New Holo Audio Cyan DAC/AMP: PCM or DSD module?

Discussion in 'Dedicated Source Components' started by doraymon, Sep 28, 2017.
?

Would you be more interested in the PCM or DSD module for the Holo Audio Cyan DAC/AMP?

  1. PCM

  2. DSD

Results are only viewable after voting.
  1. Yeang
    @Energy Thanks for the post! I have been searching for a while to see if anyone has tried Threadripper with HQPlayer. I am in the process of planning a new PC for upsampling to DSD512 with xtr, and would love some more information regarding Threadripper's performance on DSD.

    I assume you have the 1950X? Is it overclocked or at base clock? How much CPU usage at DSD512 upsampling?

    Can threadripper do DSD512 upsampling across rate families without stuttering using poly-sinc xtr (non-2s/full version)? I.e. 24/96 -> DSD512 @ 11.2 Mhz and not 12.288Mhz? I know that Jussi has a 6950X which cannot do that without stuttering, but can only do within rate families.

    Let me know! Thanks!
    Yeang
     
  2. guymrob
    I suggest get a DSD model. Why? The are numbers of reasons:

    1. Converting DSD to PCM has high ultra-sonic noise and require steep filter around 30kHz - If you getting a PCM version
    2. Converting PCM to DSD does noise shaping and more gentle filtering above 50kHz - If you are getting a DSD version

    I believe Cyan uses AKM SRC to convert PCM to DSD on the fly, similar to Spring. Moreover, converting PCM especially 44.1kHz to DSD sounds very easy to the ears; less fatigue and more 'analogue' sounding.
     
    Last edited: Oct 10, 2017
  3. guymrob
    I'm not sure scaling up the number of cores up beyond a certain limit will help, you need to ask Jussi for advise. The thing here is, graphic card based on Nvidia 'Cuda' processing core can dramatically off load and make it more easier to up-sample to DSD512.
     
  4. Energy
    It’s very rare for people to have a capable computer for the poly-sinc-xtr filter.

    I am in the process of building a computer right now that will do the job. Just waiting on the parts to come in (arrives tomorrow).
    CPU: Intel i9 7900X (10 core/20 threads)
    GPU: NVIDIA GTX 1080 Ti
    PSU: EVGA SuperNOVA 1000P (1000W) Platinum
    RAM: G.Skill Trident Z DDR4 3000MHz
    SSD: Samsung 960 PRO 512GB M.2 (for operating system)

    I thought about going for Jussi’s i7-6950X (10 core) but since it was pushed to 90% workload just to encode poly-sinc-xtr at DSD512, I reckon it would be better to get one with a little more wiggle room just in case music stutters when minimizing apps within the user interface. I went for the i9-7900X (10 core) as it is the newer flagship model and should be pushed to no higher than 70% workload when encoding PCM to DSD. When overclocked I doubt it will be any higher than 60%. I did consider the cheaper option 1920X (12 core) or paying a little more for the 1950X (16 core), but since those CPU’s are more for workstations, FPS for games would suffer. If more intensive filters are released in the future, I will simply just upgrade to a higher core count Intel i9 as it goes up to 18.

    The Holo Audio Spring sounds out of this world when fed PCM music that’s been encoded to DSD512. I really enjoy poly-sinc-xtr-mp due to the majority of my music library being either modern pop music or EDM so the mp (minimum phase) really helps with transient response. Unfortunately I have to run the two stage (2s) otherwise my previous Intel 6700k would bottleneck at single stage (non-2s). I’m excited to get all the parts in for the build and to finally give it a listen. Even when I listen to poly-sinc-xtr-mp on DSD256, there is more focus, layering, detail, and accuracy, but without the expansiveness that DSD512 brings. This build should give me the best of both worlds.

    The ThreadRipper 1920X and 1950X should easily handle DSD512 at any rate and so will the i9 7900X (6950X replacement) from Intel that I got. And if it somehow still stutters, a simple overclock to 4.6GHz should fix the problem.

    DSD encoding is heavily dependent on CPU power thus why Jussi made it to where HQPlayer can utilize up to a total of 27 cores. Even with his Intel i7 6950X and GTX 1080 Ti, it still hovered around 90% workload just to upsample to DSD512 using the single stage (non-2s) poly-sinc-xtr filter. I personally have a GTX 1080 Ti as well.
     
    Last edited: Oct 11, 2017
  5. Yeang
    Yeah I have been in close communication with Jussi about this. My experience with gpu offload has been quite mixed with little success even with a titan black with has huge DP processing speeds.

