Quick Summary (TL;DR): For a new Intel mini-PC Frigate build, try OpenVINO on the integrated GPU before buying a separate accelerator. It is supported by current Frigate and avoids an extra module, slot and driver. If you already own a working Google Coral, keep using it where its model compatibility and performance suit you; Frigate explicitly says Coral is no longer its recommended default for new installations. Choose Hailo-8LAliExpress price or Hailo-8AliExpress price when you have a supported PCIe/M.2 or Raspberry Pi 5AliExpress price route, want dedicated inference capacity and are willing to maintain its host driver and compatible compiled models. On a supported Core Ultra machine, OpenVINO may use the NPU, but Frigate currently says Intel NPU detection does not work under Home Assistant OS. These are not three interchangeable USB accessories: video decoding, object detection, model accuracy, host compatibility and total ownership cost are separate decisions. This is a comparison based on current official documentation, not our own benchmark.
The Buying Verdict in One Table
| Your situation | First choice | Why | Check before spending |
|---|---|---|---|
| Already have an Intel N100/N150 mini PC | OpenVINO GPU | Uses existing iGPU; usually the simplest initial test. | Expose /dev/dri; confirm driver, model and detection headroom. |
| Buying a modern Intel Core Ultra host | OpenVINO NPU or GPU | Potentially reserves GPU resources for decoding or other Frigate tasks. | Host NPU firmware and Linux support; not HAOS for NPU. |
| Have a working Coral USB | Keep Coral | No need to replace a functioning detector solely because a newer option exists. | USB power, stable device mapping and Edge TPU-compatible models. |
| Starting a Raspberry Pi 5AliExpress price Frigate system | Hailo AI HAT+ worth evaluating | Documented Pi 5AliExpress price hardware path with 13 or 26 TOPS variants. | Pi video-decode load, Hailo driver, case clearance and recording storage. |
| Expandable x86 host, heavier inference workloads | Hailo or Intel GPU/NPU | Choice depends on model, driver support, available PCIe lanes and real workload. | Run the same model and scene workload before concluding one is faster. |
| Old PC with no usable GPU acceleration | Coral if already owned; otherwise assess host upgrade | USB Coral can extend some older machines, but may be a poor new-system investment. | Compare full host replacement with adapter, hub and maintenance costs. |
The critical change in the current Frigate recommended-hardware guide is explicit: Coral is supported, but no longer the first recommendation for a new installation except particular low-power or host-compatibility cases. That is a much better buying signal than old tutorials that insist every Frigate system needs a USB Coral.
First Understand the Three Jobs Frigate Performs
A Frigate installation is not a single AI workload. It ingests and decodes camera video, detects motion in decoded frames, then sends selected regions to an object-detection model. It may also record original video, provide live views, and run optional enrichment features. A detector accelerator handles model inference; it does not magically decode all camera streams, accelerate disk writes or eliminate the need for a capable host.
- Video decode: FFmpeg, VAAPI, Intel Quick Sync or another supported decoder processes H.264/H.265 frames. This can be the limiting factor with multiple high-resolution streams.
- Motion processing: determines which regions warrant AI inference; masks, motion settings and sensible detect streams reduce unnecessary work.
- Object detection: OpenVINO, Edge TPU or Hailo evaluates candidate regions using a specific compatible model.
- Recording and live view: require network bandwidth, reliable storage and sometimes extra CPU/GPU work even when object detection is fast.
A 4K camera does not necessarily require 4K object detection. Configure a lower-resolution substream for detection and a separate full-resolution recording stream where supported. Frigate recommends starting with a 320×320 detector model for most scenes, because it crops and enlarges motion regions before inference. See the current detector/model-size documentation.
A frequent purchasing mistake is to compare “13 TOPS Hailo” against “4 TOPS Coral” and assume the former will run exactly 3.25 times as many cameras. TOPS is a theoretical arithmetic rating, not a camera-count or application throughput benchmark. Models, precision, preprocessing, memory transfers, host decode and batching all matter.
