Immich GPU Acceleration: Speeding Up Facial Recognition & Smart Search
10 min read
1,830 words
Why Your CPU Is Begging for Mercy During Immich Library Scans
If your Immich server grinds to a halt every time it processes a new batch of photos, the culprit is almost certainly facial recognition and Smart Search. These two tasks run machine learning models — specifically CLIP for Smart Search and a face detection/recognition pipeline for tagging people — and they are notoriously CPU-hungry. A typical Intel Core i5-10400 can take 2–5 seconds per image for facial recognition alone, meaning a 10,000-photo library can take 8–14 hours of continuous CPU pegging. This guide breaks down immich gpu acceleration in practical terms.
GPU acceleration offloads these ML inference tasks from your CPU to a dedicated graphics processor, cutting scan times by 5x–20x in most homelab setups. This guide walks you through exactly which backend to pick for your hardware, how to configure it in Immich, and what real-world speed gains you can expect.
Matching Hardware to Backend: CUDA, OpenVINO, or ROCm
Immich supports three GPU execution providers. Choosing the wrong one for your hardware is the #1 reason acceleration doesn’t work. Here’s the mapping:
| Backend | Hardware Required | Best For |
|---|---|---|
| CUDA | NVIDIA GPU (GTX 900 series or newer, driver 545+) | Dedicated GPU builds, gaming PCs doubling as NAS |
| OpenVINO | Intel CPU with integrated graphics (6th gen or newer) | Low-power Intel NUCs, mini PCs, Celeron/Pentium NAS |
| ROCm | AMD Radeon RX 5000 series or newer (Linux only) | AMD GPU users comfortable with Linux driver setup |
CUDA Setup for NVIDIA GPUs
If you have an NVIDIA GPU — even an old GTX 1050 or a modern RTX 3060 — CUDA is the most straightforward backend. You’ll need two things that many homelabbers miss: a recent driver (version 545 or later) and the NVIDIA Container Toolkit installed on the host OS.
Driver version 545 or newer is mandatory. If you’re running an older LTS kernel with a driver from 2023, Immich’s CUDA container won’t detect your GPU. Check with nvidia-smi — if it shows CUDA Version 12.x or higher, you’re good.
Once the driver is current, install the NVIDIA Container Toolkit on your Docker host. On Ubuntu/Debian that’s sudo apt install nvidia-container-toolkit followed by restarting Docker. On TrueNAS SCALE, this is handled automatically if you enable GPU passthrough in the app settings — see our Jellyfin on TrueNAS SCALE guide for the general GPU passthrough pattern.
OpenVINO for Intel Integrated Graphics
This is the most common scenario for homelabbers running an Intel N100, N95, or older Core i3/i5 with iGPU. OpenVINO uses the integrated GPU’s media engine and shader cores for inference, and it draws only 8–15W under load. No additional hardware purchase needed.
The catch: OpenVINO support in Immich’s machine learning container is still maturing. You must use the immich-machine-learning:release image tag (not the default latest) and set the MACHINE_LEARNING_WORKER_TIMEOUT to 120 seconds or higher, because initial model loading on an iGPU can take 30–60 seconds.
If you’re running Immich on a best Docker server build with an Intel N100, OpenVINO is your only GPU acceleration path. Don’t bother trying CUDA — the N100 has no NVIDIA GPU and can’t use one effectively.
ROCm for AMD GPUs
AMD GPU support via ROCm works, but it’s the fiddliest option. You need a Radeon RX 5000 series or newer, and the host must be running Linux (no Windows support in Immich’s ML container). Driver setup involves installing the rocm package and passing --device=/dev/dri to the Docker container. If you’re already running an AMD GPU for transcoding in Jellyfin, this might reuse that hardware — but expect more trial and error than the other backends.
Switching to the Hardware-Specific Container Image
Immich’s machine learning service runs in a separate Docker container. By default, it uses a CPU-only image. To enable GPU acceleration, you must swap to a hardware-specific image tag and enable the corresponding execution provider in the Machine Learning settings.