    He says all cores will be used as the code was optimized for parallelization for CUDA offload.

    He does not know with 7900x or threadripper will work.



     
  6. Energy
    So he recommends CUDA OFFLOAD being on? I’ve never noticed it to help me much in the past but then again that was with an older computer and not using single stage xtreme filters.

    Hopefully the 7900X will work. If it’s still stutters I’ll switch it for the 7920X, 7940X, 7960X, or 7980XE. If there’s other issues with code I’ll just have to wait and downgrade to the 6950X in the meantime.
     
    Last edited: Oct 11, 2017
  7. Yeang
    He has not said one way or another, but I get the sense he feels that if you can get it all done with CPU only, the better.

    So you do not have a Threadripper? I thought from your first post in this thread you had one. Remember that Threadripper is the same price as the 7900X, for 6 more cores...

    I was hoping to get some confirmation from someone who has tried it that Threadripper will handle DSD512 for sure.
     
  8. bluesaint
    I'm looking for that info as well! on CA forum there's a thread on it, but no new updates :frowning2: For now, i'm sticking with my 6700 laptop and DSD256 XTR 2S-MP. I'm finding DSD256 XTR 2S better than DSD512 2S-MP

    And to keep on subject, I think i get a good sense the R2R DSD section will be fine, but don't see too much info on the AMP portion...
     
  9. Energy
    I think CPU by itself would be better too.
    I bought a ThreadRipper 1950X but returned it because they do not perform well in FPS games and loses 20-30 FPS when compared to Intel offerings, which is why I returned it and went with an Intel i9.

    This PC isn’t built just for the audiophile hobby so I felt it was not justifiable to spend $1000 on a CPU alone just for DSD Encoding.

    That’s interesting. Even with a 4th generation laptop I was able to play DSD256 in poly-sinc-xtr and poly-sinc-xtr-mp without the 2S. It stutters like crazy at DSD512 which ends up needing the 2S.

    I found the non-2S to sound better even on DSD256 however the sound stage takes a hit. I’m building my computer tonight and will test DSD512 with poly-sinc-xtr-mp tomorrow. Hopefully this CPU is fast enough. If not i’ll just overclock it or go for the 7920X.
     
    Last edited: Oct 11, 2017
  10. Energy
    The Intel i9 7900X did not work and i have doubts the AMD ThreadRipper 1950X would be any different.

    • With CUDA offload is disabled, all cores are being used at it's maximum frequency yet only 60% workload shows up. Songs exhibit more stuttering with CUDA offload disabled.
    • With CUDA offload is enabled, the entire filter workload gets sent to the GPU in which it hovers around 60-85% while the CPU hovers around 12%. HQPlayer will not use more than 4 cores at max frequency. Every other core remain at idle frequency.
    I managed to get POLY-SINC-XTR & DSD512 to work on certain songs that has lower bit rate (kbps) but I had to overclock both the CPU and GPU. This is with CUDA offload enabled.

    CPU (Intel i9 7900X)
    - 3.3GHz overclocked to 4.8GHz

    GPU (GTX 1080 Ti)
    - 1500MHz overclocked to 2050MHz [Core Clock]
    - 5000MHz overclocked to 5500MHz [Memory Clock]

    It’s a lot of work. Sometimes it will still skip on certain songs. At this point I'm not sure if it's a CPU utilization issue (due to new processor release) or a buffer related problem. Regardless of whether or not CUDA offload is enabled or disabled, both the CPU and GPU are never used to their maximum potential so I am thinking HQPlayer needs a patch to support the newer processors.

    While POLY-SINC-XTR on DSD512 sounds extremely good (pun intended), there may still be micro-stutters in my listening sessions that is degrading my overall judgement. At this point I can only see it getting better once this thing starts working. If I knew how important GPU processing was for the filter I would have gone with the newer 8th generation Intel CPU for cheaper and spent the extra money on a NVIDIA Titan Xp instead.
     
    Last edited: Oct 13, 2017
    gr8soundz likes this.
  11. Yeang
    Energy,

    Did you deactivate all the power saving c-states and set windows to max performance? Also use Throttlestop to set cores to max frequency. That worked wonders for me with my 4770k - I could not do dsd256 without stuttering before I tried that. Look for a thread on computeraudiophile re throttlestop. There’s about a registry edit you do to turn off c-states.

    I know that Jussi said he too had to set his windows to max performance in the power settings before his 6950x worked with dsd512 xtr non2s.

    Try it before you return the processor and let me know.

