Both boxes run the same OS stack (Debian-based NAS distro, Docker), so the only variables are the drive and the switch port.

Box A — 2011 laptop: Celeron B815 (2C/2T, 1.6 GHz), 4 GB DDR3, Realtek 100 Mbps, 256 GB SATA SSD + 320 GB 5400 rpm in the optical bay. Box B — 2011 mini PC: i5-2520M, 8 GB DDR3, gigabit, 128 GB SSD + 3 TB WD Red Plus.

Same library, same client:

  • Immich first load: 24 s on the 5400 rpm 2.5-inch, 0.25 s after moving the DB + UI to the 256 GB SSD. ~95x, on identical hardware.
  • SMB to Box A: 11.4 MB/s up, 11.3 MB/s down (85-91 Mbps — the link is saturated, the disk is not the limit).
  • hdparm -t on the SATA SSD: 506 MB/s.
  • WD Red Plus sequential write: 107 MB/s. 5400 rpm 2.5-inch buffered read: 88 MB/s.
  • Box B idle 20 W, under sustained load ~40 W.

The takeaway that surprised me: the photo blobs are fine on spinning rust. What kills perceived speed is putting the Postgres DB and the thumbnail cache on a 5400 rpm drive — every page load turns into a scatter of small random reads. Move those two, leave the originals where they are.

The other half is the network. Once the SSD was in, the 100 Mbps port became the bottleneck for everything except thumbnail generation, which is pure CPU (about 91 thumbnails a minute, so a 12k library is roughly two hours of solid work).

I keep the full numbers, and the method used to get them, here: lowspechq.com/best/cheap-ssd-for-nas/ Source for the site builder if you want to check the claims against the code: github.com/lowspecai/lowspec-ai

Disclosure: I run lowspechq.com. That page carries one clearly labelled AliExpress affiliate link (an SSD listing) and nothing else; every number above was measured before any link existed.