Don't Overspend on Silicon: The Truth About Immich Hardwa…

The dream of a self-hosted, private Google Photos alternative is a cornerstone of the modern homelab movement. Immich has emerged as the gold standard for this endeavor, offering a feature-rich interface that handles everything from facial recognition to sophisticated "Smart Search" via CLIP-style embeddings. However, as users begin to migrate tens of thousands of high-resolution images into their private clouds, a technical panic begins to set in.

The prevailing narrative in the AI era is simple: If you want to do AI, you need a GPU.

While this is often true for training large language models or running heavy generative diffusion pipelines, applying this logic blindly to Immich can lead to a massive waste of capital. Before you go shopping for a dedicated NVIDIA card, you need to understand the distinction between the two most important metrics in machine learning: throughput and latency.

Understanding the Immich Workload

Immich isn't just a photo gallery; it is a continuous machine learning pipeline. When you upload a photo, the system doesn't just save a file; it triggers several intensive processes:

Facial Detection and Recognition: Identifying human faces and grouping them into identities.
CLIP Embeddings: Generating mathematical representations of your images so you can search for concepts like "dog on a beach" or "sunset over mountains."
Metadata Extraction: Parsing EXIF data and other technical information.

These tasks are computationally expensive. However, the way these tasks impact your user experience depends entirely on your ingestion habits. This is where most users make their "expensive mistake."

Throughput vs. Latency: The Great Decoupling

To make an informed hardware decision, you must distinguish between how fast the system processes a mountain of data (throughput) and how fast it responds to a single request (latency).

Throughput: The Initial Indexing Hurdle

Throughput refers to the volume of work completed in a given timeframe. When you first install Immich and point it at a 500GB directory of legacy photos, your server enters a high-throughput phase. The machine learning containers will run at 100% utilization for hours, or even days, as they attempt to index every single image. A GPU is a powerhouse here; it can process batches of images simultaneously, slashing the time it takes to reach a "steady state."

Latency: The Daily User Experience

Latency is the time elapsed between a user action and the system's response. Once your library is fully indexed, your daily workload changes. You might upload 20 new photos a day. At this stage, you aren't looking for massive throughput; you are looking for the system to quickly process those 20 photos and allow you to search your existing library.

Crucially, a high-end GPU does not necessarily make your search queries feel "snappier." Searching an existing database of embeddings is largely a database lookup task, not a heavy compute task. If your goal is to make the search bar respond faster, a GPU might be a lateral move rather than a vertical upgrade.

The Hardware Tiers: From Budget to Beast

If you are auditing your hardware, you likely fall into one of three categories.

1. The CPU Baseline (The "Wait it Out" Strategy)

You do not need a GPU to run Immich. It is entirely possible to run the machine learning containers on a standard CPU. However, there are strict minimum requirements to ensure the ML containers don't crash or stall your entire system.

Minimum Requirements: At least 2 CPU cores and 6GB of RAM.
Recommended Specs: 4 cores and 8GB of RAM for a smoother experience.
The x86-64-v2 Trap: This is a critical technical detail. Immich 3 requires the x86-64-v2 instruction set. If you are trying to run this on an ancient server or a very low-end Celeron processor that lacks these instructions, the ML containers will fail to launch regardless of whether you have a GPU attached.

2. Intel iGPU (The "Efficiency" Play)

If you are running a modern Intel-based NUC or a consumer desktop, you may already own a powerful tool: the integrated GPU (iGPU). Through the use of OpenVINO, Immich can offload machine learning tasks to Intel’s integrated graphics.

This is often the "sweet spot" for homelabbers. It provides a significant boost over pure CPU processing without the power draw, heat, or cost of a dedicated card. It is an excellent way to increase throughput during initial indexing while keeping your power bill low.

3. NVIDIA GPU (The "Power User" Play)

NVIDIA remains the king of AI acceleration due to its CUDA ecosystem. If you use the NVIDIA container toolkit, Immich can leverage CUDA to accelerate the embedding and face detection processes.

When should you actually buy an NVIDIA card?

You are constantly importing massive datasets (e.g., you are a professional photographer or a digital archivist).
You are running Immich on a shared server where other users are also demanding compute resources.
You have a massive backlog that, if left to the CPU, would take weeks to process.

If you only add a few dozen photos a day, even a budget card like an RTX 3050 LP is likely an unnecessary expense.

Common Pitfalls and Technical Confusion

Before you finalize your build, beware of two common misconceptions that lead to wasted hardware spend.

The Transcoding Confusion

Many users see "hardware acceleration" mentioned in Immich documentation and assume it refers to video. This is a mistake. While Intel Quick Sync or NVIDIA NVENC are vital for video transcoding (making sure your videos play smoothly on various devices), they are entirely separate from ML acceleration. A card that is great at transcoding might not be the most efficient for CLIP embeddings, and vice versa. Ensure you are solving for the specific bottleneck you are experiencing.

The Benchmark Mirage

It is important to note that there is currently a lack of controlled, cross-platform benchmarks comparing specific hardware speeds for Immich. While older comparisons between OpenVINO and CPU exist, they may be outdated. Immich is evolving rapidly, and recent optimizations may have changed the performance delta between these hardware tiers. Do not rely on "old" data to justify a "new" purchase.

Summary: A Decision Matrix for Immich Users

To decide your path, audit your ingestion habits using the following guide:

Conclusion

The most important rule of budget AI builds is to avoid solving a one-time problem with permanent hardware. If your only "problem" is that your initial 10,000-photo import took three days to index, the solution isn't a $300 GPU—it's patience. Only invest in dedicated silicon when your daily workflow demands the throughput that only a discrete accelerator can provide.

# Quick check for x86-64-v2 compatibility on Linux
/lib/ld-linux-x86-64.so.2 --help | grep -E "v2|v3|v4"

---

### Sources

- [Does Immich need a GPU? CPU vs Intel Graphics vs NVIDIA for private AI photo search](https://www.popularai.org/p/does-immich-need-gpu) — Popular AI