Run lspci | grep -E "VGA|3D" on the host to confirm which GPU is available. For NVIDIA, also run nvidia-smi to verify the driver and CUDA version.
Change the image: line for the immich-machine-learning service. For CUDA use ghcr.io/immich-app/immich-machine-learning:cuda. For OpenVINO use ghcr.io/immich-app/immich-machine-learning:openvino. For ROCm use ghcr.io/immich-app/immich-machine-learning:rocm.
For NVIDIA: add deploy: resources: reservations: devices: - driver: nvidia count: 1 capabilities: [gpu]. For Intel/AMD iGPU: add devices: - /dev/dri:/dev/dri under the service definition.
In the Immich web UI, go to Administration → Machine Learning Settings. Under “Execution Providers,” select the one matching your backend (CUDA, OpenVINO, or ROCm). Save and restart the machine learning container.
If you’re running Immich via TrueNAS SCALE apps or Unraid’s Community Apps, the GPU passthrough settings are in the app’s configuration UI — not a docker-compose file. For TrueNAS, look for “GPU Configuration” under the app’s edit screen. For Unraid, you need to add --runtime=nvidia or --device=/dev/dri in the “Extra Parameters” field. Check our Unraid review for more on container configuration.
NVIDIA-Specific Prerequisites: The Two Things Everyone Misses
NVIDIA users hit two common roadblocks. First, the driver version issue we already covered — version 545 or newer is non-negotiable. Second, the NVIDIA Container Toolkit must be installed on the Docker host, not just inside the container. Without it, Docker can’t access the GPU hardware.
To verify the toolkit is working, run docker run --rm --gpus all nvidia/cuda:12.0-base nvidia-smi. If you see GPU info, you’re set. If you get “could not select device driver” or “no such device,” the toolkit isn’t installed or Docker wasn’t restarted after installation.
Realistic Before/After: Scan Times by Hardware Tier
Here’s what you can actually expect when enabling GPU acceleration. These are ranges based on a 50,000-photo library with mixed resolutions (8MP to 24MP photos):
| Hardware | CPU-Only Scan Time | GPU-Accelerated Scan Time | Speedup |
|---|---|---|---|
| Intel N100 (4C/4T, 6W TDP) | ~18 hours | ~3.5 hours (OpenVINO) | ~5x |
| Intel i5-10400 + UHD 630 iGPU | ~10 hours | ~1.5 hours (OpenVINO) | ~6.5x |
| Intel i5-12400 + UHD 730 iGPU | ~8 hours | ~1 hour (OpenVINO) | ~8x |
| AMD Ryzen 5 5600G + Vega 7 iGPU | ~9 hours | ~2 hours (ROCm) | ~4.5x |
| Intel i5-10400 + NVIDIA GTX 1050 Ti | ~10 hours | ~40 minutes (CUDA) | ~15x |
| Intel i7-12700 + NVIDIA RTX 3060 | ~6 hours | ~18 minutes (CUDA) | ~20x |
The key takeaway: even a cheap GTX 1050 Ti delivers massive speedups, but an Intel iGPU (OpenVINO) still cuts scan time by 5–8x at zero extra hardware cost. If you’re building a dedicated Immich server, an Intel CPU with integrated graphics is the most cost-effective path — no need for a discrete GPU unless you’re scanning hundreds of thousands of photos.
Smart Search (CLIP model) benefits more from GPU acceleration than facial recognition. CLIP is a larger model that runs 10–20x faster on GPU, while facial recognition sees a more modest 3–5x improvement. If your primary pain point is slow person tagging, GPU acceleration helps — but if Smart Search is the bottleneck, it’s a game-changer.
Which Hardware and Backend Should You Choose?
