Coral vs Hailo vs Intel NPU for Frigate Detection: Power and Latency in 2026

Choosing Frigate detection hardware comes down to three questions: which box, how many watts, and how many dollars. Frigate’s documentation answers the compute and latency questions clearly, the power question only partially, and the price question barely. This guide walks through what the official docs and vendor spec pages establish, where they fall silent, and which detector fits which home. If Frigate itself is new, start with our Frigate and Home Assistant setup guide.

A Slower Detector Can Serve More Cameras Than a Faster One

A slower detector can serve more cameras than a faster one: Frigate’s docs explain that a GPU can run multiple instances of the same model concurrently, so roughly 30 ms still keeps up with several cameras, while a Google Coral runs a single instance and needs around 10 ms to keep up (Frigate object detector docs).

That inversion flips the usual shopping logic: milliseconds are an input to the decision, not the decision. The Coral’s ceiling follows a rule of thumb from Frigate’s documentation — maximum throughput equals 1,000 divided by the reported inference speed, so a detector reporting 10 sustains roughly 100 frames per second (Frigate hardware docs). A homeowner with four cameras should ask two questions of any candidate: what does it report per inference, and how many instances can it run at once. Vendor TOPS figures answer neither.

The Three Detectors, Defined

Three detector families cover most homeowner builds: the Google Coral EdgeTPU, the Hailo-8 and Hailo-8L accelerators, and the Intel NPU reached through OpenVINO. All three appear on Frigate’s supported detector list, alongside MemryX MX3, AMD ROCm and XDNA2, Apple Silicon, TensorRT/Jetson and Rockchip (Frigate hardware docs).

Coral EdgeTPU

Coral is Google’s dedicated edge-inference chip, sold as a USB accelerator and as an M.2 module (Frigate hardware docs). It supports primarily ssdlite and mobilenet architectures, with Mobiledet the recommended model; in configuration it is type edgetpu with device usb or pci (Frigate object detector docs). It remains the fallback for hardware that cannot take a modern accelerator card.

Hailo-8 and Hailo-8L

Hailo’s accelerators ship in M.2 format, with a HAT option for Raspberry Pi boards, and support many model architectures (Frigate hardware docs). Frigate’s integration identifies the hardware automatically and selects YOLOv6n as the default for both chips; the model comes from the Hailo Model Zoo, is cached under /config/model_cache/hailo, and runs fully offline after the first download (Frigate object detector docs). The Hailo-8 itself carries 26 TOPS with fully integrated memory and no external DRAM required (Hailo-8 product page).

Intel NPU via OpenVINO

The Intel NPU is a low-power hardware block introduced with the Core Ultra CPU generation (formerly Meteor Lake) that offloads neural network computation from other devices; OpenVINO exposes it as device NPU 3720, requires an NPU driver, and caches models by default to improve time to first inference (OpenVINO NPU documentation). Frigate reaches it through the OpenVINO detector, which also runs IR models on Intel and AMD CPUs and Intel GPUs (Frigate object detector docs). Intel’s Core Ultra datasheet adds the hardware detail: the NPU IP carries two Neural Compute Engine tiles with 4k MACs in total and 4 MB of near-compute memory, running INT8 or FP16 primitives (Intel Core Ultra datasheet, NPU section).

The Latency Numbers: What Is Actually Measured

The latency figures below come from two different measurement worlds: Hailo and Intel times are reported in Frigate’s documentation, while Coral times are Google’s own model-execution benchmarks produced with a C++ harness (Coral Edge TPU benchmarks). Because they measure different things on different models, read the table as ranges, not head-to-head results.

