Tihon

You also mentioned earlier your custom kernel for Ubuntu. Isn’t it necessary?

Oh, remote access makes things a lot easier. I assume my Mac’s terminal is the app to use for such connection, right? Will I need to setup some username/password on server for that SSH, to be able to login it from my Mac later? If so, anfter installing SSH there will be some kind of interface for that, or I need to know other commands to do that?

Hmmm… With Ubuntu for HQP I will need to download and install Ubuntu from official website, than install Miska’s kernel, than NVIDIA drivers and than HQP Embedded? And all it need to be done on the computer intended for HQP? Problem is, I don’t have any monitor (I use laptops for work), so each time to do something (like setup bios et.c.) I have to wait till children go to bed and carry my 20+ kg case to a living room, where I can connect it via HDMI to TV. Therefore, it is not an option after each step to write to the forum and wait for an answer. I would really appreciate if somebody can explain or point me to an article/guide how to install Ubuntu, HQP, drivers and whatever else is needed to get in the end a server that will work like my present HQP OS build (no need for keyboard or monitor, just a press power button and it’s working).

Whish I knew how to Install it. I understand how to create bootable USB with standard desktop Ubuntu using my MacBook. It seems to be easy, just download .iso and use Balena Etcher. But that’s standard desktop OS for people who want to use computer like… well, like normal computer 🤣 How to create a bootable ISO that will contain server Ubuntu with your kernel, HQP Embedded, NVIDIA drivers and CUDA toolkit - no idea. Pages with software, like yours: https://www.signalyst.com/embedded-install.html? or NVIDIA’s https://docs.nvidia.com/datacenter/tesla/tesla-installation-notes/index.html are obviously written for those who already know what they are doing. For a Linux-newbie like me this is totally incomprehensible 😳

What does this real-time kernel give us from a practical point of view? I don't doubt it makes some sense since you're using it, just wondering what the advantage is. More efficient use of the CPU, faster interface or better sound, something else? I was curious and tried googling realtime kernels. I came across some forum about Linux, where someone asked about it. He was advised to use low-latency and leave real-time kernel for guided missiles and industrial equipment 🤣

I thought your HQP OS with its realtime kernel is the most effective from performance perspective. But I don’t know much about how Linux works. Ubuntu or Fedora with low-latency kernel might be better for running hardware at max load? Your screenshot says it’s playing Sinc-MGa, ASDM7EC-super at 512x48. My system also plays that (plus convolution) perfectly, even without GPU. Can you play -super or -light modulator with short filter at 1024x48 with the help of GPU? Can you play 512x48 with -super and short filter if CUDA offload is off?

With short filters and no convolution it plays and it is listenable, but not 100% perfect. Occasional dropouts from time to time. At 512 everything plays perfectly in all conditions: any source, -super modulator, convolution, longer filters (except for monstrous sinc-L).

I've checked everything: all cores, AVX2 enabled/auto, OC to 6.3Ghz. No luck with ASDM7EC-super at 1024 no matter what settings are. Looks like too much is too much 🥲 Apparently we'll have to wait a couple more generations of processors. On the other hand, with already incredible sound of -super modulator at 512 waiting will not be boring 🤣

ASUS ROG STRIX z790-e

I’ll try both scenarios. Will see how it will work out with HQP. I was playing with Cinebench R23 to get some understanding how BIOS settings work. Deactivating E-cores leads to increase in single-core performance benchmark and keeps temperature lower. HQP OS knows a difference between P-cores and E-cores? In other words, will it always use P-core (or one of two “favored” P-cores) for heavy modulator task, or it might end up using E-core (which is obviously slower) for modulator?

Ok, looks like CPU is not enough for -super modulator reaching 1024 rates. Even with short filter and no convolution I get hiccups every 5sec. Will see if some OC helps.

You mean GPU RAM here, not system? I assume all 4090 have same 24Gb GDDR6. Is 32Gb of system RAM enough for my case, or I will need not only add GPU, but also add another 32Gb? As for cores, correct me if I wrong… for stereo playback we will see two cores heavily used (modulators for right and left channel) and other cores with much lighter load? Looks like it might be useful to disable “unnecessary cores” to run the rest at higher frequencies. 13900 has 8 P-cores, so it’s probably worth trying to disable 2-4 cores and run HQP OS on 4 or 6.

So if modulator works at 1024 with short filter and no convolution, GPU should be able to make this rate work with heavier filter? Good place to start, thank you! How many cores HQP OS will normally use? What it depends on? I want to try a special OC strategy to increase multiplier for only necessary cores. It is not typical gamers OC, but might work here. Well, 4090 is the fastest possible option for the task, right? I wasn’t planning to go with budget options. If I add a GPU, I want the most powerful one so as not to regret later. Price is a secondary concern, especially since even with GPU my server will still cost less than the CD transport I sold after hearing DSD upsampling 🤣

Computer performance depends on many factors, even the same hardware may or may not be able to handle the load depending on cooling, power, etc. I understand how it works and I'm not going to blame you later that you said GPU would work 🤣 I just want to roughly predict the chances and understand whether in my case it is worth buying and installing a GPU, learning how to install Linux, etc. Right now I’m listening to ASDM7EC-super 512+ fs / sinc-MGa at 512 and the sound is really great. I wonder what are the chances that 4090 will give me ability either to go up to 1024 with -super modulator and keeping my sinc-MGa+convolution, or to stay at 512, but use the sinc-L filter + convolution. I’ve tried both this configs and while my system wasn’t able to play it flawlessly, but the sound between interruptions was so unbelievably good, that I can’t stop thinking of it. 2all Gentlemen, anybody here has an HQP embedded server with 13900K/KS + 3090/4090 GPU? What rates you were able to achieve? P.S. Both DSD1024 and sinc-L are not possible with currently available hardware, right?

OK. Understood the interface. It's really the same thing, just organized differently. I've build a custom PC for HQP OS: 13900KS 32Gb DDR5 7200 no GPU (only built-in) ASDM7EC-super 512+ fs / sinc-MGa / convolution = plays any source files at 48x512, stuttering when limit set to 48x1024 ASDM7EC-light 512+ fs / sinc-MGa / convolution = stuttering at 1024 ASDM7EC-light 512+ fs / default filters / no convolution = plays 1024 sinc-L stuttering even at 512 😱 I've tried overclocking CPU to 6,3 GHz, disabling E-cores, et.c. Nothing helped to play ASDM7EC-super at 1024. Probably not doable without GPU. @Miska how do you think... If I add 4090 to my system and run HQP Embedded under Linux will it do 48x1024 with ASDM7EC-super 512+ fs, convolution and since-MGa reliably or not? CUDA offload helps with modulators also, or only with filters/convolution? P.S. Auto rate family in Embedded means that if checked it it will play 41 as 41x and 48 as 48x and if not checked - everything will play at highest rate?