Coral vs Hailo: the best AI accelerator for Frigate

Your Frigate shopping list might be one item shorter than you think. If you already have an Intel N100 or N150 mini PC, try its integrated graphics with OpenVINO before buying an accelerator. When that handles your cameras and chosen model, you are done.

Building around a Raspberry Pi 5? Choose the 13 TOPS Hailo-8L AI HAT+ for a dedicated detector. Need an external accelerator for a compatible Linux PC with no internal slot? Coral USB is the slot-free retrofit, although Frigate now favors other options for most new installations.

This research-based guide uses maker specifications and Frigate documentation; inference figures below are the project’s published examples.

Quick picks: start with your host

Your starting point	Pick	Why it fits	Setup that matters
Intel N100/N150 mini PC already on your desk	You may not need one: OpenVINO on the iGPU	Uses the Intel graphics you already own	Give Frigate GPU access and configure its detector/model
Raspberry Pi 5	Hailo-8L AI HAT+ 13 TOPS	A Pi-shaped PCIe detector, with Frigate’s Hailo backend	Fit the HAT, install the matching host driver and expose the device
Compatible Linux PC without a suitable internal slot	Coral USB Accelerator	External USB connection; no M.2 slot needed	Pass USB into Frigate and use an Edge TPU-compatible model
PC with a suitable spare PCIe M.2 socket	Hailo-8 M.2 is worth considering	Internal installation with more Hailo compute than the 8L	Match the exact key, size, PCIe wiring and driver

The linked HAT offer is Yahboom’s AI HAT+ 13 TOPS listing. Its own product documentation identifies the Hailo-8L 13T variant for Raspberry Pi 5. Choose that variant rather than the separate 26 TOPS option.

For the broader host decision, see our mini-PC buying guide and Raspberry Pi alternatives guide. A Pi HAT belongs on its supported Pi host; it does not become compatible with every SBC that has a similar-looking header.

What are you accelerating?

Frigate has two jobs that are easy to lump together. First it decodes camera video into frames. Then its object detector examines selected regions for people, cars and other objects. An AI accelerator speeds up that second job.

Buying a Hailo or Coral does not automatically fix expensive video decoding. Configure Frigate’s video hardware acceleration separately. Intel graphics can serve both purposes when configured for each: video decoding through FFmpeg, object detection through OpenVINO.

Start with a sensible detect stream, too. Your camera’s full-resolution recording feed and its detection feed have different jobs. Our PoE camera guide for Frigate walks through the lens, lighting and main-stream/substream choices. A well-framed camera gives your detector useful pixels instead of a huge picture of nothing happening.

Intel N100 or N150: try OpenVINO first

OpenVINO is the first route to try on a supported Intel mini PC you already own. Frigate’s hardware guide lists Intel integrated graphics as a detector option and publishes N100/N150 examples. You do not need to add a USB stick or fill a storage slot just to get started.

Follow the Intel setup in Frigate’s getting-started guide:

1. Give the Frigate container access to the Intel GPU’s render device. The documentation uses /dev/dri/renderD128; use the render node belonging to your GPU.
2. Configure video decoding under Settings → Global configuration → FFmpeg with the appropriate hardware-acceleration preset.
3. Open Settings → System → Detectors and model, add OpenVINO, and select GPU. Configure a supported model as the guide describes; its supplied MobileNet example and a custom YOLO model have different settings.
4. Start with one camera, then add your planned streams and watch the detector and camera metrics under real activity.

If that setup keeps up with your workload, skip the accelerator purchase. If it struggles, look at decode errors, detect resolution and model size before buying another piece of silicon. The detector documentation explains supported model formats and the trade-off between model size and inference time.

N100/N150 graphics are an iGPU, not an Intel NPU. Pick the GPU device for this route. “OpenVINO” describes the software backend; it does not mean you need a separate Intel accelerator.

Pi 5 pick: Hailo-8L AI HAT+ 13 TOPS

For a Pi 5 recorder, the Hailo-8L HAT is the straightforward dedicated-detector choice here. It attaches through the Pi’s PCIe connection rather than using a USB port. Allow room for the board, cable and cooling when choosing your case.

The Raspberry Pi AI HAT+ product brief distinguishes the chips: Hailo-8L is rated at 13 TOPS; Hailo-8 at 26 TOPS. Those are maker compute ratings, not camera frame rates. The selected Yahboom listing is the 13 TOPS version, so that is the hardware being recommended in the Parts list.

For Frigate, fitting the board is only the physical half. Use Frigate’s Hailo installation instructions for the host driver and firmware. They distinguish Raspberry Pi OS Bookworm’s built-in-driver conflict from Trixie’s DKMS route. Follow the instructions matching your OS and Frigate release rather than mixing a general Pi demo environment with a different container runtime.