Detector	Form factor	Compute (vendor-published)	Power (vendor-published)	Documented detection time	Models in Frigate	Status in Frigate docs
Coral EdgeTPU	USB stick or M.2 (Frigate hardware)	4 TOPS (Coral benchmarks)	0.5 W per TOPS (2 TOPS/W; ≈2 W at rated throughput — derived arithmetic) (Coral benchmarks)	6.5 ms MobileNet v1 SSD 224; 2.4 ms MobileNet v1 — Google benchmark, not Frigate-reported (Coral benchmarks)	ssdlite/mobilenet; Mobiledet recommended (Frigate object detector docs)	Not recommended for new installs except low-power or incapable hardware; still supported (Frigate hardware)
Hailo-8	M.2 or HAT (AI HAT+ 26 TOPS); x86-tested (Frigate installation)	26 TOPS (Hailo-8 product page)	2.5 W typical (Hailo-8 product page)	~6 ms ssd mobilenet v1; ~7 ms yolov6n — Frigate-reported (Frigate hardware)	YOLOv6n default, Model Zoo, offline after first download (Frigate object detector docs)	Actively maintained (Frigate releases)
Hailo-8L	M.2 or HAT (AI Kit or AI HAT+ 13 TOPS) (Raspberry Pi AI Kit)	13 TOPS (Raspberry Pi AI Kit)	Not published	~10 ms ssd mobilenet v1; ~11 ms yolov6n; yolov9-tiny 320: 18 ms — Frigate-reported (Frigate hardware)	Same YOLOv6n default (Frigate object detector docs)	Actively maintained (Frigate releases)
Intel NPU (Core Ultra, via OpenVINO)	Integrated in Core Ultra CPUs (OpenVINO NPU documentation)	Not published	No wattage — “low-power” by design only (OpenVINO NPU documentation)	MobileNetV2 ~6 ms; YOLOv9 s-320: 11 ms; s-640: 30 ms — Frigate-reported (Frigate hardware)	OpenVINO IR; 320x320 recommended (Frigate object detector docs)	Requires a supported Intel platform (Frigate object detector docs)
Intel N100 (context)	Mini-PC CPU with UHD iGPU (Intel N100 specifications)	—	6 W TDP for the whole CPU including iGPU — not an accelerator figure (Intel N100 specifications)	MobileNetV2 ~15 ms (one detector instance only); YOLOv9 s-320: 30 ms — Frigate-reported (Frigate hardware)	OpenVINO (Frigate object detector docs)	Supported, with single-instance flag (Frigate hardware)

Coral times are Google’s own C++ model-execution benchmarks excluding pre-processing, not Frigate-reported inference times. “Not published” means no figure appears in the sources reviewed. Figures come from different models and harnesses and are not a same-model head-to-head. The t-/s- prefixes on YOLOv9 names are tiny/small variants as printed in Frigate’s table.

Why Coral’s Milliseconds Are Not Frigate Milliseconds

Google’s benchmark page measures model execution only, excludes input pre-processing, and uses a C++ harness — so its 6.5 ms for MobileNet v1 SSD 224 (against 109 ms on a desktop CPU) and 2.4 ms for MobileNet v1 are not Frigate inference readings (Coral Edge TPU benchmarks). Frigate-reported times, such as the Hailo-8’s ~7 ms on the default YOLOv6n, come from Frigate’s own hardware table (Frigate hardware docs). Quoting a Coral millisecond as a Frigate millisecond overstates what Google measured.

Power Draw: The Documented Numbers and the Gaps

Power draw has documented numbers and real gaps: the Coral is rated at 4 TOPS using 0.5 watts per TOPS (Coral Edge TPU benchmarks), Hailo-8 at a typical 2.5 watts (Hailo-8 product page), and the Intel NPU only as “a low-power hardware solution,” with no wattage published (OpenVINO NPU documentation).

Multiply the Coral’s two published figures — 4 TOPS at 0.5 W per TOPS — and rated throughput works out near 2 W; that is arithmetic on Google’s numbers, not a measured device figure. Frigate’s docs nevertheless call the Coral “one of the most power-efficient devices for executing object detection models” (Frigate hardware docs). The N100’s 6 W TDP covers the entire four-core processor including its UHD graphics — a CPU figure, not a detector figure (Intel N100 specifications). Hailo’s “best-in-class power efficiency” and “highest cost-efficiency (TOPS/$)” lines are vendor claims, not independent measurements (Hailo-8 product page).

