The Question Everyone Asks Wrong
“What hardware do I need for Nominatim?”
The honest answer is “it depends entirely on which extract you import and how much patience you have.” But that’s not satisfying, so people shrug and either (a) try it on whatever box is sitting around, or (b) over-provision wildly and waste money.
Both are bad outcomes. The actual sizing math is not that hard once you separate the import from the runtime — they have totally different bottlenecks. Let’s get into the numbers so you can pick the right box without buying a server you don’t need.
This post assumes you’ve read the Nominatim install walkthrough, or at least know what the Compose stack looks like. We’re focused here on hardware, not setup.
The Two Phases That Have Nothing in Common
Nominatim has two distinct workloads, and they want different hardware:
- Import — runs once (per fresh data load). CPU-bound during parsing, disk-IO-bound during indexing, hungry for RAM and shared memory. Lasts hours to days.
- Runtime — query serving. Mostly disk-cache-bound, light on CPU, modest RAM. Lasts forever.
If you size for the import, you’ll waste a lot of resources during runtime. If you size for the runtime, the import will take a week. Most people end up biased toward the import side because the import is the painful, blocking experience that defines whether the project works at all.
The pragmatic compromise: pick a box that can do the import in 12–24 hours, then accept that it’s overkill at runtime. That’s fine. The runtime overhead of an underused server is rounding error.
Disk: The First Filter
Disk size is the easiest variable to plan for. It’s a hard floor — if you don’t have the bytes, the import fails.
Rough numbers for the imported database, with IMPORT_STYLE: full and the flatnode file enabled:
| Extract | Compressed PBF | Imported DB |
|---|---|---|
| Country (e.g., Germany, Italy) | 3–5 GB | 30–80 GB |
| Continent (e.g., North America, Europe) | 10–25 GB | 150–400 GB |
| Planet (full OSM) | ~80 GB | 700–1000 GB |
Add 20% headroom for growth from replication updates and Postgres bloat over time. Then add the OS and Docker overhead — call it 20 GB for a clean Linux install plus the mediagis image plus its caches.
Practical sizes:
- A 500 GB NVMe is comfortable for any single country and tight for North America.
- A 1 TB NVMe is comfortable for any continent and tight for the planet.
- A 2 TB NVMe is comfortable for the planet with growth headroom.
Disk speed matters at least as much as disk size. A SATA SSD will work but will turn a 12-hour import into a 30-hour one. Spinning rust is borderline unusable for the import — you can do it, but expect days. NVMe is the right answer if you can.
RAM: The Sneaky One
RAM requirements split between import and runtime.
For the import, Nominatim wants enough RAM to fit the working set of Postgres sorts and indexes. The mediagis image defaults expect at least 8 GB; for a country extract you can squeak by with that. For a continent, 16 GB is the practical floor. For the planet, 32 GB is the floor and 64 GB is the right answer if you want the import to finish in a reasonable timeframe.
There’s a flatnode file Nominatim uses to track raw nodes. For the planet, this file is around 75 GB and should be on the same fast disk, not crammed into RAM. mediagis sets this up correctly by default.
For runtime, you need much less. The actual query workload is mostly hitting hot indexes that Linux will cache from disk. 8 GB total system RAM is plenty for runtime even on a continent-scale import, because Postgres only really needs ~4 GB of buffer cache for typical query patterns. The kernel page cache does most of the work.
The pattern that bites people: import on a beefy temporary box (rented VPS, borrowed workstation), then move the data volume to a smaller permanent box for runtime. That’s actually a totally reasonable workflow if you’re trying to run on a Mini PC long-term.
CPU: Less Picky Than You’d Think
Nominatim’s import scales decently with CPU threads, up to a point. The THREADS env var controls how many parallel workers do the indexing. More threads = faster import, but with diminishing returns past 8 because Postgres serializes a lot of the work and disk IO becomes the bottleneck.
Sane CPU picks:
- 4 cores / 8 threads — fine for any country extract. ~6–10 hours for a country.
- 6–8 cores / 12–16 threads — fine for a continent. ~12–18 hours for North America.
- 12+ cores — useful for the planet. Even then you’re probably disk-bound before you’re CPU-bound.
CPU at runtime is essentially nothing. A single Nominatim query takes a few milliseconds of CPU. You can saturate a modern Mini PC at runtime only if you’re hammering it with thousands of requests per second.
The big gotcha: thermal throttling on consumer hardware. Mini PCs and laptops with weak cooling will throttle hard during a 12-hour import. Watch your temperatures. If the box drops to 60% clock speed because the fan can’t keep up, your 12-hour import becomes a 24-hour import. A USB desk fan blowing on the case is not a joke.
Reference Builds: What Actually Works
A few concrete profiles people run successfully.
Profile A: The Mini PC, Single Country
- N100 / Ryzen 5 mini PC
- 16 GB RAM
- 1 TB NVMe
- Country extract (e.g., a single European country, or one US state)
- Import time: 4–8 hours
- Runtime headroom: more than enough; this box can run other services too
This is the right answer for ~80% of self-hosters. You don’t need a continent if you only care about your own country.
