Hailo-8L vs Google Coral: the honest homelab comparison
Both of these are small, cheap edge-AI accelerators that homelabbers bolt onto a Raspberry Pi or mini PC to take machine-learning load off the CPU. They are not interchangeable, and — despite what a lot of search results imply — they do the same core job: real-time object detection for a self-hosted NVR like Frigate. Here is how they actually differ, based on running a homelab with cameras and a self-hosted photo stack every day.
| Google Coral (Edge TPU) | Hailo-8L | |
|---|---|---|
| Raw performance | ~4 TOPS (int8) | ~13 TOPS (int8) |
| Power draw | ~2 W — sips power | A few watts — still low |
| Form factors | USB, M.2 A+E, M.2 B+M, PCIe | M.2 (Raspberry Pi AI Kit / AI HAT+) |
| Model support | Int8 TFLite only, compiled for the Edge TPU — a fixed, aging set | Broader, more modern models via the HailoRT toolchain |
| Frigate support | Mature — the classic, best-documented detector | Supported in recent Frigate releases; newer, less battle-tested |
| Development pace | Largely static; the Edge TPU is old now | Active; backed by the Raspberry Pi partnership |
| Rough price | ~$60–80 (USB) | ~$70 (13 TOPS AI Kit) |
| Speeds up Immich? | No | No |
First, the myth: neither one speeds up Immich
A lot of people land here after searching “immich coral tpu” expecting to make Immich’s face recognition and smart search faster with a $60 stick. It does not work that way. As of 2026, Immich‘s machine-learning container runs on the CPU by default and accelerates on GPU/NPU backends — CUDA (NVIDIA), OpenVINO (Intel), ARM-NN (Mali) and RKNN (Rockchip). The Coral Edge TPU and the Hailo-8L are not drop-in Immich ML backends. Immich uses ONNX models that these detection-focused accelerators are not built to run.
So if your goal is faster Immich search, a cheap second-hand NVIDIA card (or an Intel iGPU via OpenVINO) is the answer — not one of these. Where Coral and Hailo genuinely earn their place is NVR object detection: watching camera streams in Frigate and flagging “person”, “car”, “package” in real time without melting your CPU. That is the fair fight, so that is how I will compare them.
Google Coral: the low-power veteran
The Coral Edge TPU has been the default Frigate accelerator for years, and for good reason. It draws about 2 W, detects objects on several camera feeds without breaking a sweat, and the Frigate documentation treats it as a first-class citizen. The USB Accelerator drops onto anything — a Raspberry Pi, an old mini PC, a NUC — with no PCIe slot required, which makes it the easiest possible entry point.
The catch is age. The Edge TPU is a fixed-function ASIC that only runs int8 TFLite models compiled specifically for it, and Google has not meaningfully updated the platform in a long time. You get roughly 4 TOPS and a frozen model set. For a two-to-four-camera setup that is plenty; for anything more ambitious it starts to show its limits. Supply has also been famously patchy, though it has eased.
Hailo-8L: the newer, roomier option
The Hailo-8L is the entry tier of Hailo’s edge-AI line, most commonly bought as the Raspberry Pi AI Kit (an M.2 HAT+ carrying the 13-TOPS Hailo-8L) or the AI HAT+ (13 or 26 TOPS). That is roughly three times the raw throughput of a Coral, a modern and actively maintained toolchain (HailoRT), and support for a broader, newer range of models. Recent Frigate versions added Hailo detectors, so it now competes directly on Coral’s home turf.
The trade-offs: it needs an M.2/PCIe slot, so in the Raspberry Pi world that means a Pi 5 with the AI HAT rather than any old board; it pulls a little more power than a Coral; and because Frigate support is newer, you will occasionally hit rougher edges than the Coral’s years of accumulated documentation. None of that is a dealbreaker — it is the normal cost of being on the newer, faster platform.
Which should you buy?
- A few cameras, lowest power, least fuss, or a Pi 4 / mini PC with no free M.2 slot → Google Coral (USB). It is mature, sips power, and Frigate loves it.
- More cameras, modern models, future-proofing, and you already run a Pi 5 or have an M.2 slot → Hailo-8L. More headroom and active development.
- You actually want faster Immich search → neither. Use an NVIDIA GPU (CUDA) or an Intel iGPU (OpenVINO). See the self-hosted photos comparison.
In my own setup the split is exactly that: an accelerator handles Frigate’s camera detection, while the photo stack leans on a GPU for its machine learning. Treating them as one problem — “which stick makes all my AI faster” — is the mistake that sends people down the wrong path. They are two different jobs.
- Immich vs Synology Photos vs PhotoPrism — picking a self-hosted photo stack
- Immich and the Coral TPU — what the Coral actually does (and doesn’t) for Immich
- Homelab hardware — the rest of the build
Specs and framework support move quickly on both platforms — check the current Frigate and Immich documentation before you buy, especially for exact model and detector compatibility.