After installation, the documented device is /dev/hailo0. Make that device available to the Frigate container. In Settings → System → Detectors and model, select Hailo-8/Hailo-8L and the PCIe device. Frigate can select and download the default YOLOv6n model for the detected hardware, then use the cached model offline.

You can move on to custom models later. First get the default detector working, walk through the camera view, and check that objects appear in Frigate. That gives you a useful starting point before you change models and chase accuracy improvements.

Coral USB: best external retrofit

Choose Coral USB when the connection is the deciding feature: you have a compatible host, no suitable internal slot, and want an external Edge TPU detector. It is also a sensible way to keep an existing Coral-based setup working without rebuilding the hardware around a new interface.

Google’s USB Accelerator datasheet specifies a USB 3.0 Type-C connection and 4 TOPS peak INT8 performance. The Coral setup guide covers its runtime and connection requirements. For this article’s Frigate setup, plan a supported Linux host and the documented container access—not a plug-and-play promise for any operating system.

Frigate’s current hardware recommendation keeps Coral support but favors other detectors for most new installs, with exceptions for particularly low-power needs or hosts unable to use alternatives. That is why Coral wins the USB retrofit role here rather than the overall new-build badge.

The USB version needs no Coral kernel driver on the host. Frigate still needs access to the USB device and its Edge TPU software backend. Follow the installation guide for the /dev/bus/usb mapping, then select EdgeTPU, device usb, in the detector settings. Frigate includes a default Edge TPU model; custom models must be compiled for that hardware.

Hailo M.2: a socket is only the beginning

A Hailo-8 M.2 module can make sense inside a PC, especially when you want Hailo without a Pi-shaped board hanging outside it. Hailo’s module specifications describe M, B+M and A+E key variants, with different dimensions and PCIe lane arrangements. M-key uses four PCIe lanes; the other listed variants use two.

Match the module to the socket’s physical key, available space and PCIe wiring. An M.2 connector intended only for SATA or USB is the wrong route. Also account for the storage drive you might otherwise put in that socket.

We have kept the shopping links focused on the identified USB and HAT products. For an M.2 purchase, get the exact module part number and key from the seller, then match them to Hailo’s specification. The chip name alone leaves too much of the installation undecided.

Models and published inference times

The model has to match the accelerator. Frigate’s detector guide uses TensorFlow Lite with the Edge TPU delegate for Coral and Hailo-compiled HEF models for Hailo. A random PyTorch download cannot be dropped into either backend. Hailo-8 and Hailo-8L also need the correct compiled model for their respective hardware.

Frigate’s published hardware examples include:

Hardware	Named model	Published inference time
Hailo-8	YOLOv6n	About 7ms
Hailo-8L	YOLOv6n	About 11ms
Intel N100 iGPU	MobileNetV2	About 15ms
Intel N150 iGPU	MobileNetV2	About 15ms

These are separate project examples, not a controlled ranking across one identical model and host. In particular, the Intel rows use a different model from the Hailo rows. Use them to understand the available routes, then watch your own metrics with your cameras.

Avoid turning the TOPS numbers into a camera-count calculator. Stream decoding, activity, model choice and inference demand all affect the recorder. Start with one working detector backend; Frigate does not support mixing different backend types for object detection in the same instance.

What changes under Proxmox?

There are two hardware-access layers to solve: host → guest, then guest → Frigate container. Passing a device to a VM does not automatically give the container inside it access.

For Coral USB, configure USB passthrough to the guest, then expose USB to Frigate. For Hailo M.2, plan PCIe passthrough and a compatible guest driver before mapping the detector device into Docker. For Intel OpenVINO, the guest must actually have access to the Intel graphics device.

Frigate’s installation guidance prefers direct hardware access over a VM deployment. If you want the simplest first recorder, use a dedicated Linux host. If Proxmox is your chosen platform, work through device ownership and isolation first; our passthrough requirements guide explains the PCIe/IOMMU groundwork. Its GPU examples help with that groundwork; the Hailo driver setup comes from Frigate’s own instructions.

Make the first camera useful

Choose your PoE camera and streams, give Frigate its decoder and detector, then test the scene you care about. Verify that a person is tracked, a recording plays, and the system stays healthy while your planned cameras are active.

For an Intel mini PC, start with OpenVINO. For Pi 5, choose the Hailo-8L HAT. Use Coral USB when an external Edge TPU is the practical fit. Then head back to the Home Assistant hub and put that local video to work.

Before choosing the detector, give the cameras a power plan: our PoE switch guide for Frigate matches port counts and maker camera wattage to four switch options.