Intel OpenVINO: The Sensible First Test on an Intel Mini PC
OpenVINO is a software inference runtime, not a separate hardware dongle. Frigate documents an OpenVINO detector for supported Intel integrated GPUs, Intel Arc GPUs, Intel NPUs and CPUs. It lists Intel 6th-generation Skylake and newer as supported platforms; the exact GPU or NPU path still depends on the processor, Linux drivers, kernel and device access. An Intel-branded CPU does not guarantee that every OpenVINO execution target is available.
On an N100 or N150 mini PC, the practical first experiment is device: GPU, with /dev/dri passed into the Frigate container. On supported Core Ultra systems, device: NPU can be considered for object detection, potentially leaving the iGPU available for video decode and other enrichments. Frigate warns that the NPU firmware belongs on the host and that Intel NPUs cannot currently be used with Home Assistant OS because the necessary firmware is absent. Verify this before paying a premium for an NPU-based design.
OpenVINO advantages
- No accelerator purchase when a suitable Intel GPU is already in the server.
- Flexible deployment across several Intel generations, subject to supported devices and drivers.
- Model choice is broader than the Edge TPU-specific compilation path, though models still need the format and settings required by Frigate.
- Easy initial troubleshooting: test CPU versus GPU availability and inspect Frigate metrics before buying hardware.
OpenVINO limitations
- The GPU is shared with decoding, display or other compute tasks; contention can reduce headroom.
- A weak or old integrated GPU may struggle with demanding models or busy scenes.
- An NPU is not automatically supported by the operating system or virtualisation stack.
- Some models require conversion/export, correct input shape, colour format, label map and paths.
Frigate provides documented OpenVINO configurations and model-export guidance. Its recommended hardware page also publishes example inference times for specific Intel devices and models; those are Frigate-published examples, not results from our lab. Compare only identical models and settings when making a purchasing decision. OpenVINO configuration and hardware examples.
Google Coral: Still Useful, But No Longer the Default New Purchase
Coral is a dedicated Edge TPU accelerator for compatible TensorFlow Lite models compiled for the Edge TPU. Google’s official USB Accelerator specification lists 4 TOPS (INT8) and USB 3.0 Type-C. The USB model can also work over USB 2.0, but at reduced inference speed. Frigate supports USB and PCIe/M.2 Coral devices with the edgetpu detector.
Coral remains attractive when you already own one, need a portable USB device or have a low-power host with no usable Intel GPU. The Frigate hardware guide notes that the USB version is broadly compatible and avoids a host PCIe driver, whereas PCIe/M.2 versions need a host driver. The USB version also lacks the automatic throttling of the other form factors. Avoid treating every Coral product as electrically identical: the dual Edge TPU board and different M.2 keying can complicate motherboard compatibility.
Where Coral loses ground
- Model constraints: a general ONNX model cannot simply be copied to Coral; it needs an Edge TPU-compatible TensorFlow Lite path and compilation.
- Upgrade flexibility: new detector/model features may be easier to deploy on Intel GPU/NPU or Hailo, depending on the model.
- USB power: Raspberry Pi and mini-PC USB buses can be marginal with several peripherals; use a properly powered hub where necessary.
- Purchase rationale: Frigate no longer recommends Coral as the default for new installations, so an apparently cheap accelerator should not be evaluated in isolation.
Coral is not obsolete or unsupported. Frigate states that it intends to support Coral for as long as practicable. The sensible recommendation is to keep a stable, adequately performing Coral installation rather than replace it to chase a marketing number. Source: Frigate hardware guidance and Coral USB product specifications.
Hailo-8L and Hailo-8: Dedicated PCIe Inference, More Setup
Hailo offers dedicated inference hardware commonly used through M.2/PCIe modules or Raspberry Pi 5 AI HAT+ boards. The official Raspberry Pi AI HAT+ product documentation identifies Hailo-8L at 13 TOPS and Hailo-8 at 26 TOPS. These are vendor arithmetic ratings, not Frigate frame-rate guarantees. Hailo-8 and Hailo-8L compiled model binaries are not universally interchangeable: the model must target a compatible accelerator architecture.
Frigate supports Hailo-8 and Hailo-8L. Its detector uses the hailo8l type for the documented Hailo integration and can select the appropriate default compiled model based on detected hardware. The current documentation says that the default YOLOv6n model is downloaded on first use, then cached for offline operation. The host needs the correct Hailo PCIe driver, firmware and access to the resulting device, normally /dev/hailo0.