What No Vendor Publishes

The gaps are as instructive as the numbers. In the sources reviewed, no Hailo-8L wattage appears, the Intel NPU publishes neither watts nor TOPS, and no wall-plug or whole-system measurement exists for any complete Frigate build — host CPU, RAM, storage and accelerator combined. The 2.5 W and 0.5 W-per-TOPS figures also come from different vendors’ pages under different conditions, not under Frigate load, so a watts-per-camera comparison is not possible from public documentation.

Why 320x320 Beats 640x640 for Camera Detection

For camera detection in Frigate, 320x320 beats 640x640 for the vast majority of setups: Frigate’s docs describe 320x320 as optimized and the best choice because Frigate crops a region of motion and zooms in before detection, making 320x320 better at small and distant objects than it appears, while 640x640 is slower and uses more resources (Frigate object detector docs).

The crop-and-zoom pipeline is why low resolution isn’t a compromise here. Start every new setup at 320x320 and treat 640x640 as a specialist escalation that costs inference time and resources across every camera, every frame.

Raspberry Pi 5 Owners Have a Clear Path: The AI HAT+

Raspberry Pi 5 owners have a clear path: the AI HAT+ starts at $70 and comes in 13 TOPS (Hailo-8L) and 26 TOPS (Hailo-8) versions (Raspberry Pi AI HAT+), and Raspberry Pi states new customers should purchase the AI HAT+ rather than the older AI Kit (Raspberry Pi AI Kit).

The superseded AI Kit pairs a 13 TOPS Hailo-8L module in M.2 2242 form with the M.2 HAT+ that connects it to the Pi 5’s PCIe 2.0 interface (Raspberry Pi AI Kit). One software detail decides whether the detector appears at all: on Raspberry Pi OS Bookworm the kernel ships an older Hailo driver that is incompatible with Frigate, so the built-in driver must be disabled and the correct version installed; on Trixie the driver installs via DKMS and no conflict applies (Frigate installation docs). Frigate’s docs list verification commands, including ls -l /dev/hailo0 and cat /sys/module/hailo_pci/version. For the full build, see our guide to local object detection on the Raspberry Pi AI HAT.

Intel Box Owners May Already Have a Detector

An Intel box owner may already have a detector: OpenVINO runs on Intel platforms from 6th Gen (Skylake) onward, covering CPUs, Intel GPUs and Intel NPUs, though using the GPU or NPU device requires a supported Intel platform (Frigate object detector docs). It adds no hardware and no purchase.

Frigate’s hardware table shows what to expect: an Intel UHD 730 iGPU handles MobileNetV2 in about 10 ms and YOLOv9 s-320 in 24 ms, Iris XE manages MobileNetV2 in ~10 ms and YOLOv9 t-320 in 6 ms, while the HD 530 (15–35 ms) and the N100 (~15 ms on MobileNetV2) are both flagged as able to run only one detector instance (Frigate hardware docs). Frigate also reports that x86 platforms with dual PCIe lanes deliver further improvements in FPS, throughput and latency compared with the Raspberry Pi setup, even with multiple cameras running concurrently, and OpenVINO lists Intel Arc A-series and B-series Battlemage GPUs among its run targets (Frigate hardware docs).

Existing Coral Owners Should Keep It — For Now

Existing Coral owners should keep it for now: Frigate’s docs no longer recommend the Coral for new installations except where power requirements are particularly low or the hardware cannot use alternative accelerators, yet commit to supporting it as long as practicably possible because it remains one of the most power-efficient detection devices (Frigate hardware docs).

That pairing — a nudge away for new builds, a long-term support promise for existing ones — narrows the Coral’s role from default pick to specialist tool for low-watt builds. Frigate also notes a single Coral can handle many cameras with the default model, so a working installation has no emergency. Migrate on a natural hardware refresh, not on a headline.

You Cannot Mix Detectors for Object Detection

Detectors cannot be mixed for object detection in Frigate: OpenVINO and the Coral EdgeTPU, for example, cannot run object detection at the same time (Frigate object detector docs). Plan around one detection backend per installation rather than expecting a fast detector and a frugal one to share the job.

