This comes up constantly because the Google Coral USB Accelerator is famous in homelabs — it’s the go-to for Frigate object detection. And since Immich also does “object detection” and smart search, people naturally assume a Coral will accelerate it too. It won’t, and understanding why saves you $60 and a lot of confusion.
Why the Coral can’t run Immich’s ML
The Coral Edge TPU is a tiny, fixed-function accelerator built to run small, quantized convolutional models (like MobileNet) for lightweight object detection. That’s exactly what Frigate needs — draw a box around a person or car, fast, at low power.
Immich’s machine learning is a different beast. It runs large transformer models through ONNX Runtime:
- CLIP embeddings for smart/semantic search (“show me photos of the beach”) — a vision transformer.
- Facial recognition — detection plus a recognition embedding model.
These models don’t fit the Edge TPU’s instruction set or its tiny memory. You can’t compile a CLIP transformer down to run on a Coral. The Immich team looked at this and declined Coral support for exactly this reason — it’s the wrong class of hardware for the job.
What actually accelerates Immich
Immich supports real hardware acceleration for its ML container — just not via Coral. Your options, best to simplest:
- NVIDIA GPU (CUDA) — fastest. Any modern NVIDIA card dramatically speeds up CLIP and face processing.
- AMD GPU (ROCm) — works on supported Radeon cards.
- Intel (OpenVINO) — uses an Intel iGPU or CPU. This is the sleeper pick for mini PCs (N100, etc.) that have no discrete GPU.
- Apple Silicon / ARM — Immich also has ARM-NN and Rockchip (RKNN) builds for SBC-class hardware.
You enable it by using the hwaccel.ml.yml compose file and switching the immich-machine-learning image to the matching tag (-cuda, -openvino, -rocm, -armnn, or -rknn). The official Immich docs have the exact, version-current YAML — follow those rather than a copy-paste that may be stale.
Do you even need acceleration?
Honestly, for most people: no. CPU-only ML works fine. The catch is the initial pass over a big library — on CPU, processing 50,000+ photos for smart search and faces can take a day or more. But it runs in the background and catches up, and once it’s done, ongoing processing of new photos is trivial. If your library is under ~20k photos, don’t bother with a GPU. If it’s huge and you want the smart-search index built quickly, a GPU (or Intel OpenVINO) is worth it.
Already bought a Coral? Repurpose it
A Google Coral USB isn’t useless — it’s just for the wrong app here. Put it where it belongs: Frigate NVR, where an Edge TPU chews through real-time camera object detection at a few watts. That’s the setup it was designed for. See the Self-Hosted Apps hub for the Frigate track.
What about Hailo?
If you specifically want an accelerator for Immich down the road, the more promising hardware is the Hailo line (as used in the Raspberry Pi AI Kit) — it’s far more capable than a Coral, and the Immich team has pointed to it as a more realistic future target than the Edge TPU. Nothing official yet, but it’s the one to watch.
Frequently Asked Questions
Does Immich support the Google Coral TPU?
No — and it never has. The developers declined it because the Coral is too weak for Immich’s CLIP and face-recognition models. Coral is for Frigate.
What hardware accelerates Immich machine learning?
NVIDIA (CUDA) is fastest; AMD works via ROCm; Intel iGPUs/CPUs via OpenVINO. CPU-only works too, just slower on the first big scan.
Do I need a GPU for Immich?
No. CPU is fine for most libraries. A GPU only meaningfully helps the initial ML pass on very large (50k+) libraries.
Related
- My full Immich review & migration off Google Photos
- Self-Hosted Photos hub (Immich vs the alternatives)
- Self-Hosted Apps hub (Frigate — where the Coral belongs)