high cpu and igpu usage #21158
Replies: 4 comments 9 replies
|
Your configuration has a few issues that need correction based on the Intel N150 chipset requirements. Configuration IssuesFor the Intel N150, the recommended detector configuration uses OpenVINO with the GPU device explicitly specified(1). Your current detector configuration should be updated to: detectors:
ov:
type: openvino
device: GPUFor the model configuration with OpenVINO and the default SSDLite MobileNet v2 model, use: model:
width: 300
height: 300
input_tensor: nhwc
input_pixel_format: bgr
path: /openvino-model/ssdlite_mobilenet_v2.xml
labelmap_path: /openvino-model/coco_91cl_bkgr.txtHardware Acceleration for FFmpegThe ffmpeg:
hwaccel_args: preset-vaapiThe documentation notes that for Intel N150 specifically, it "Can only run one detector instance"(3). The N150 has an inference time of approximately 15ms with the default model and approximately 25ms with YOLO-NAS at 320 resolution(3). Performance ExpectationsThe Intel N150 has known performance characteristics documented in Frigate. With the default MobileNetV2 model, expect around 15ms inference time(3). High CPU and iGPU usage is expected behavior for this hardware when running object detection, especially with the integrated GPU handling both video decoding and AI inference workloads. Driver RequirementsYou mentioned seeing Recommended Configurationffmpeg:
hwaccel_args: preset-vaapi
detectors:
ov:
type: openvino
device: GPU
model:
width: 300
height: 300
input_tensor: nhwc
input_pixel_format: bgr
path: /openvino-model/ssdlite_mobilenet_v2.xml
labelmap_path: /openvino-model/coco_91cl_bkgr.txtIf you continue experiencing issues with GPU detection, check the troubleshooting guide which notes that some Intel iGPUs may require additional configuration, such as ensuring the correct 📚 Sources: Have another question? Just tag @inkeep. |
|
We will need some screenshots of the metrics to see what is actually using your CPU. You should also provide logs as there is no way for us to guess what the error might be for qsv |
Based on the AI response i explicitly put GPU under openvino. If i click on the test video and enable detection, after a little while the CPU ramps up to 57% and igpu has a low amount of 4% or less, however the metrics don't seem to show usage at all. The logs don't show anything: my LXC configuration seems good: lxc.cgroup2.devices.allow: a root@proxmox:/etc/pve/lxc# ls -l /dev/dri I do apologize but not sure how to share any other logs if needed. I clicked on system logs and system metrics. |
|
iHD driver is correct for that hardware, could be an outdated or problematic driver version though. Difficult to say because yeah something is wrong with the communication with the iGPU |
|
I tried apt install i965-va-driver on proxmox, also setting lxc.environment: LIBVA_DRIVER_NAME=i965 in the specific LXC config file and still ffmpeg doesn't like it. I will keep playing with this and will try on a VM to see if i get better results. I wanted to use LXC to consume the less amount of resources possibles on proxmox since i only have 16 GB of ram and want to run other things |
|
Yes, like I said you need iHD. i965 is a legacy driver. |
|
dmesg reports i915 [ 4.610571] i915 0000:00:02.0: [drm] Found alderlake_p/alderlake_n (device ID 46d4) integrated display version 13.00 stepping D0 i'll digging up further |
|
yes, i915 is iHD |
|
Experiencing the same, iHD, ffmpeg processes eat too much cpu resources model: |
|
hmm seems like setting cpu governor did the trick |
|
Oh man thanks for cpu governor hit, cpu usage and temperature are now low!!!! Big thank you! |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Hello, i just got a mini pc with the intel n150 chipset and are trying to set frigate as privileged LXC container. I am not using Coral, just intel gpu. I understand this is not supported, and might try out later the other methods in the documentation (i.e qemu VM or using docker). I also read through #5773
After installing it with the test video (/media/frigate/person-bicycle-car-detection.mp4), I turned on detection and performance is quite slow, both cpu and igpu, despite giving it 4 cores and 4gb of ram (and nothing else running, since this a brand new proxmox 9.1.1 installation.
My main question for this thread is whether my config looks fine:
the documentation says I should use preset-intel-qsv-*, but ffmpeg doesn't like this parameter. I also tried preset-intel-qsv-h264 with the same result. If i use preset-vaapi then it does work but with the same performance issue.
So i guess the question is, how do I get QS to work? Do I need to install any drivers on the proxmox host? I already confirmed all steps in the Passing through the iGPU section are performed correctly but I do not have any drivers. I am using kernel 6.17 and if i run lsmod I do see xe,i915 showing for video, drm and a lot others.
All reactions