Frigate supports multiple detector families

Current Frigate documentation lists dedicated accelerators, Intel OpenVINO, NVIDIA/AMD ONNX paths, Apple Silicon, Rockchip and other community-supported options. The existence of many choices is useful, but it makes generic rankings misleading.

Start by narrowing to hardware that your host and preferred model actually support. Only then compare cost and inference performance.

Camera count is not enough

Object detection demand follows active motion regions, not simply the number of configured cameras. A quiet installation can place surprisingly little load on a detector, while several busy entrances can create bursts of simultaneous work.

Use camera count together with activity level and detect FPS. That combination gives a more realistic view of detector headroom.

Hailo is a dedicated low-power direction

Hailo-8 and Hailo-8L are current supported Frigate acceleration modules and fit systems that can expose the required M.2 or HAT interface. They are especially attractive when you want object detection isolated from the main GPU.

The host still needs to decode streams and run the rest of Frigate, so a Hailo purchase should not distract from CPU and media-engine requirements.

Intel OpenVINO can eliminate the add-in detector

Modern Intel iGPUs, Arc GPUs and NPUs can run OpenVINO object detection. For many users this is the cleanest path because the hardware is already inside the host.

On Core Ultra systems with both NPU and GPU, Frigate recommends the NPU for object detection and GPU for enrichments. That is a powerful consolidation strategy.

NVIDIA provides broad GPU flexibility

NVIDIA GPUs can run supported ONNX detectors through the TensorRT-enabled Frigate image and can also accelerate video decoding and supported enrichments.

That flexibility is valuable when the NVR shares the GPU with other AI workloads, but it may use more power and require more driver maintenance than a dedicated accelerator.

Coral is now a special-case recommendation

Coral remains supported, but current Frigate guidance no longer makes it the default for new installations. It still makes sense for existing deployments, very low-power requirements and hosts where newer detector paths are not practical.

This distinction prevents a common buying mistake: treating long-standing community popularity as proof that the same hardware is still the best new purchase.

CPU-only is a fallback, not the target

Frigate includes a CPU detector for testing, but current documentation does not recommend it for normal use. OpenVINO CPU mode can also be more efficient on compatible systems when no GPU or accelerator is available.

If object detection is a core feature, budget for a supported hardware-accelerated path instead of designing around sustained CPU inference.

Model support can override hardware preference

Different detector backends support different model families and input sizes. Choose the model family and accuracy target first when that requirement is fixed.

A hardware platform that benchmarks well with MobileNet may not be the best choice if your desired deployment depends on a different supported YOLO or DETR-family model.

Detection hardware does not solve enrichments automatically

Semantic search, face recognition and other enrichments are independent from object detection. Some GPUs and NPUs can accelerate both areas, but dedicated detectors such as Coral do not provide general enrichment acceleration.

Plan these workloads separately so the system does not become unbalanced after enabling advanced Frigate features.

Interface compatibility is a hard constraint

M.2 keying, PCIe lanes, USB stability, GPU slot size and container device passthrough are all binary compatibility checks. If the host cannot expose the device reliably, theoretical performance is irrelevant.

The product catalogue keeps adapters as a separate class so you can see when an accelerator requires extra hardware rather than mistaking the carrier for the detector.

For mini PCs, inspect whether the advertised M.2 connector is wired for PCIe and whether the BIOS exposes it consistently under Linux. For GPUs, confirm slot width and PSU headroom. For USB Coral, avoid unstable hubs and marginal power delivery. These checks turn a theoretical compatibility list into a build that survives reboots and upgrades.

Power should be evaluated at the wall

Dedicated accelerators can be efficient, but the host platform, storage and GPU idle state dominate total NVR power in many builds.

Estimate yearly energy from the entire system. This matters especially when comparing a Hailo add-on with replacing the host to gain an integrated NPU.

Use a two-stage buying process

First use the selector to identify a detector architecture. Then compare live current products within that architecture. This prevents price from pulling you toward hardware that does not match the workload.

Finally, verify the current Frigate detector documentation immediately before purchase because supported images, drivers and models can change between releases.

This two-stage process also makes future upgrades easier. If the host, decode path and storage remain healthy, you can replace only the detector layer when model requirements change. If enrichments become the new bottleneck, you can add or upgrade a supported GPU without discarding a perfectly adequate Hailo or Intel detector. Designing the system as separate workload layers avoids the expensive habit of replacing the whole server whenever one Frigate feature grows.