I think HQplayer doesn’t even take advantage of the M Apple GPU
Yes, the GPU is not supported, because in it’s current form it is not suitable for the task. Or in general tasks like scientific computing. Maybe in future it will and then it could be supported.
Computing needs of AI are pretty much opposite of HQPlayer. HQPlayer requires support for high precision arithmetic. While AI operates primarily on reduced precision arithmetic, like 8-bit integers and half-precision 16-bit floating point math.
Thanks for the detailed answer! 
I think the stronger performance/efficiency cores should provide more reserves, if I understand correctly?! They are about 15% stronger than the M4, which should be enough for smooth playback with heavier filters and more serious noise shaping.
Another question:
Which is the best NAA for DSD1024? The Holo Red is a very good unit, but what else is also recommended, but not too expensive?
Thx.. 
Probably any SBC that is supported by NAA OS, with images on Jussi’s repository, paired with a good USB isolator like Intona.
1 Like
Thx, I have a Holo Audio Cyan 2 Dac, its usb input is completely galvanically isolated. 
Is an isolator necessary? Which one supports DSD1024?
This USB 3.0 SuperSpeed Isolator
Model 7055-C
417,83 € incl. VAT (not cheap)
This insulator is not cheap, wouldn’t it be easier to buy the Holo RED?
I read somewhere that the RPI5 is a better choice than the RPI4 because it has better LAN and USB handling and stability.
Wait for Jussi to jump in, you will get a more accurate answer. He did mention that 7054 model is cheaper and delivers basically same result. And yes, Holo Audio DACs carry very good designed USB audio interfaces, optically isolated, and adding an separate isolator doesn’t improve much or at all.
There are various Intel based small devices. And the basic RPi4. I use for example UP Gateway. But it is not in production anymore. Now it predecessor is UP 4000 which should work fine too (I have one of those, but I have not tested it as a NAA recently)
In my case UP Gateway is driving Holo Spring 3, Gustard A26 and Musical Fidelity M3x DAC. Holo Red is driving TEAC UD-701N.
Thx!
@jussi_laako
For now I’ll look at the RPI 4 / 5 line. Thanks for the quick reply!
I3-8100T Mini PC in server room (Roon Server) -
Apple Mac Mini M4 Base model (HQPlayer 5 + DAC Correction) it’s Super silent!
RPI4 HQPlayer NAA
Holo Audio Cyan2
So far it works perfectly! The RPI4 will be moving to a new all-aluminum case. (Akasa Gem Pro) A-RA09-M2S)
2 Likes
If you start with a flac file that has a higher sample rate (96, 192, etc.) will it still work? My M4 Studio will pause randomly with those files.
Interesting. Mine definitely has intermittent pauses when starting with hi res files. Maybe it’s because I have DAC Correction turned on with my DAC 200.
It is also important to remember enable High Power Mode:
Otherwise it will thermal throttle before reaching maximum performance.
1 Like
I don’t think the M4 Mac Studio has the option to do that.
I also have DAC correction turned on.
Maybe some service is running in the background? Mine is a brand new machine, with a brand new system. There is almost nothing installed, just HQPlayer and TG Pro. I bought this machine specifically to run HQPlayer. This is my dedicated PCM/DSD upscaler.
Apple “M-Scaler”. 
This is my 16-port Gigabit switch:
Continuous, stable operation! (24/7) All wired network devices are connected here. 
2 Likes
Could you please elaborate on that? If we have high rate multichannel DSD, can resampler use more resources that convolution does ? since its at source level ? I’ve added an 5090 to existing 5080, wondering how to split operations. Also there are only three cuda related attributes - cuda_dev, cuda_cdev. I guess i use cuda_dev for resampler offload and cuda_cdev for convolution offload. I’ve added a second GPU only because sometimes 16G VRAM is not enough.
Yes, although cdev is used for all large RAM hungry operations and dev is used for the smaller ones. So it is not absolute split between these operations, but rather on size of the operations. So you can have one GPU with less RAM and another with a lot of RAM. This split also means that the dev is more latency sensitive and cdev is less latency sensitive.
These settings point to the GPU id that is shown in nvidia-smi and is also printed out by HQPlayer at startup.
I just managed this but its not easy. Fanless would be impossible. You need more then 32GB of DDR5-8600+, preferably a KIT of 2x24GB single sided Hynix M-die.
It consumes 23GB of VRAM and 37GB of DDR5. So an RTX5090 is a must, even this spikes to 100% GPU load sometimes.
And even if it plays fine without hickups and scratches etc.. the delay is insane!!! from the time i lower the volume on my iphone, i must wait like 5+ seconds for the volume to go low on speakers. Sinc-L is out of the question for realtime applications.
1 Like
I know this is asking a lot, but will any of these off the shelf PCs do it?
https://www.bestbuy.com/site/searchpage.jsp?id=pcat17071&st=14900+pc
These are some of the higher spec 5090 machines:
From what I’ve been taking from this thread, this one looks to be one of the better choices:
For sinc-L, the most sensitive bottleneck is the RAM bandwidth… even with DDR5-8400 there are still hiccups. if you really need this, I would advise feeding HQPlayer with 96k source… that would be dsd512 but at 49mhz frequency.