Size the workload, not the hype.
Immich's official baseline is refreshingly clear: at least 6GB RAM and 2 CPU cores, with 8GB RAM and 4 cores recommended. Start there. Upgrade only for a workload you can name.
Linux or another 64-bit Unix-like OS is recommended. Immich supports amd64 and arm64. A smooth experience—especially during upload—requires at least 6GB RAM; 8GB is recommended. The official CPU recommendation is four cores.
What actually creates load
- Initial ingestion: thumbnail generation, metadata extraction, facial recognition and video transcoding arrive together.
- Video: large uploads and incompatible codecs create sustained transcoding work.
- Machine learning: face and semantic-search jobs can be moved to another machine; this can matter more than buying a larger always-on server.
- Concurrency: several people browsing while background jobs run changes the experience more than total library size alone.
A practical planning ladder
| Profile | Planning starting point | Why |
|---|---|---|
| Personal / small family | 8GB RAM, 4 modern cores | Matches official recommended baseline. |
| Large library or 10+ active users | 12–16GB, 6–8 cores | Planning judgment for overlapping jobs; validate under your workload. |
| Heavy video / large household | 16GB+, hardware acceleration or remote worker | Transcoding strategy matters more than raw core count. |
Do not put the database on slow network storage
Immich says stable Postgres performance is critical. Its database is typically 1–3GB and should ideally live on a local SSD, never a network share. If Docker resource limits are used, Postgres requires at least 2GB RAM.
Buying rule
Buy for the steady state, not the first import. Temporary ingestion can run overnight, use more workers, or use a second machine. If browsing remains responsive, a long job queue is not automatically a hardware failure.
按负载规划，不跟风堆配置。
Immich 官方基线很明确：至少 6GB 内存和 2 核 CPU，建议 8GB 与 4 核。先从这里开始，只有在能说清工作负载时才升级。
建议使用 Linux 或其他 64 位类 Unix 系统。Immich 支持 amd64 与 arm64。为保证流畅体验，尤其是上传期间，至少需要 6GB 内存，建议 8GB；CPU 建议 4 核。
真正产生负载的环节
- 首次导入：缩略图、元数据、人脸识别与视频转码会同时运行。
- 视频：大文件与不兼容编码会持续占用转码资源。
- 机器学习：人脸与语义搜索可移到另一台机器；这有时比购买更大的常开服务器更有效。
- 并发：多人浏览并同时运行后台任务，比媒体库总量更影响体验。
实用规划阶梯
| 场景 | 规划起点 | 原因 |
|---|---|---|
| 个人／小家庭 | 8GB，4 个现代核心 | 符合官方建议基线。 |
| 大型媒体库或 10+ 活跃用户 | 12–16GB，6–8 核 | 针对任务重叠的规划判断，需实测验证。 |
| 大量视频／多人家庭 | 16GB+，硬件加速或远程工作节点 | 转码策略通常比单纯增加核心更重要。 |
不要把数据库放在慢速网络存储上
Immich 指出，Postgres 的稳定性能非常关键。数据库通常为 1–3GB，最好放在本地 SSD，绝不能放在网络共享上。若设置 Docker 资源限制，Postgres 至少需要 2GB 内存。
采购原则
按长期稳定负载采购，而不是按首次导入的峰值采购。首次导入可以夜间运行、增加工作进程或借用第二台机器。只要浏览仍然流畅，队列较长不一定代表硬件不足。