Hi Ryan,
Nice to see you this morning in Rugby’s Chatfest, hope the music festival this evening goes well.
I wondered which Linux nVidia driver you are using for your GPU compute assist to HQPlayer ?
nvidia-smi tells me that I am running: Driver Version: 535.183.01 CUDA Version: 12.2
I recently upgraded to Ubuntu 24.04.1 LTS (GNU/Linux 6.8.0-44-generic x86_64) (Noble Numbat) because I was apparently cursed at birth by some bad fairy to fix things in computers that aren’t broken. Anyway, to cut a long story short, after I got everything working again I installed the above nVidia driver.
I understand the Noble Numbat generic headers default to low latency, which is nice except for folks who prefer higher latency (… Bueller ? … Bueller ?).
What I have noticed is the GPU now working a bit harder than it used to with sinc-MGA and occasionally spiking to 100%. Conversely the CPU seems to be working a bit less, so this could be either better intentional load shifting or nature’s way of telling me to get a better (more expensive) GPU.
I think my NVIDIA GeForce RTX 3060 has a Compute Capability Number of 8.6 and your RTX 4080 (?) has 8.9.
So what driver are you using and what sort of % spikes are you seeing with sinc-MGA ?
My HQP machine is off right now as my system is in process of going back together. I shall get back to you later in the week but, if you don’t hear from me, please send me a friendly reminder.
Sadly, didn’t make it to the festival. More on that another day.
I just upgraded to Noble but I decided to go with realtime kernel. Here’s some screenshot about my AMD7950 + RTX4090. I’m using in-tree nvidia-headless-550.
The sinc-MGa is one of the integer ratio filter but MCH 44.1 sampling rate contents are very rare in my library . Here’s the 6ch 44.1 Eagles: Hell Freezes Over:
Driver version 560 is latest and recommended for the latest build as it is matching the build. It may work also with earlier driver versions, although there could be potentially some problems. When the driver is too old, CUDA stack just doesn’t detect the GPU at all anymore and the offload feature falls off. (5.7.0 and 5.7.1 releases use CUDA 12.6)
There’s no version 560 driver on the Ubuntu 24 in-tree list so far. Must download and install it from Nvidia website. Users must know how to disable nouveau driver first…
It is available from the CUDA package repository, also as headless server flavor. And no, you don’t need to do anything like disable nouveau or something, since this is all automatically handled by the Nvidia’s driver packaging for the DKMS driver.
There’s a package “nvidia-headless-560-open”, which is their fully open source driver, and a meta package for the driver called “nvidia-headless-560”.
Headless driver’s shouldn’t pull in any of the X.org graphics stuff that is unnecessary for headless server use.
Regardless of load spikes, the overall power usage (56W here) gives rough average idea of how much the GPU is pushed. Although reaching the full power usage with HQPlayer is likely not something possible while having smooth playback.
$ nvidia-smi
Sat Sep 28 09:25:42 2024
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 560.35.03 Driver Version: 560.35.03 CUDA Version: 12.6 |
|-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA GeForce RTX 3090 On | 00000000:01:00.0 Off | N/A |
| 36% 59C P2 140W / 420W | 416MiB / 24576MiB | 23% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+
+-----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=========================================================================================|
| 0 N/A N/A 714 C /usr/bin/hqplayerd 406MiB |
+-----------------------------------------------------------------------------------------+
24/96 → DSD256 poly-sinc-guass-xla ASDM7ECv3
$ nvidia-smi
Sat Sep 28 09:31:16 2024
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 560.35.03 Driver Version: 560.35.03 CUDA Version: 12.6 |
|-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA GeForce RTX 3090 On | 00000000:01:00.0 Off | N/A |
| 38% 58C P2 136W / 420W | 3174MiB / 24576MiB | 38% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+
+-----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=========================================================================================|
| 0 N/A N/A 714 C /usr/bin/hqplayerd 3164MiB |
+-----------------------------------------------------------------------------------------+
24/96 → DSD256 sinc-MGa ASDM7ECv3
I don’t notice a ton of difference in load between guass-xla and MGa on the GPU but CPU I could get guass-xla but not MGa. I don’t think I can find a combination which won’t work at DSD256. I need a new DAC before I need new GPU at this point.
. Here’s the 6ch 44.1 Eagles: Hell Freezes Over:
