Skip to content

AI accelerator setup

By default, Frigate runs object detection on the NVR's CPU. This works, but each detection frame takes 80–150ms to process, which limits how many cameras you can run at high detection rates before the CPU becomes the bottleneck.

A hardware AI accelerator moves that inference workload to dedicated silicon. Detection time drops to 5–15ms on a Hailo-8, and the CPU stays available for recording, transcoding, and running other services.

Which hardware is in your system

Tier Accelerator
Vigil Optional add-on (Hailo-8)
Sentinel Hailo-8 PCIe M.2 included
Warden Hailo-8 PCIe M.2 included

If your system shipped from LAN Foundry with an accelerator installed, the drivers and runtime are already set up. Skip to Part 2: Verify the accelerator is recognized.

Part 1 — Install the accelerator drivers (post-purchase upgrade)

Follow this section if you are adding a Hailo-8 accelerator to a Vigil system after purchase.

If the accelerator was included when your system shipped, skip to Part 2.

The Hailo-8 driver and HailoRT runtime are distributed by Hailo directly. Follow the installation guide at hailo.ai/developer-zone for the current Ubuntu installation steps. The guide covers adding the Hailo apt repository, installing HailoRT, and loading the kernel module.

After installation, reboot the NVR:

sudo reboot

Then verify the device node is present:

ls /dev/hailo0

Part 2 — Verify the accelerator is recognized

Confirm the device node exists:

ls /dev/hailo0

Confirm the PCI device is visible to the kernel:

lspci | grep -i hailo

A result containing Hailo Technologies confirms the hardware is recognized.

If either check returns nothing or a "No such file or directory" error, see the troubleshooting section below before continuing.

Part 3 — Configure Frigate

Open config.yml on the NVR. The file is typically at /opt/lanfoundry/config/frigate/config.yml or the path shown in your Docker compose file.

Add a detectors block to config.yml:

detectors:
  hailo:
    type: hailo8l
    device: PCIe

Add the device passthrough in Docker compose:

services:
  frigate:
    devices:
      - /dev/hailo0:/dev/hailo0

Apply the changes

Restart Frigate after editing either file:

docker restart frigate

Check the logs for any errors:

docker logs frigate --tail 50

A successful startup will include a line indicating the detector type that initialized. If you see a warning about falling back to CPU detection, check the troubleshooting section below.

Part 4 — Verify Frigate is using the accelerator

Open the Frigate web interface and navigate to System in the top menu, then select Detectors.

The detectors panel shows the inference time for each active detector. Use these as a baseline:

Detector Expected inference time
CPU (no accelerator) 80–150ms
Hailo-8 5–15ms

If the inference time shown is in the CPU range, Frigate is not using the accelerator. See troubleshooting below.

You can also confirm from the logs:

docker logs frigate 2>&1 | grep -i detector

A correctly initialized accelerator will appear by name in the startup output.

Troubleshooting

Device node not present after installation

Confirm the system was rebooted after driver install. Then check the kernel log for device initialization messages:

dmesg | grep -i hailo

A line showing the device being registered confirms the driver loaded. No output suggests the driver did not load. Re-run the installation steps and check for any errors during apt install or modprobe.

Frigate falling back to CPU detection

Two common causes: the devices: entry is missing from the Docker compose file, or the runtime library is not installed. Verify both, then restart Frigate.

Permission denied on the device node

Docker may not have the correct group access to the device. Restart Docker and Frigate:

sudo systemctl restart docker
docker restart frigate

Still not working

Reach out to LAN Foundry support at support@lanfoundry.com with the output of docker logs frigate --tail 100 and dmesg | grep -i hailo.

Where to go from here

Was this page helpful?

Found something wrong or out of date? Let us know below, or open a GitHub issue if you have a GitHub account.