The AI HAT+ is specifically a Raspberry Pi 5 accessory, not a generic USB stick. A Hailo M.2 module needs the right physical key, length, PCIe lanes, carrier and cooling. A vacant M.2 SSD socket is not automatically a compatible Hailo slot; inspect the host manual before ordering.
Hailo advantages
- Dedicated inference compute that does not have to consume the Intel iGPU used for video decoding.
- Supported integration for Hailo-8/8L in current Frigate releases.
- A documented Raspberry Pi 5 hardware route via AI HAT+; a possible x86 route where M.2/PCIe and drivers are compatible.
- Potential for more demanding models and concurrent inference, depending on the compiled model and host.
Hailo disadvantages
- Host driver/firmware management is an additional maintenance responsibility.
- PCIe slot, physical fit, heatsink and kernel compatibility must all be checked.
- Compiled
.hefmodels must match the device and the detector configuration. - A Raspberry Pi 5 still needs enough resources to decode, process, store and serve the camera streams.
Follow Frigate’s current Hailo installation instructions rather than an old blog post: Frigate documents a specific driver installation path, including different treatment of Raspberry Pi OS Bookworm and Trixie. See also Raspberry Pi AI HAT+ specifications.
Head-to-Head Comparison: What You Are Actually Buying
| Criterion | Intel OpenVINO | Google Coral | Hailo-8L / Hailo-8 |
|---|---|---|---|
| Hardware purchase | No extra accelerator if compatible Intel GPU/NPU already exists | USB or PCIe/M.2 Edge TPU hardware | PCIe/M.2 module or Pi 5 AI HAT+ |
| Frigate detector type | openvino | edgetpu | hailo8l (documented Hailo integration) |
| Device selection | CPU, GPU or NPU | usb or pci variants | PCIe |
| Model format | Compatible OpenVINO IR/ONNX path and settings | Edge TPU-compiled TensorFlow Lite | Compatible Hailo .hef |
| Host prerequisites | Intel driver, supported device and Docker mapping; NPU firmware for NPU | USB power/device access; PCIe driver for M.2 | PCIe electrical fit, driver, firmware, device mapping |
| Typical integration friction | Low to moderate on supported Intel Linux | Low for USB; higher for PCIe | Moderate to high |
| Frigate new-install recommendation | Strong starting point on compatible Intel hardware | Not recommended as default for new installs | Supported alternative when hardware route fits |
| Video decode included? | Intel GPU may separately accelerate decode when configured | No | No |
| Best purchase logic | Use existing compute first | Retain an existing working device | Buy for a confirmed compatible host and model |
These entries describe supported architectures, not a speed ranking. Intel GPU versus NPU versus Hailo inference speed depends on the same model family, image size, camera activity, driver version and other running workloads. Comparing Coral’s small default model with a larger Hailo model is not a fair benchmark.
Model Compatibility Matters More Than TOPS
A detector is only useful if it runs the model you actually want. Frigate’s current Frigate+ documentation lists support for CPU, Coral, OpenVINO, ONNX, Hailo and other detector backends, with model choices dependent on hardware. A Frigate+ subscription or custom model is a separate software decision, not a guaranteed performance upgrade from buying a more powerful NPU.
| Question | OpenVINO | Coral | Hailo |
|---|---|---|---|
| Can I reuse an arbitrary ONNX model? | Not automatically; verify supported export, model type and input settings. | No; needs compatible TFLite/Edge TPU compilation. | No; needs suitable Hailo compilation and post-processing. |
| Can I switch detector type without touching the model? | Only if a compatible version of the model exists. | Generally not; different model artefact. | Generally not; compiled .hef required. |
| Can I run two detector families at once? | No for mixed object-detection families in current Frigate. | Same restriction. | Same restriction. |
| Do I need cloud inference? | No, inference can run locally. | No, local inference. | No after initial required model download/cache. |
Frigate explicitly says mixed detector types such as OpenVINO and Coral cannot be used simultaneously for object detection. Multiple detectors of the same supported type may be possible, and using a GPU for other Frigate enrichments is a separate question. Do not purchase a Coral as a “fallback” expecting Frigate to balance the same object-detection queue across Coral and OpenVINO automatically. Source: Frigate detector documentation.
