Best Immich Server Build 2026: Self-Hosted Photo & Video Backup

If you are planning a self-hosted replacement for Google Photos or iCloud in 2026, the best Immich server build balances storage capacity with the hardware needed for its demanding machine learning features. Immich can run on surprisingly modest hardware for basic photo browsing, but its facial recognition, object detection, and smart search require a significant CPU or GPU investment. This guide breaks down the exact hardware requirements for each feature tier, then walks through three complete builds — budget, mid-range, and GPU-accelerated — so you can match your server to your photo library and budget.

Base Hardware Needs for Photo and Video Storage

Immich, at its core, is a web application that stores and serves your media files. The base requirements for simply uploading, browsing, and downloading photos are moderate. You need a reliable CPU, enough RAM to keep the stack running, and sufficient storage for your photo library.

The CPU handles the web server, database queries, and image thumbnail generation. A modern quad-core processor like an Intel N100 or an older i3-8100 is perfectly adequate for a single user or a small family library. For video thumbnails and transcoding, Intel Quick Sync Video (on 7th-gen or newer CPUs) provides excellent hardware acceleration without needing a discrete GPU.

RAM is more important than you might expect. Immich runs as a Docker stack with multiple containers: the main server, a Postgres database, a Redis cache, and the machine learning container (if enabled). Postgres alone can consume 1-2 GB of RAM with a moderate library. For a basic setup without ML features, 8 GB of RAM is the realistic minimum for a smooth experience.

When You Need Extra CPU/GPU for Immich’s Machine Learning

Immich’s standout features — automatic face grouping, object/tag recognition, and smart search — are powered by machine learning models. These are computationally expensive. Running them on a CPU is possible but slow, especially for large libraries.

If you plan to use facial recognition and object tagging, you have two paths:

• CPU-only (budget): A modern 6-core or 8-core CPU (like an Intel i5-12400 or Ryzen 5 5600) will work, but expect slow initial processing and higher CPU usage during library scans. Fine for small libraries under 5,000 photos.
• GPU-accelerated (recommended for ML): An NVIDIA GPU with CUDA cores dramatically accelerates inference. An NVIDIA T4 (16 GB VRAM) or an RTX 3060 (12 GB VRAM) are excellent choices. Avoid AMD GPUs for now — Immich’s ML pipeline is optimized for CUDA.

RAM Sizing for the Immich Stack

The Immich Docker stack is more RAM-hungry than a typical media server like Plex or Jellyfin. Here is a realistic breakdown of RAM usage per component:

For a basic setup (no ML): 8 GB is the minimum, but 16 GB is strongly recommended for headroom. For an ML-enabled server: 16 GB is the practical minimum; 32 GB is comfortable for libraries over 50,000 photos. Postgres benefits greatly from extra RAM for caching frequently accessed metadata.

Storage Growth Planning for Photo and Video Libraries

Photo libraries grow faster than most people estimate. A single user with a modern smartphone can generate 50-100 GB of photos and videos per year. A family of four can easily hit 300-500 GB annually, especially with 4K video recording.

Immich stores original files on disk and generates thumbnails and compressed versions. The thumbnails add about 5-10% overhead on top of the original file sizes. For a 5-year plan, size your storage accordingly.

For RAID, a two-drive RAID 1 mirror gives you fault tolerance but halves raw capacity. A four-drive RAID 5 or RAIDZ1 gives three drives usable — good for capacity-focused builds. Remember: RAID protects against drive failure, not accidental deletion or ransomware. RAID is not a backup. Maintain a separate backup of your Immich library to an external drive or cloud service.

For a deeper dive on sizing your storage, see our guide on How Much Storage Do You Need for a NAS?

Three Build Tiers for the Best Immich Server Build in 2026

Below are three complete builds covering different budgets and feature requirements. Each tier specifies CPU, RAM, storage, GPU (if needed), and approximate cost.

Budget Build: CPU-Only for Basic Browsing

This build uses a low-power Intel N100 mini PC. It handles basic photo upload, browsing, and thumbnail generation without issue. It can run Immich’s ML features, but expect very slow processing — hours for a few thousand photos. Ideal for a single user who just wants self-hosted photo backup without smart search.

Mid-Range Build: i5 with Quick Sync

An Intel i5-12400 with 16 GB of RAM and a 4-bay NAS case. Quick Sync handles video transcoding and thumbnail generation. The CPU can run ML tasks at a moderate pace — expect 4-8 hours for a 10,000-photo library. This is the sweet spot for most homelabbers who want decent performance without a discrete GPU.

For power efficiency tips on this build, check out our guide on How to Reduce Your NAS’s Power Consumption.

ML-Accelerated Build: GPU for Full Features

This build pairs a high-core-count CPU with an NVIDIA RTX 3060 (12 GB VRAM) for fast ML inference. Face detection and object tagging complete in minutes, not hours. The 32 GB of RAM ensures Postgres and the ML container never compete for memory. This is the best Immich server build for families or power users who want the full Google Photos experience at home.

Mobile App Auto-Backup Considerations

Immich’s mobile app provides background photo and video backup, similar to Google Photos. This feature does not directly affect server hardware requirements because the upload process is light on the server side. However, consider these indirect effects:

• Network bandwidth: Multiple users backing up large 4K videos simultaneously can saturate a 1 Gbps link. A 2.5 Gbps NIC is a worthwhile upgrade for family setups.
• Storage write speed: SSDs handle concurrent writes better than HDDs. If your family uploads 50+ photos per day, an SSD cache or all-SSD storage improves responsiveness.
• Thumbnail generation: Each uploaded photo triggers thumbnail generation. A faster CPU or GPU reduces the lag between upload and availability in the web UI.

Which Build Should You Choose?

For most homelabbers replacing Google Photos in 2026, the mid-range build with an i5-12400 and 16 GB of RAM offers the best balance of cost, performance, and power efficiency. It handles a family library of 10,000-20,000 photos with acceptable ML speed, and Quick Sync handles video transcoding without a GPU. If you have a large library (50,000+ photos) and want fast face recognition, invest in the ML-accelerated build. The budget N100 build is a capable entry point for single users who prioritize low power over ML features.

For more on choosing between operating systems for your Immich server, read our comparison of TrueNAS vs Unraid. If you are building a TrueNAS-based server, our Best CPU for TrueNAS guide will help you pick the right processor.