NPU Performance - Frigate 0.17-beta2, Core Ultra 2 285H

[mabnz]

I'm seeing poor performance when using Frigate 0.17-beta2 with a Core Ultra 2's NPU. I've updated the host drivers to the latest Intel drivers (v1.28, 20251113MTL_CLIENT_SILICON-NVR+NN-deployment72[snip]907ffc78) but still seeing poor performance when compared to GPU.

iGPU outperforms by about 15-20ms. 18 cameras, processing about 115fps.

iGPU: ~13-18ms
NPU: 30ms

Has anyone had this issue and solved it?

[inkeep-ai-assistant[bot]]

Based on the available documentation, your Core Ultra 2 285H includes an NPU that should be supported in Frigate 0.17+.(1) The Intel 125H is mentioned as including an NPU for more efficient detection in 0.17+, suggesting Intel Core Ultra processors with NPUs are supported.(1)

However, the performance you're experiencing (NPU at ~30ms vs iGPU at 13-18ms) appears suboptimal. According to the hardware acceleration documentation, OpenVINO can run on Intel Arc GPUs, Intel integrated GPUs, and Intel CPUs to provide efficient object detection.(2) OpenVINO will automatically be detected and used as a detector in the default Frigate image when a supported ONNX model is configured.(2)

For Intel-based CPUs, the recommended hardware acceleration preset depends on your CPU generation.(3) For gen13+ processors (which includes your Core Ultra 2 285H), the recommended preset is preset-intel-qsv-* with the iHD/Xe driver.(3)

The NPU performance characteristics may vary based on the model being used. The documentation notes that OpenVINO runs best with tiny, small, or medium models.(2) If you're using a larger model, this could explain the performance difference you're seeing.

Without specific NPU tuning guidance in the current documentation, you may want to verify:

• Your model size and type
• Whether your detector configuration is explicitly targeting the NPU device
• Your OpenVINO detector configuration settings

The iGPU performance you're achieving (13-18ms) is actually quite good and falls within expected ranges for Intel integrated graphics.(1)

📚 Sources:

1. Recommended hardware
2. Supported Hardware
3. Video Decoding

---

Have another question? Just tag @inkeep.

[NickM-27]

Which model are you running?

[mabnz]

yolov9s 640x640

[NickM-27]

That seems to match what I see on the 125H, so it seems correct to me. Note that in the majority of cases using 320x320 is recommended.

[shjips]

A bit off-topic, but have you been able to get 125H/185H to work properly when using vmware + ubuntu + docker?

[yangminglong]

I hope this is useful to you #21148 (reply in thread)