Hailo-8 vs Google Coral TPU for Frigate NVR: which edge AI accelerator wins in 2026

Hailo vs Coral for a new Frigate build in 2026 has a short answer: buy a Hailo-8L or Hailo-8 and treat the Google Coral Edge TPU as a legacy device. Frigate’s own hardware guide no longer recommends Coral for new installs. Google has moved the brand to a new RISC-V chip, and the Edge TPU software repos are archived. The newer edge AI accelerator costs less per camera, runs YOLO-family models, and gets first-class support in Frigate. Coral still makes sense in two cases: you already own one that works, or you need the lowest power draw possible.

Hailo vs Coral in October 2026: where Coral stands

Frigate runs object detection on your own hardware, so no frames leave the house. That job needs dedicated AI hardware, because a CPU alone runs out of room after a few cameras. As of 16 October 2026, each line below has its own source:

• Frigate’s advice. The recommended hardware page now carries a warning: “The Coral is no longer recommended for new Frigate installations, except in deployments with particularly low power requirements or hardware incapable of utilizing alternative AI accelerators.” It also promises support “for as long as practicably possible”, since the Edge TPU is still one of the most power-efficient detectors.
• Google’s site. coral.ai now redirects to developers.google.com/coral, which covers the open-source Coral NPU (released 15 October 2025) and the Synaptics Coralboard. The USB, M.2 and Dual Edge TPU product pages are gone from it.
• Google’s code. On GitHub, libedgetpu, edgetpu, pycoral and the gasket PCIe driver are all archived (read-only). The last libedgetpu release, “grouper”, dates from July 2021.
• The PCIe driver. For M.2 and PCIe cards, Frigate’s docs now point to the community gasket-builder instead of Google’s package.

Google hasn’t posted a formal end-of-life notice that we could find, so “discontinued” overstates what is on record. “Unmaintained, and no longer the recommended path” is accurate. Retailers still stock the Edge TPU hardware. If you only need presence and not video, RuView senses breathing through a wall with a $9 chip.

Hailo vs Coral price per camera

A sticker price says little until you divide it by cameras. Prices were checked on 16 October 2026. The Pi Hut lists in pounds with 20% UK VAT included, so those prices are converted to US dollars at that day’s ECB reference rate of 1 GBP = 1.3335 USD, VAT included. The camera ceiling is computed, not measured: detections per second = 1000 / inference ms from Frigate’s hardware docs, divided by Frigate’s default of 5 detect FPS per camera (camera setup docs).

Frigate publishes no Edge TPU inference time. Its sizing example uses 10 ms, which works out to 100 detections per second, but that is only an illustration. Per TOPS at the same shop, the Coral USB now costs $37.57 against $5.42 for the 26 TOPS AI HAT+, so the old “the Edge TPU is the cheap AI option” line is out of date.

Read the ceiling as an upper bound. Frigate only sends motion regions to the detector, so a quiet camera costs far less than 5 inferences per second. A busy one can cost more. One Hailo-8L user in Frigate discussion #21587 saw skipped frames at 30 to 40 detections per second, well below the 90 the latency implies. Plan with headroom.

Hailo-8 and Hailo-8L in Frigate

Hailo makes two chips for this job. The Hailo-8 delivers 26 TOPS at a typical 2.5W. The Hailo-8L delivers 13 TOPS. Hailo doesn’t publish a power figure for the 8L. Both come as bare M.2 modules and as the Raspberry Pi AI HAT+, and both run models compiled to Hailo’s HEF format. Frigate added them in version 0.15 (February 2025). In 0.16 the Hailo team rewrote the detector for better performance, and the standard Docker image gained Hailo support. The current release is 0.18.2, out on 15 October 2026.

On first start, Frigate downloads YOLOv6n from the Hailo Model Zoo for both chips. Any Model Zoo model with HailoRT post-processing also works, so you can swap in a larger model later. The detector docs give this block:

detectors:
  hailo:
    type: hailo8l
    device: PCIe

model:
  width: 320
  height: 320
  input_tensor: nhwc
  input_pixel_format: rgb
  input_dtype: int
  model_type: yolo-generic
  labelmap_path: /labelmap/coco-80.txt

YOLOv6n is a 640x640 model, but Frigate recommends a 320x320 input and lets the detector scale it up. A maintainer says in #21587 that this “should not use much CPU at all on modern systems.”

The host driver has to match the version Frigate ships. Frigate 0.16 moved to HailoRT 4.21.0 and told users to update the host driver with its install script. On Raspberry Pi OS Bookworm the kernel ships an older driver that clashes with Frigate, so the installation docs have you disable it first. On Trixie the driver comes through DKMS and that step goes away.

Hailo-8 vs Hailo-8L

The 8 has twice the TOPS. Frigate’s own model timings show about 7 ms instead of 11 ms on YOLOv6n, or 6 ms instead of 10 ms on SSD MobileNet v1. At list price the cost per camera is almost the same. Pick the 8L accelerator for a few cameras. Pick the 8 if you plan to grow, want to run a larger model, or want spare performance for other AI work.

AI HAT+ or a bare M.2 module on a Pi 5

The Raspberry Pi AI Kit is out of production, and Raspberry Pi tells new buyers to get the AI HAT+. The HAT+ switches the Pi 5 to PCIe Gen 3 by itself. A bare M.2 module on a third-party M.2 HAT also works, but it takes the slot you might use for NVMe. On an x86 mini PC, a bare module goes in a spare M.2 slot.

