What hardware is required for a huge set of photos/videos?

[McBenac]

Having 500'000 photos and 250'000 video clips, approx 5TB, on a Synology NAS, and running Immich as a Docker container on the same NAS, and adding the folder with the material as an External Library is probably too much to ask of either the Synology NAS or of Immich. The NAS still has plenty of free HDD space.
First, is Immich capable of managing this amount of material?
Second, what kind of hardware would be a good choice for this setup?

Having the files added through an External Library is necessary.

ChatGPT suggests:

Short answer: Immich itself will “run” on modest hardware, but for a smooth experience at your scale (≈750k assets; ~5 TB), you’ll be much happier if you size for three hotspots: (1) ML/face-&-object processing, (2) search (Typesense in-RAM index), and (3) fast local storage for Postgres + thumbnails.

Here’s a practical target that works well for libraries in the ~0.5–1.5 M asset range:

Recommended build (single box or VM host)

• CPU: 8 cores / 16 threads (modern Intel or AMD). This gives headroom for parallel thumbnailing/transcoding and ML jobs. Immich’s docs only require 2–4 cores, but that’s for small installs. ([docs.immich.app]1)

• RAM: 64 GB (32 GB minimum). Typesense keeps its index in RAM; real-world guidance is ~2–3× the size of the fields you index. At 750k assets with names/paths/tags/faces, 32–64 GB RAM prevents swapping and speeds up search. ([typesense.org]2)

• GPU (strongly recommended): NVIDIA with CUDA 12.3+ support (e.g., RTX 3060/4060 12 GB, RTX T4/A2000) or Intel Arc / iGPU via OpenVINO, or AMD via ROCm. Offloads face/object embeddings and can accelerate video transcode. ([docs.immich.app]3)

• System/storage layout:

  • NVMe SSD (fast, local): 1 TB for Postgres, Typesense data, and thumbs/ cache. Immich explicitly advises local SSD for Postgres (not a network share). ([docs.immich.app]4)
  • Bulk library storage: your 5 TB originals on HDD array/NAS is fine; mount to Immich as an external library.
  • Encoded videos: put encoded-video/ on fast storage if you transcode a lot, or keep on HDDs and accept slower first plays. (Hardware transcode can produce larger files than CPU with the same quality settings.) ([docs.immich.app]5)

What to expect for space & sizing overhead

• Postgres DB: usually a few GB (1–3 GB), but give it plenty of headroom (20–50 GB) for maintenance, migrations, and growth (there have been reports of larger DB use after upgrades). Keep it on local NVMe. ([docs.immich.app]4)
• Thumbnails (thumbs/): commonly ~7–9% of your photo data size (varies). For 5 TB, budget ~350–450 GB if you enable all sizes. ([Reddit]6)
• Transcoded videos (encoded-video/): highly variable by codec/settings; hardware transcode can be bigger than software at similar quality—tune policies and codecs in Admin → “Video Transcoding Settings”. ([docs.immich.app]5)

Deployment tips for large libraries

• Run on Linux (VM or bare-metal) with Docker; avoid Docker-in-LXC unless you know the caveats. ([docs.immich.app]1)
• Keep Postgres & Typesense on local SSD; mount your originals via NFS/SMB/ZFS dataset as an external library. Do not put Postgres on a network share. ([docs.immich.app]4)
• Tune ingestion: for first import of 750k assets, you can temporarily disable video transcoding and let thumbs/ML finish first; then enable transcode with the policy/codec/resolution you actually want. ([docs.immich.app]7)
• ML acceleration: enable CUDA/ROCm/OpenVINO per docs (driver + container toolkit for NVIDIA). This drastically speeds up face/object embedding on big sets. ([docs.immich.app]3)
• Typesense RAM watch: if searches feel sluggish or restarts are slow, it’s often RAM pressure from the Typesense index—add memory before CPU. ([typesense.org]2)
• Backups: back up Postgres dumps + profile/ + config paths; if you move file locations, update backup paths (see “Files Custom Locations”). ([docs.immich.app]8)

“Good / Better / Best” quick picks

• Good (budget, no GPU): 8c/16t CPU, 32 GB RAM, 1 TB NVMe (DB+thumbs+tsdata), HDD array for 5 TB library.
• Better (balanced): 8c/16t, 64 GB RAM, 1–2 TB NVMe, RTX 3060/4060 12 GB (or Intel Arc A380/A750).
• Best (headroom): 12–16 cores, 64–128 GB RAM, mirrored 2 TB NVMe for DB/cache, discrete GPU (NVIDIA 20–60-series w/12–24 GB VRAM) + HDD pool.

Reality check vs official minimums

Immich’s official minimums are tiny (2 cores, 4–6 GB RAM) and it can run on low-end boxes/NAS, but at your library size you’ll want the above to avoid slow imports, long ML jobs, and RAM pressure from Typesense. ([docs.immich.app]1)

[bo0tzz]

is Immich capable of managing this amount of material?

Yes

what kind of hardware would be a good choice for this setup?

It mostly depends on how much patience you have; you can do this on fairly minimal hardware (within reason), it'll just take much longer (potentially weeks) to process everything.

ChatGPT suggests:

LLMs aren't usually worth listening to about Immich stuff; there's a good bit of outdated nonsense in there, eg we haven't used typesense for a while.

[mbuzina]

My external library is about 1/3 your size running on a synology 925. I extended to 32gb memory, but the rest is plain. I use my desktop with a 4080 as machine learning for the initial load, it still took about a week. Now everything is extremely snappy, synology photos still running on the same box feels outdated slow by comparison.

I had to preload the models for quick search, but once that is in memory all is great

[webysther]

My immich on docker have:

services:
  immich:
    image: ghcr.io/imagegenius/immich:1.131.3-openvino
    deploy:
      resources:
        limits:
          cpus: '6.00'
          memory: 8G
    devices:
      - /dev/dri:/dev/dri
    group_add:
      - 6661
    device_cgroup_rules:
      - 'c 189:* rmw' # openvino

About same size of yours, using all as external library took about 1 week, my job settings:

Things important I always consider:

• I only use immich as read-only
• The network over linux use nfs over 10gbps in NAS and 2.5gbps on server
• The amount of CPU are TOTAL-2 to make sure immich will not stop the server to work
• I use iGPU intel
• The data directory of immich stay on server, not in NAS.
• limit the resources also to pgvecto because both drag lot of resources.
• My suggestion is to add the external folder in steps for what is more important, for instance, the more recent photos from mobile was ingested first and the last was the scan of old family photos.

If you have so many images and have RAW+JPEG, consider filter out the RAW to speed the process