An upgrade is therefore a switch, not an addition: moving from Coral to Hailo means changing the detection backend wholesale.

FAQ

The questions homeowners ask most about Frigate detectors come down to buying, power and camera count. The answers below stick to what Frigate’s documentation and vendor pages state, flag what is not published in the sources reviewed, and skip rankings the evidence cannot support (Frigate hardware docs).

Is the Google Coral still worth buying for a new Frigate install?

Frigate’s docs say no for most new installs: the Coral is no longer recommended except where power needs are particularly low or the hardware cannot use alternative accelerators (Frigate hardware docs). If one is already installed, keep it — support continues for as long as practicably possible, and its power efficiency remains a documented strength.

How many cameras can one detector handle?

Frigate avoids fixed camera counts but gives a throughput rule for the Coral: maximum fps equals 1,000 divided by the reported inference speed, so a 10 ms Coral sustains about 100 frames per second (Frigate hardware docs). For GPUs and NPUs, concurrency matters more than the millisecond figure — a detector at ~30 ms can keep up with several cameras.

Does the Intel NPU use less power than Hailo or Coral?

Public documentation cannot answer that: no Intel NPU wattage or TOPS figure is published in the sources reviewed, and OpenVINO describes the NPU only as “a low-power hardware solution” (OpenVINO NPU documentation). Hailo-8 lists 2.5 W typical and the Coral 0.5 W per TOPS, but those figures come from different vendors’ pages and weren’t measured under Frigate load.

What’s the cheapest way to add detection to a Raspberry Pi 5?

The documented entry point is the Raspberry Pi AI HAT+, starting at $70 in a 13 TOPS version with a 26 TOPS Hailo-8 option above it (Raspberry Pi AI HAT+). Raspberry Pi directs new customers to the AI HAT+ over the older AI Kit (Raspberry Pi AI Kit). Budget for the Bookworm driver fix, or run Trixie.

The Bottom Line

No overall winner emerges from this documentation, because the sources reviewed cannot rank Coral, Hailo and Intel NPU against each other on power, latency and price at once (Frigate hardware docs). What the evidence supports is a conditional call, made on the hardware already in the house.

Raspberry Pi 5 owner: buy the AI HAT+ — 13 TOPS for the documented default models, 26 TOPS for headroom — starting at $70, and sort out the Bookworm driver or run Trixie (Raspberry Pi AI HAT+).
Intel 6th-gen-or-newer owner: start with OpenVINO on the silicon already installed; it adds no purchase and no card, though the GPU and NPU devices require a supported Intel platform (Frigate object detector docs).
Existing Coral owner: keep it running; support continues for as long as practicably possible, so migration can wait for a natural hardware refresh (Frigate hardware docs).
Hard low-power constraint, or hardware that cannot use alternatives: the Coral remains the documented exception for exactly that case (Frigate hardware docs).

Detectors cannot be mixed for object detection, so pick one path and commit. And whatever computes the detections, the frames matter just as much — see our guide to the best outdoor security cameras for 2026 before spending on silicon.

How This Guide Was Built

This guide is built from Frigate’s official documentation, vendor specification pages and the Coral Edge TPU benchmark documentation. We did not run these accelerators hands-on. Every latency figure is either Frigate-reported or labeled as Google’s model-execution benchmark, every watt is a vendor-published specification or marked as derived arithmetic, and every gap is marked “not published” (Frigate hardware docs).

Sources consulted include Frigate’s hardware, object detector and installation pages, the OpenVINO NPU documentation, Intel’s N100 specification page, Hailo’s product page, Raspberry Pi’s AI HAT+ and AI Kit pages, and the Frigate release changelog, where entries such as “Update hailo installation instructions” show the Hailo detector is actively maintained (Frigate releases). No bench testing, lab measurement or long-term ownership informs any claim in this guide.

CodeIntel Log — code quality, debugging, and software engineering benchmarks
ToolBrain — tool reviews, LLM comparisons, and AI workflow guides

Cross-links automatically generated from SmartHome Field Guide.

← Back to guides