 8-core chiplet while 5900x is on dual 6-core chiplets, there might be cross chiplet latency penalty on the 5900x although its number of core, number of L3 cache, max turbo frequency are all higher than the 5800x.

 

Thanks

[Miska]

11 hours ago, k6davis said:

You mentioned that your testing was done with Redbook. Are you having any success, with any chip, with 24-192 => DSD 512 ECv2 x 48?

 

Seems to work without issues on the same AMD machine:

[Miska]

15 minutes ago, chipvn said:

@Miska With your new core allocation code, will it be any benefit of 5800x over 5900x? Understand that the 5800x is based on one 8-core chiplet while 5900x is on dual 6-core chiplets, there might be cross chiplet latency penalty on the 5900x although its number of core, number of L3 cache, max turbo frequency are all higher than the 5800x.

 

I don't have one to test with, but it should work equally fine, since different channels end up on different chiplets, so there's not much communication needed. I picked 5800X likely based on base clock which roughly represents achievable clock rate under full load condition.

 

[chipvn]

1 hour ago, Miska said:

 

I don't have one to test with, but it should work equally fine, since different channels end up on different chiplets, so there's not much communication needed. I picked 5800X likely based on base clock which roughly represents achievable clock rate under full load condition.

 

So does it mean HQPlayer could split different channels to different chiplets and assign the same channel related tasks to cores allocate on the same chiplets?

[Miska]

3 hours ago, chipvn said:

So does it mean HQPlayer could split different channels to different chiplets and assign the same channel related tasks to cores allocate on the same chiplets?

 

Yes, that's the idea.

 

This whole thing is very complex, because

• You have chips with number of CCX (AMD has 2, 3 and 4 core CCXs)
• You have chips with number of dies (multiple chips/chiplets in one package)
• You have chips with different types of processors, performance cores and efficiency cores
• Processors with and without threading
• You can have multiple sockets
• You can have NUMA systems where different portions of RAM are behind different packages

 

[Carlos269]

Question specifically pertaining to HQPLAYER CUDA Offloading:

 

I notice that the Effective Memory Clock for all NVIDIA 3090 GPU cards is 19500 MHz regardless of the specified “GPU Boost Clock” speed.

 

My question is, is there any benefit for CUDA Offloading on HQPLAYER in getting a GTX 3090 card with a bost clock of 1920 MHz versus getting a card with a boos clock of 1695 MHz? In other words, does the GPU boost clock rate matter?

[Miska]

10 hours ago, Carlos269 said:

My question is, is there any benefit for CUDA Offloading on HQPLAYER in getting a GTX 3090 card with a bost clock of 1920 MHz versus getting a card with a boos clock of 1695 MHz? In other words, does the GPU boost clock rate matter?

 

What matters is number of CUDA cores combined with clock rate. You get sort of performance factor number if you multiply clock rate with number of CUDA cores. You can then use this figure to compare GPUs within the same architecture (generation).

 

[elan120]

Is ASUS ROG Maximus Z690 Apex a good choice for the new HQPlayer server build?  It is interesting that there is only 2 memory slots instead of typical 4 memory slots.

[Miska]

9 hours ago, elan120 said:

Is ASUS ROG Maximus Z690 Apex a good choice for the new HQPlayer server build?  It is interesting that there is only 2 memory slots instead of typical 4 memory slots.

 

Should be pretty nice, if you can also get hold of DDR5 RAM (at least here hard to get).

 

I guess reason for 2 slots is that DDR5 by itself has double width compared to DDR4.

 

[elan120]

4 minutes ago, Miska said:

 

Should be pretty nice, if you can also get hold of DDR5 RAM (at least here hard to get).

 

I guess reason for 2 slots is that DDR5 by itself has double width compared to DDR4.

 

Thank you for the confirmation.  Will using DDR4 be poor in performance?  Especially SQ?

[Miska]

2 minutes ago, elan120 said:

Thank you for the confirmation.  Will using DDR4 be poor in performance?  Especially SQ?

 

It won't affect sound quality. My 12900K system uses DDR4 RAM and works nice. But likely you can squeeze some extra performance out of the 12900K + HQPlayer by using DDR5. Which remains to be verified, but could be the case based on some benchmarks I've seen.

 

[elan120]

16 minutes ago, Miska said:

 

It won't affect sound quality. My 12900K system uses DDR4 RAM and works nice. But likely you can squeeze some extra performance out of the 12900K + HQPlayer by using DDR5. Which remains to be verified, but could be the case based on some benchmarks I've seen.

 

Thank you again, I will start with DDR4 and try DDR5 next year when they are more available.

[StreamFidelity]

8 hours ago, Miska said:

My 12900K system uses DDR4 RAM and works nice. But likely you can squeeze some extra performance out of the 12900K + HQPlayer by using DDR5.

 

I currently recommend DDR4 RAM for several reasons:

 

1. DDR5 are poorly available, expensive and
2. have too high latencies. One of the fastest is for e.g .: G.Skill Trident Z5 RGB DDR5-6000MHz CL36-36-36-76 1.30V 32GB (2x16GB) - CAS Latency CL: 12.00ns.
3. DDR4 are mature and have very low latencies, e.G .: G.Skill Ripjaws V DDR4-4000MHz CL14-15-15-35 1.55V 16GB (2x8GB) - CAS Latency CL: 7.00ns and
4. DDR4 still benefits from 2 IMC (Integrated Memory Controller):

 

Source: https://www.igorslab.de/ddr5-deep-dive-kingston-im-interview-zum-neuen-arbeitsspeicher-standard-und-beispiele-aus-der-praxis/2/

[LoryWiv]

I have a new 12900 PC with 32 GB DDR5 expected in 4 weeks or so and will report performance with HQ Player.

 

@StreamFidelity you are correct on multiple fronts. The DDR5 is pricey and availability is the rate limiting step in getting my build completed., expected late December. I went with 32 GB DDR5 rather than 64 GB of DDR4 at similar cost based on the recommendation of @Miska. I do wonder if the DDR5 clock speed increase may prove more helpful than the negative impact of higher latency. My prior PC's RAM was KINGSTON 32GB Kit (2 x 16GB) HyperX Fury DDR4 2133MHz, PC4-17000, CL14 and new build will have 32GB (2 x 16GB) FURY Beast DDR5 5200MHz, CL40. More to follow.