    Yeang

     
  12. Energy
    I have disabled power saving c-states in both BIOS as well as windows registry command (https://www.computeraudiophile.com/...-rid-of-dropouts/?tab=comments#comment-603849)

    I've set the latest version of Throttlestop to mac core frequencies (33) and my cores still fluctuate below stock frequency (3300MHz) during audio playback. I even disabled hyper threading.
    At this point I don't even think that the CPU isn't strong enough but due to the frequency difference probably causes some latency issues between cores during the DSD encoding process.

    I don't know what to do now. Either get the Intel i9 7960X 16 core or find another way to solve this issue. I feel like spending extra money won't help this case since Jussi's 6950X is already working fine and the 7900X I have should be better than that CPU.
     
  13. guymrob
    The sonic difference between DSD512 and DSD256 is minimal but the processing power is more than double! In my opinion a i5 can just do DSD256 but i7 has no problem. It doesn't make sense to invest such a power PC to do DSD512; the noise, heat contributed to the over all performance. Not worth doing that unless you use an endpoint. DSD256 in my opinion is good enough.
     
    Last edited: Oct 14, 2017
  14. Yeang
    Well, that’s distressing... I would message Jussi at CA (he responds very quickly) and see what he says.

    I believe he said the one thing that was worse on the 7900x compared to 6950x was the cache amount which he thinks is quite important. Threadripper has a lot more cache I think.

    Did you ever try dsd512 xtr with your Threadripper before you returned it for low FPS reasons?

     
  15. Energy
    DSD512 compared to DSD256 is fairly noticeable for me but definitely not worth the money in terms of fresh investments. If someone were to build a computer for gaming purposes then it becomes much more justified to add an addition few hundred dollars on a more powerful CPU to use poly-sinc-xtr with DSD512. That same few hundreds can also be used on better audio gear so it's up to each their own. For me however the couple hundred (or even thousands) on a new DAC didn't show the improvements I wanted to hear. The money spent on a more powerful computer for better filtering proved to be superior through the Holo Audio Spring regarding PCM file up-sampling.

    Good looking out. I did some research on this and think you're on to something.

    My message to Jussi

    Jussi,

    I think I found the problem.

    (source: http://blog.logicalincrements.com/2017/06/intel-i9-7900x-vs-i7-6950x-vs-ryzen-7/)
    It looks like even though the 7900X is superior in terms of processing power, they have chose to decrease the L3 cache and increase L2 cache.

    Intel 6950X vs 7900X
    L3 Cache:
    6950X = 25MB
    7900X = 13.75MB

    L2 Cache:
    6950X = 10 x 256KB
    7900X = 10 x 1MB

    The link shows how different programs are more optimized for certain architectures. I'm guessing the DSD Encoding process of HQPlayer takes up more L3 cache. That even if the 7900X has stronger compute power, it's missing the "real-time" cache required by your software.

    Important Quote: (from the link)
    "The lower L3 cache should in theory be a good thing, but in reality it changes the performance of the new Skylake-X CPUs in cache-bound benchmarks, where lower performance can be seen in L3 cache-dependent tests. This approach is very similar to the “victim cache” used in Ryzen processors. As a result of this, Skylake-X has a similar problem to Ryzen: the latency between core-to-core and thread-to-thread communication is higher, but not as high as it is within Ryzen processors."

    So not only is the L3 cache less than the 6950X, but there seems to be more latency between core and thread communication, perhaps something fairly as important in PCM to DSD encoding, both of which results in a combining explanation for the stutters that I'm getting. It doesn't seem like processing power is the issue as given from my overclock tests, but having more cores does equal more cache so that's all I can do for now. At this point I am going to try and upgrade to the 7920X, which gives an additional 2 cores (totaling 12) and another 2.75MB worth of L3 cache. I can't justify spending more than $1000 on a CPU so if the 7920X doesn't work, I'm switching over to AMD because who knows how many more cores or 2.75MB increments I'll need until it hits the non-stuttering threshold. If the AMD ThreadRipper 1950X has latency issues and still stutters in music, I'll meet you in the 6950X realm. :)

    AMD 1920X vs 1950X
    L3 Cache:
    1920X = 32MB
    1950X = 32MB

    L2 Cache:
    1920X = 12 x 512KB
    1950X = 16 x 512KB

    Anyways these are all just theories but they make sense to me. Based on the information above, the 32MB offered by AMD should suffice as it's even well over the 25MB from the 6950X. However the architecture on AMD processors work differently so at this point we'll know if we don't try. I'll be the guinea pig. Let me know if you think i'm wrong or if anything came to mind during your read. Thanks.

    Cheers,
    Thanh
     
    Last edited: Oct 16, 2017

Share This Page