Which Option Works with Your Host?
| Host / operating system | OpenVINO | Coral | Hailo |
|---|---|---|---|
| Intel N100/N150 mini PC, Linux Docker | Good first choice if iGPU exposed | USB possible; PCIe needs compatible slot/driver | Possible only with suitable PCIe/M.2 slot and host driver |
| Intel Core Ultra Linux host | GPU; NPU if firmware/device support verified | USB/PCIe possible | PCIe possible if slot and driver supported |
| Home Assistant OS | Intel GPU route may be available in supported HA App deployments; verify device access | USB commonly used; check host and add-on | Depends on OS and driver support; verify current Frigate instructions |
| Home Assistant OS + Intel NPU | Not supported for NPU in current Frigate docs | Independent Coral path may work | Independent host support required |
| Raspberry Pi 5 + Raspberry Pi OS | Intel GPU path not applicable | USB possible, watch power | AI HAT+ is a documented hardware route; follow Frigate driver instructions |
| Proxmox VM | Needs reliable GPU/NPU passthrough | USB/PCIe passthrough possible | PCIe passthrough and guest driver required; assess complexity |
Frigate recommends Docker on bare-metal Debian-based Linux for straightforward hardware access. Its installation guide cautions that virtual machines add device-passthrough complexity, and that Docker-in-LXC on Proxmox is not officially supported by Frigate. If the camera recorder must be highly reliable, a dedicated Linux host may be worth more than an elaborate virtualisation arrangement. Official installation and Proxmox guidance.
Practical Configuration Examples (Starting Points, Not Tested Here)
The following examples show the documented detector selection only. They are not complete Frigate configurations: you must separately configure cameras, recordings, storage, model selection where needed, device mappings and network access. Current Frigate releases also allow much of this through Settings → System → Detectors and model.
Intel OpenVINO GPU
detectors:
intel_gpu:
type: openvino
device: GPUFor a Docker host, check ls -l /dev/dri and pass the correct render device into the container. If the host lacks GPU access, device: CPU is a troubleshooting fallback, not an assurance of equivalent performance. For custom YOLO models, set the model path, dimensions, tensor layout and label map according to the official OpenVINO example.
Google Coral USB
detectors:
coral_usb:
type: edgetpu
device: usbExpose the USB bus to Docker as documented by Frigate. If the Coral is intermittently detected, investigate the USB cable, USB 3.0 connection, power budget, hub and container permissions before buying a replacement.
Hailo-8 / Hailo-8L
detectors:
hailo:
type: hailo8l
device: PCIeThis is the detector selector, not a complete custom-model configuration. Frigate can select its default Hailo model; if you supply a different .hef, follow the exact model settings in Frigate’s Hailo section. First verify /dev/hailo0 exists on the host after driver installation and is mapped into the container.
Video decoding is a separate setting
ffmpeg:
hwaccel_args: preset-vaapiThis VAAPI example applies to compatible Intel/AMD GPU decode paths; it does not enable Hailo or Coral object detection. Use the documented Quick Sync H.264/H.265 presets when appropriate to your Intel hardware and camera codec. See Frigate video decoding guidance.
How to Measure Before You Buy an Accelerator
- Record a baseline: note the Frigate version, host model, kernel, camera count, codecs, detect resolutions, motion masks and active object model.
- Separate decode from inference: inspect CPU, GPU decode and detector metrics. If FFmpeg decoding is saturated, a faster TPU may not fix the problem.
- Use the same scene and model: compare detector latency and missed detections under similar real-world activity. Do not compare unlike model families or input sizes.
- Look for sustained load: short periods with no motion do not represent a driveway during rush hour or a garden with wind-blown foliage.
- Check queueing and skipped detections: low average inference time can coexist with overload when multiple cameras become active at once.
- Test recording simultaneously: full-quality recordings, live views, snapshots and enrichment features should be enabled during the trial.
- Only then decide: tune detection streams and motion masks first; upgrade host, GPU, TPU or storage based on the measured bottleneck.