Hailo-10H and the USB option

There is now a USB option. The ASUS UGen300 puts a Hailo-10H (40 TOPS INT4, 20 TOPS INT8, 8GB LPDDR4) on USB-C for $299. Frigate’s docs list only the Hailo-8 and 8L, so treat it as untested for Frigate. ASUS rates it at 2.5W typical, the same figure our Hailo-10H review used for the M.2 card. Its 8GB of memory is there for small generative AI models.

Power draw and performance per watt

Hailo rates the 8 at 2.5W typical, and the TPU spec of 2 TOPS per watt puts the Coral at about 2W. That is about 10 TOPS per watt against 2. The extra 0.5W running all year adds up to 0.5 x 8,760 hours = about 4.4 kWh, a dollar or two a year at typical home power rates. Power draw isn’t a reason to pick either chip for a home NVR.

Google Coral Edge TPU in Frigate

Coral’s Edge TPU is a 4 TOPS chip rated at 2 TOPS per watt, so it draws about 2W of power. It comes as a USB stick, M.2 A+E and B+M cards, a Mini PCIe card and the M.2 Dual Edge TPU. The dual card holds two Edge TPUs for 8 TOPS, and you only get both if the slot wires two PCIe lanes. The config below is the same two-line block our Frigate with Coral guide starts from.

detectors:
  coral:
    type: edgetpu
    device: usb

The default model is MobileDet, bundled in the container. Frigate’s hardware page says Coral “supports primarily ssdlite and mobilenet model architectures”. Since 0.17, Frigate can also run a quantized YOLOv9 model on Coral, with only part of the COCO labels. Frigate+ offers a YOLOv9 model for it with all 41 of its labels.

The USB version needs no host driver and takes its power from the port. Frigate also warns that it “lacks the automatic throttling features of the other versions”, so give it airflow. PCIe and M.2 cards need the gasket driver, now built by the community.

Where Hailo loses to Coral

Night under IR. In discussion #21587 a user moved from a Coral USB to a Hailo-8L on Frigate 0.16.3. Daytime accuracy felt better, but at night “detection is hopeless compared to Corals MobileDET”. A second user on the full 8 missed people more than 7 to 8 feet away under IR. The maintainer’s answer was Frigate+ YOLOv9s: “I have had no issues with YOLOv9s with my IR or color night cameras.” If your cameras run on IR at night, budget for Frigate+ or test the stock model’s performance first.

Host CPU load. A post in the Frigate NVR Users Facebook group reported detector CPU load of 4% on Coral and 18% on a Hailo-8L, rising above 70% at night with dropped frames. The hosts differed (J4125 against N150) and no one found the cause, so treat it as a warning.

Alternatives to Google Coral

Frigate’s hardware page lists several other AI paths off Coral, with its own inference times:

• Intel iGPU through OpenVINO. The N100’s integrated GPU runs MobileNetV2 in about 15 ms. You may already own this GPU, which makes it the cheapest AI option.
• Intel NPU through OpenVINO. Supported since 0.17 on Core Ultra chips. Frigate suggests the NPU for detection and the GPU for other AI tasks such as face recognition, because NPU performance is close to the GPU at lower power.
• Nvidia GPU through ONNX. An RTX 3050 runs YOLOv9 tiny at 320 in about 8 ms. A discrete GPU draws far more power, so only use one that is already in the box.
• AMD GPU through ROCm. A Radeon 780M iGPU runs YOLOv9 tiny at 320 in about 14 ms. Apple Silicon works through a host-side detector.
• Rockchip RKNN. Community-supported object detection on RK3562, RK3566, RK3568, RK3576 and RK3588 since 0.14. An RK3588 with three NPU cores runs YOLOv9 tiny in about 35 ms.
• MemryX MX3, a community-supported M.2 AI module whose default model is YOLO-NAS-Small.

Frigate can’t mix detector types for object detection. You cannot run the two side by side, so an old Edge TPU cannot become a second detector next to a new AI accelerator. It can still serve a second, low-power box.

Which accelerator to buy

• New Raspberry Pi 5 NVR: the AI HAT+, 13 TOPS for a few cameras, 26 TOPS for more AI performance.
• New x86 mini PC with a free M.2 slot: a bare Hailo-8 or Hailo-8L module, or try OpenVINO on the Intel GPU first since it costs nothing.
• Working Edge TPU already installed: keep it until a rebuild or a kernel upgrade breaks the driver.
• IR-heavy night scenes on Hailo: plan for Frigate+ YOLOv9, or test performance on your own cameras first.
• Box already has an Nvidia or AMD GPU: use it through ONNX or ROCm before buying any AI hardware.
• Only USB available: a Coral USB still works, or the $299 UGen300 if you accept that Frigate does not list it.

Whichever chip you buy, it needs clean camera streams, and you can keep Reolink cameras local with go2rtc to feed it. Old Android phones can serve as extra sensors and cameras too. Frigate alerts then land in Home Assistant, which you can run in a Proxmox VM.

Hailo vs Coral FAQ

Is Google Coral discontinued?

Not formally, as far as public records show. Google’s site now covers a new RISC-V NPU, the Edge TPU repos are archived, and Frigate no longer recommends Coral for new installs. Retailers still sell the hardware.

What are some good alternatives to Coral?

For Frigate: a Hailo-8 or 8L, an Intel iGPU or NPU through OpenVINO, an Nvidia or AMD GPU you already own, or a Rockchip RK3588 board.

Is Coral AI owned by Google?

Yes. It is a Google brand, and its site now lives at developers.google.com/coral. The new Coral NPU is open-source AI hardware IP that Synaptics builds into its edge chips.

Does Coral AI cost money?

The software is free and open source. The edge AI hardware is not: on 16 October 2026 The Pi Hut listed the USB Accelerator at $150.29 and the M.2 A+E card at $54.54, converted from £112.70 and £40.90 with UK VAT included.