Profile B: The Home Lab Workstation, Continent
- 6–8 core CPU (older Xeon, modern Ryzen)
- 32 GB RAM
- 1 TB NVMe (with 1.5 TB+ if you want growth room)
- Continent extract (North America, Europe)
- Import time: 12–24 hours
- Runtime headroom: comfortable; can serve multiple apps
This is the sweet spot for people who want continental coverage without going overboard.
Profile C: The Used Server, Planet
- 12+ core CPU (used Xeon, EPYC if you can find it cheap)
- 64 GB RAM
- 2 TB NVMe
- Full planet
- Import time: 1–3 days
- Runtime headroom: absurd; this is overkill for runtime
If you actually need the planet — and most people don’t — used enterprise gear is the most cost-effective path. A 5-year-old dual-Xeon with 128 GB and a couple of NVMe drives is often cheaper than a new mid-range desktop.
Profile D: The Don’t-Do-This-Trap
- Raspberry Pi 4 / 5
- 8 GB RAM
- USB SSD
- Country extract
Honestly, you can technically pull off a single small country extract on a Pi. The import is painful (24+ hours for a small country) and the runtime is acceptable. But you’re better off spending the same money on a used Mini PC that imports in 4 hours and runs everything else you’d want too.
Network: The Forgotten Variable
Pulling the PBF takes time. Geofabrik publishes from EU mirrors at decent speeds, but:
- A 25 GB North America PBF over a 100 Mbit residential link is ~30+ minutes
- A planet PBF (~80 GB) over the same link is multiple hours
If you have flaky internet, run the import on a box with a wired gigabit connection or download the PBF in advance and mount it into the container. The mediagis image accepts a local PBF path via PBF_PATH instead of PBF_URL.
Replication: Ongoing Cost
After the initial import, daily replication updates pull diffs from Geofabrik. These are tiny — usually single megabytes per day for a country, tens of MB for a continent. Disk growth is gradual: budget for ~10–20% data size growth per year if you keep replication running indefinitely.
Periodic vacuum and reindex helps keep the database lean. The mediagis image runs autovacuum by default, but if you’ve been running for a year you might want to do a manual VACUUM FULL during a maintenance window.
Practical Buying Guide
Cheapest viable build for “I want to play with Nominatim and learn”:
- Used Mini PC (N100, Beelink, similar) — $200-ish
- 16 GB DDR5 — already in most Mini PCs at this price
- 1 TB NVMe — $50–80
- A regional extract (your country)
That’s a $300 build that runs Nominatim plus a half-dozen other home lab services without breaking a sweat.
If you want continental coverage, jump to a used SFF workstation or a tower with 32 GB RAM and a real NVMe. Plenty of refurbished Dell OptiPlex / HP EliteDesk options under $500 with the specs you want.
If you want the planet — really, do you? If yes, used 1U enterprise servers from eBay are the best dollar-per-spec, with the caveat that they’re loud and power-hungry. Otherwise rent a beefy VPS for the import only, then move the data volume to a smaller permanent box.
Things That Will Bite You
- Underestimating disk speed. SATA SSD on paper is “fine” but in practice doubles or triples the import time. NVMe matters more than you think.
- Running other heavy stuff on the same box during import. Don’t run Plex transcodes during a Nominatim import. They’ll fight for IO and your import will crawl.
- Tight
shm_size. Setting Docker’s shared memory to less than 1 GB causes mysterious OOMs during indexing. Don’t. - Forgetting the flatnode file. It needs ~75 GB for the planet. mediagis configures this by default but if you’re rolling your own, easy to miss.
- Replication on too tight a timer. Hourly replication on a busy region creates avoidable IO churn. Daily is plenty for most use cases.
- Power loss mid-import. A UPS is not optional if your power is even slightly flaky. A failed import means starting over.
Wrapping Up
For 80% of people the right answer is a 16 GB Mini PC with a 1 TB NVMe and a country extract. For the next 15% it’s a 32 GB box and a continent. The 5% who need the planet should be using used enterprise gear and doing the math seriously.
The biggest single mistake I see: people sizing for the planet on consumer hardware because “it’s only 80 GB compressed.” It’s not 80 GB; it’s 700–1000 GB imported, and the import itself is a 1–3 day commitment that wants real RAM and real disk. If you don’t actually need the planet, don’t import it.
Pick the smallest extract that covers your needs. You can always import a bigger one later. That’s the whole strategy.
Related posts
- Nominatim: Self-Hosted Geocoding — Docker install walkthrough
- Nominatim vs Photon vs Pelias — choosing the right engine
- Reverse Geocoding for Home Assistant — privacy-friendly device tracking
- Full Self-Hosted Maps Stack — Nominatim + PostGIS + tiles
- Mini PCs for Home Lab: N100 vs N305 vs Ryzen 7000 — pick a host