Frigate notes that a Coral runs a single model instance, so latency and queue utilisation matter differently from a GPU capable of concurrent inference. Its detector guide discusses model resolution, model size and the importance of detection headroom. The hardware page publishes example timings, but these are not a substitute for measuring your own installation. Official performance interpretation.
What Will the Complete System Cost Over Five Years?
For a purchasing decision, the accelerator price alone is misleading. A Hailo module may need a PCIe carrier, heatsink or a different host; a Coral USB may need a powered hub; OpenVINO may cost nothing extra if you already own a compatible Intel PC, or it may mean buying a new mini PC. Camera storage, backup, network and power protection often exceed the cost of a detector. The following worksheet deliberately uses no unverified shop prices: enter current UK/EU quotations for the exact SKU and region.
| Five-year cost component | OpenVINO on existing Intel host | Coral add-on | Hailo add-on / Pi 5 |
|---|---|---|---|
| Host computer and RAM | £0 incremental if already suitable; otherwise enter full host cost | Enter existing host upgrade or new-host cost | Enter compatible x86 host or Pi 5, RAM, cooler and case |
| Detector hardware | £0 incremental when using existing GPU/NPU | Coral USB/M.2 unit + any PCIe adapter | Hailo module or AI HAT+ + carrier if needed |
| Driver/power accessories | Usually no separate detector power accessory | Powered hub/cable if required | Carrier, heatsink, cabling, compatible case |
| Recording storage | SSD for config/database + HDD/SSD for footage | Same | Same |
| Camera networking | PoE switch/injectors, Ethernet, uplinks | Same | Same |
| Resilience | UPS, backup disk or NAS, replacement reserve | Same | Same |
| Electricity for 43,800 hours | Measured whole-system watts × 43.8 × £/kWh | Same; include host/hub power | Same; include host/HAT/adapter power |
| Five-year total | Sum every row, including replacement costs | Sum every row, including replacement costs | Sum every row, including replacement costs |
The electricity calculation is watts ÷ 1,000 × 24 × 365 × 5 × tariff (£/kWh). For example, at a purely illustrative electricity tariff of £0.25/kWh, an additional measured 5W continuously costs £54.75 over five years; 10W costs £109.50; 20W costs £219.00. These are arithmetic scenarios, not measured power draws for Coral, Hailo or Intel systems. Measure the complete machine at the wall with your real cameras running. Electricity tariffs and power consumption vary.
Use a consistent scope: if the OpenVINO option reuses your existing mini PC, compare incremental cost for all three choices; if buying a whole new recorder, compare the entire new system for each. Do not charge one option for an existing PoE switch and quietly omit it from the others. A UPS battery replacement, off-site backup, replacement drive and spare storage capacity also belong in the five-year budget.
Storage and networking can dominate the budget
For a quick recording capacity estimate, one continuously recorded 4Mbit/s camera uses roughly 43.2GB/day before filesystem and metadata overhead: 4 Mbit/s × 86,400 seconds ÷ 8 ÷ 1,000. Eight such cameras generate about 345.6GB/day, or 10.37TB for 30 days in decimal units. This is a bitrate assumption, not a claim about any camera model. Event-only recording, different codecs and variable bitrates change the result substantially. Plan retention from actual camera output, not accelerator TOPS.
A small N100 mini PC can have ample inference capacity yet be a poor recorder if its only storage is a small soldered SSD. Before buying, check supported drive capacity, cooling, endurance, and whether you need a NAS or dedicated surveillance HDD. A 1GbE connection is often sufficient for modest camera bitrates, but PoE port count, isolation and uplink reliability are still important.
Common Buying Mistakes
- Buying Coral because a 2022 tutorial says it is mandatory. Current Frigate guidance explicitly changed; start from today’s supported detector list.
- Comparing TOPS as if it were cameras. An inference accelerator is only one part of the video pipeline.
- Buying Hailo for a slot that merely looks like M.2. Confirm PCIe wiring, keying, length, power, firmware and cooling.
- Assuming every Intel NPU works under Home Assistant OS. Current Frigate documentation says it does not.
- Using a large detection model before tuning streams. Start with a supported small model and the recommended 320×320 input where appropriate.