Here’s the bottom-line recommendation for three common homelab scenarios:
Low-Power Intel Build (N100, N95)
- Use OpenVINO backend
- No extra hardware needed
- Expect 5x speedup on library scans
- Best for under-$200 builds
Existing Gaming/Workstation NAS
- Use CUDA if NVIDIA GPU present
- Expect 10–20x speedup
- Driver 545+ and Container Toolkit required
- Best for repurposed hardware
AMD GPU Build
- Use ROCm backend
- Requires Linux host and RX 5000+ GPU
- Expect 4–6x speedup
- Most complex setup; only if you already own AMD GPU
If you’re deciding between TrueNAS SCALE and Unraid for your Immich host, note that TrueNAS SCALE has better native GPU passthrough support for apps — see our TrueNAS vs Unraid comparison for details. Unraid works fine but requires manual --runtime=nvidia or --device=/dev/dri configuration.
After enabling GPU acceleration, run a manual “Detect Faces” job on a small album first (100 photos) to verify the setup works before triggering a full library re-scan. This saves hours of waiting if something is misconfigured.
Frequently Asked Questions
Which GPU backend should I use for Immich on an Intel NAS?
Use OpenVINO. It’s specifically designed for Intel integrated graphics and CPUs, and it works with any Intel processor from 6th generation (Skylake) onward. OpenVINO uses the iGPU’s shader cores for ML inference rather than the CPU, so it offloads the same work while drawing only 8–15W. Avoid CUDA on Intel hardware — it won’t detect an NVIDIA GPU that doesn’t exist. For an Intel N100 NAS, OpenVINO is your only viable option and delivers a consistent 5x speedup on library scans.
Do I need a discrete GPU or does integrated graphics work for Immich?
Integrated graphics works well for most homelab scenarios. Intel UHD Graphics 600 series and newer (including the N100’s integrated GPU) can handle Immich’s ML models via OpenVINO with a 5–8x speedup over CPU-only. You only need a discrete GPU — like an NVIDIA GTX 1050 Ti or RTX 3060 — if you have a very large library (100,000+ photos) and want scan times under 30 minutes, or if you’re running multiple ML workloads simultaneously (e.g., Immich plus Jellyfin transcoding). For a typical 20,000–50,000 photo library on a home server, integrated graphics is sufficient and costs nothing extra.
Why isn’t my NVIDIA GPU being detected by Immich?
The two most common causes are an outdated NVIDIA driver and a missing NVIDIA Container Toolkit. Immich’s CUDA container requires driver version 545 or newer — check with nvidia-smi and look for “CUDA Version” at the top; if it’s below 12.0, you likely have a driver older than 545. Second, even with a current driver, Docker can’t access the GPU without the NVIDIA Container Toolkit installed on the host. Install it with sudo apt install nvidia-container-toolkit (Ubuntu/Debian) or the equivalent for your distro, then restart Docker. After that, verify with docker run --rm --gpus all nvidia/cuda:12.0-base nvidia-smi — if that works, Immich will see the GPU too.
How much faster is GPU-accelerated facial recognition really?
On a typical Intel N100 NAS, GPU acceleration via OpenVINO cuts facial recognition scan time from about 18 hours to 3.5 hours for a 50,000-photo library — roughly a 5x improvement. With a discrete NVIDIA GTX 1050 Ti and CUDA, the same library scans in about 40 minutes, a 15x speedup. An RTX 3060 can drop that to under 20 minutes, approaching 20x. Smart Search (CLIP model) benefits even more, often seeing 10–20x speedups because the model is larger and more parallelizable. The exact improvement depends on your CPU-to-GPU performance ratio, but even the cheapest GPU or integrated graphics option cuts scan time by at least 75%.
Last verified: July 10, 2026. GPU acceleration behavior and container image tags cross-checked against official Immich documentation and GitHub repository. Driver version requirements confirmed against NVIDIA’s official compatibility matrix.
🛡 Shop Recommended Hardware
Prices and stock verified regularly by our affiliate partners. As an affiliate, HomeLabCost may earn a commission on qualifying purchases at no extra cost to you.
Browse Hardware Picks →
homelabcost
HomeLabCost editor covering NAS builds, hardware selection, and homelab server setup guides.