Intel Arc A380 Performance Test

[michael-pptf]

Test Platform

CPU: Intel i3-12100
Motherboard: AsRock B660 ITX
RAM: Teamgroup DDR4-2666 64GB
GPU: AsRock Arc A380 Challenger 6GB
Frigate: 0.15-rc1
Test Clip: https://www.youtube.com/watch?v=MNn9qKG2UFI
Model: YOLO-NAS-s-320.onnx
Detector: OpenVINO with GPU

Test is done on a Ubuntu 24.04.1 LTS VM running in Proxmox 8.3. The VM is assigned with 4 vCPU and 8GB of RAM with the Arc A380 passed through.

Methodology

Transcoded the test clip to 480x270, 640x360, 854x480, 960x540 and 1280x720 for testing, as well as using the original 1920x1080 resolution to simulate detection load with various stream resolutions and detector resolutions.

Used a shell script which:

1. Modifies the detection source and detector dimensions based on the testing clip
2. Starts the Frigate container
3. Queries intel_gpu_top for performance data
4. Runs for 5 minutes
5. Stops the intel_gpu_top query
6. Stops the Frigate container
7. Repeats tests until all resolutions are tested.

Results

Terminology:
IRQ/s: Indicates how many IRQs are sent from the CPU requesting work done by the GPU. The higher the IRQ, the higher the CPU Utilization.
RC6 %: Amount of time GPU spent in idle, waiting for work. The lower the number, the lower the GPU Utilization.
VCS %: Amount of time the Video Engine is busy. The higher the number, the higher the load on video decoding/encoding.
CCS %: Amount of time the Compute Engine is busy. The higher the number, the higher the load on compute (In this context, OpenVINO workloads).

Detector Resolution	GPU Freq MHz (avg)	IRQ/s (avg)	RC6 % (avg)	VCS % (avg)	CCS% (avg)	Inference Speed (ms)	CPU Utilization %
480x270	386	128	39.06	0.29	5.61	9	13
640x360	509	154	34.25	0.41	7.24	9	17
854x480	646	207	31.86	0.61	10.77	9	23
960x540	735	232	29.61	0.70	12.17	9	29
1280x720	1266	309	8.97	0.96	16.56	10	35
1920x1080	1316	390	16.55	1.80	21	11	42

Note: Inference speed is very dependent on the complexity of the frame. The clip I used should be considered as such - heavy traffic with pedestrians. I consider this is a "worse case scenario" for a home NVR. This test is meant to stress the subsystems to find the limits of the hardware used.

Conclusion

Based on the tests performed, I believe both 854x480 and 640x360 are suitable for detection streams with a GPU like the Arc A380. The 854x480 resolution for detection excels at far away objects since the detector has more pixel to work with, while the 640x360 or even the 480x270 resolutions should be used in near field detections like doorbell and porch areas, where cameras are positioned relatively close to the field of interest.

Since I've only given 4 vCPUs to this VM (while the i3-12100 has 4C8T), CPU resources should not be a concern for any modern x86 hardware that has more than four physical cores when a suitable lower resolution detection stream is used.

[michael-pptf]

Couple of examples for detections at various resolutions

1080p

Objects are being clearly detected even at the edge of the frame.

720p

Slightly lower confidence for the further away objects, but they are being picked up by the detector.

480p

Detection at distance started to degrade a bit, but still maintained high confidence when classifying the parked black car on the far left. I think this is the resolution that fits the best for a wide field of view camera pointing at an open area.

360p

Detection performance degraded significantly when the object is far away, evidenced by the confidence in classifying the parked car. Given the result, this resolution is best used for monitoring areas that are closer to the camera.

[ProsjektX]

Hi! I'm curious to know, what's the inference speed on the 12100 using the same yolonas model but running openvino on CPU, and also in the case of using the built-in iGPU?