- Assuming an accelerator fixes poor H.265 decoding. Check FFmpeg decode acceleration separately.
- Mixing Coral and OpenVINO detectors. Current Frigate does not support mixing detector families for the object-detection queue.
- Forgetting software updates. A working Hailo PCIe driver can require attention after host kernel changes; document recovery and rollback.
Frequently Asked Questions
Is Intel N100 OpenVINO enough for Frigate?
It can be a good starting point for a modest camera system, particularly with low-resolution detect streams and properly configured hardware video decoding. The correct answer depends on camera activity, model size, decoding and recording load. Start with the iGPU and examine Frigate metrics; do not assume a fixed maximum camera count from the CPU name alone.
Should I buy a Coral USB in 2026?
Not automatically. Frigate still supports Coral and it remains useful on certain low-power or incompatible hosts, but the current recommended-hardware page explicitly advises other supported detectors for most new installations. If you already own one that meets your needs, replacing it is unnecessary.
Is Hailo-8 twice as fast as Hailo-8L in Frigate?
The published hardware ratings are 26 versus 13 TOPS, but that does not establish a universal 2× application-speed improvement. Real inference latency depends on the model and host. Check the same compiled model family, Frigate’s published examples and your actual workload.
Can a Hailo AI HAT+ accelerate video decoding?
No. The Hailo accelerator handles supported neural-network inference; the host remains responsible for ingesting and decoding video. A Raspberry Pi 5 must still cope with the chosen streams and recording requirements.
Can I run OpenVINO and Coral together?
Current Frigate documentation says you cannot mix detector types for object detection. You can use different hardware for other functions such as video decoding or enrichment, but do not design a new recorder around automatic Coral/OpenVINO detector mixing.
Does Frigate+ require Hailo or Coral?
No. Frigate+ documents models for several detector backends, including Intel/OpenVINO and Coral/Hailo. Confirm the current model offering, supported hardware and any subscription terms before purchasing an accelerator for a particular model.
What is the lowest-risk upgrade path?
On a suitable Intel host: configure VAAPI or Quick Sync video decoding, start with OpenVINO GPU, tune the detect streams and review metrics. Buy another accelerator only if those measurements reveal a genuine inference bottleneck and you have verified the model and host compatibility.
Related ESP32 and Home Assistant Guides
If you are building the wider smart-home system, these existing site guides cover adjacent hardware and integration questions:
- Install Home Assistant OS on an N100/N150 mini PC — dedicated Home Assistant hardware setup; note the separate Intel NPU/HAOS limitation above.
- ESP32 smart doorbell with Home Assistant notifications and snapshots — a complementary camera and automation use case.
- LD2450 multi-target presence sensor with Home Assistant — local presence detection, distinct from Frigate camera object recognition.
- UGREEN DXP4800 Plus vs Pro — relevant when comparing a separate Frigate recorder with a NAS used for storage or other services.
Datasheets and Official Resources
- Frigate: Recommended hardware — current detector recommendations and Frigate-published inference examples.
- Frigate: Object detectors — current OpenVINO, Edge TPU and Hailo configurations and model requirements.
- Frigate: Installation — Docker device mapping, Hailo drivers and virtualisation caveats.
- Frigate: Video decoding — FFmpeg VAAPI and Intel Quick Sync presets.
- Frigate+: Supported model families — detector and model compatibility.
- Raspberry Pi: AI HAT+ product — Hailo-8L/8 hardware variants and physical requirements.
- Google Coral: USB Accelerator specifications — Edge TPU performance rating, USB and software requirements.
Bottom line: Choose the detector that your existing or planned host can actually support, then validate it with your chosen model and camera workload. For most new Intel mini-PC builds, OpenVINO is the first option to try. Coral remains sensible when already owned or needed for a specific host. Hailo can be a strong dedicated inference path, provided the PCIe hardware, driver and compiled models are confirmed before purchase. The best investment is often better decoding, storage or camera configuration rather than a second accelerator.
Editorial verification: 10 October 2026. This is a documentation-led technical comparison, not a hands-on test. Product availability, host kernels, Frigate versions, accelerator firmware, supported models and regional prices can change. Confirm the exact device, software release and retailer terms before buying.