toolchest_

Measured on a real homelab

Immich RAM & CPU requirements (measured)

On our homelab it uses about 714 MB of RAM on a normal hour, 1,008 MB in busy hours, and peaked at 4.3 GB over the last 30 days (one household: 11,520 photos + 4,673 videos (46 GB); machine learning loads its models on demand).

What it really uses (process memory)

Typical (median)
714 MB
Average
710 MB
Busy hours (95th percentile)
1,008 MB
Highest recorded
4.3 GB
As docker stats shows it (with disk cache)
median 1.2 GB · peak 6 GB
Window
720 hours measured, up to Sep 26, 2026
Version
3.2.2
Stack (all summed)
server + machine-learning + PostgreSQL + Valkey
Workload
one household: 11,520 photos + 4,673 videos (46 GB); machine learning loads its models on demand
Hardware
Intel i5-10600K · 23 GB RAM · Docker (Coolify)

Caveat (Oct 2, 2026): these figures count process memory held in RAM. When the server is under memory pressure, idle apps get part of their memory swapped out and it was not being counted, so the "typical" figure may be too low; the peak is reliable. We now measure RAM plus swap and will update these numbers.

What makes it grow

Thumbnails and transcoded video add 10-20% on top of the original media library size, and very large libraries (750k+ assets) benefit from significantly more RAM and a GPU for ML processing.

Sizing it

Our rule of thumb: plan for the highest process memory recorded plus 25% headroom: 5.3 GB. Our workload is small — scale it with the factors above before buying hardware.

Combine it with other apps in the resource calculator →

Build it: Rent a VPS · Buy a mini-PC

Self-hosting photos, files or passwords? Keep an off-site copy (3-2-1 rule): pCloud lifetime backup

How we measure

Memory per container from Netdata on the same server, summed across the whole stack (app, database and cache) hour by hour. The main figures are the processes' own memory (anonymous memory: heap, stacks). The docker stats figure is higher because it also counts active disk cache — which Linux gives back when it needs it — and shared memory such as PostgreSQL's shared buffers. Hours where the main container had no data are left out, not counted as zero. Re-measured with a script, not by hand.

Other measured apps