I’m now looking at GPUs 
Do you have recommendations on RAM? Looking at these graphs, 12GB should be good for 512 but not enough if I tried to double that? Would 768 fit in 12GB?
A user I trust said 4070 super should be the card to obtain. That gives me 12GB. At some point I’ll need to rebuild the whole thing anyway so trying to find the right value here before just starting over. Thanks again.
If you’d like to do something like sinc-L at DSD512, you may run out of RAM with 12 GB. So I would recommend to seek for 16 GB. Currently cheapest option for that is the new 4070Ti Super.
I would instead look for 4070Ti Super.
Hejlo.
Which filter/mod/rate?
So this works to DSD1024 without overclocking:
1x = poly-sinc-gauss-long
Nx = poly-sinc-gauss-hires-lp
Modulator = ASDM7EC-super
While this doesn’t:
1x = poly-sinc-gauss-long
Nx = poly-sinc-gauss-hires-lp
Modulator = ASDM7EC-super 512+fs
I’m just a bit curious if the difference between 14900K and 14900KS is enough to close the gap and make the second one work too. Or if it still needs some overclocking. This case shouldn’t be RAM bound, so it should solely depend on the CPU itself (P-cores).
This is from 44.1/48 PCM > 1024?
This is no problem on my stock machine. 14900k and 3090.
Jussi, thank you from the bottom of my heart for such a technically detailed answer!! This makes us all a little more knowledgeable in this area. And now I can see that comparing filter operation by means of spectrum gives more noticeable differences and I understand why the client part has such visualisation.
I just wanted to share what pushed me down the path of analysis and comparison. I have a DAP with R2R architecture and 256 FPGA taps. One of the filters has a very comfortable and non-fatiguing sound for me and the noted community. The manufacturer spices its description with words like “ultra” “secret” etc :). In search of the truth and “grail” I tried to take a picture of its AFC and impulse (attached below). How would you rate this filter? Which of the current list of available HQP filters is it closest to from the technical and scientific point of view?
Looks like asymFIR
Yes, on my 14900K + RTX A4500. And AVX2 build on Ubuntu Desktop 22.04 to be exact.
I need to check also with the Intel build, maybe it is different.
What is your environment / HQPlayer build?
Looks like combination of FIR and IIR which is what ESS does. And looks like 4x oversampling since you can can see lower half of the first image at the right side of the plot.
From that plot it looks like quite a bit of NOS-like roll-off and ripple in top octave (IOW, some pre-echo too). Followed by steeper roll-off. Depending on how that measurement came to be.
I am on 5.6.1 W11 networked to Red and has been possible since around 5.1 time if iirc.
Closes to that is the Intel build on Linux. I need to check that out on this particular hardware as well. (I usually use this workstation for the AVX2 builds)
I think my new favorite is poly-sinc-gauss-long / hires-lp ASDM7EC-light 512+fs
I’m sorry @jussi_laako to react so lately. You answered to @Michael_Laviolette that for its Gustard R26 a good starting point was DAC Bits set to 20.
From a past recommendation, I captured the Bits set to 16 for this DAC. Did this value revisited or I misunderstood, may be 
Othewise, @J_de_Hond, I’m a big fan of the 705.6/768k PCM and Gauss filters + LNS15 (as suggested by Jussi) on my main setup with a R26. I tested several time SDM and definitively prefer PCM on this DAC. I use its internal streamer+NAA.
On my Topping E50 (headphone setup), I prefer to stream in SDM.
Have you ever considered to integrate in HQPlayer an AI with music style recognition and/or using other parameters like user’s preference to select automatically the best filters, modulator ?
Such solution can integrate a learning tool interacting with the user like what Roon is doing with the Roon radio.
Those cloud backends are using a different method than AI for the recommendations. And those are not based on recognizing the music content, but instead based on metadata relations.
Modulator depends on the DAC and one’s preferences more. While filter is more preference and genre specific.
But yes, I have thoughts around this. But it doesn’t require AI.
Yes Jussi the hype AI term is a bit over or under used, it depends
There’s probably some algos and sometime not the more hypes and last one that can fit the needs. Some old school technics deserve to be studied may be. But just a thought, we all already appreciate HQPlayer as it is
On the other subject, any consideration about the right R26’s bits set ?
Would need to have it on a test bench to give exact value. But at least with 16 bit you should be safe.
can someone tell me where I’m going wrong? aiming for DSD512 but output shows DSD256.
Rockna wavelight server
Rockna Wavelight dac
via i2s connection
if I use the wavelight server upsampling, the dac can do DSD512.
You have SDM pack “DoP” selected. This is DSD over PCM. It’s allows for DSD to be sent across drivers which only understand PCM but takes twice the PCM bandwidth. That means 512 native becomes 256 DoP.
Some DACs only support native DSD via their own drivers. I don’t know your DAC. But I’m going to assume this is why you’re only seeing 256 with your NAA endpoint. Does it work without DoP?
