How Much RAM and CPU Does n8n Need? We Tested It Ourselves

We ran n8n 2.44.6 on a Lenovo M720q we already had at home, inside an isolated Debian container with 2 CPU cores and 4 GB of RAM. n8n itself ran in Docker with a 3 GB limit. It’s not a rented VPS, so take the numbers as a guide to how different workflows behave, not as a promise about any provider.

What to start with

These are starting points, not guaranteed minimums. We measured on a 4 GB machine and never repeated the tests with strict 1 GB or 2 GB limits. The memory figures are for the whole machine (Debian, Docker and n8n together), because that’s what a VPS plan has to cover.

Simple workflows barely need anything

Our first workflow read four RSS feeds, removed duplicates and sent a daily email through Gmail. Idle, the n8n container used about 408 MiB. Running the workflow, it peaked at 447 MB, and CPU stayed low.

Our first benchmark said 854 MB and 182% CPU for that same workflow. It looked wrong, and it was: launching it with “n8n execute” starts a second n8n process inside the container, so we were measuring two instances. Triggering it through the server that was already running gave the 447 MB. If you benchmark n8n yourself, check how your test launches the workflow.

Most of our non-video tests barely used a second core

Apart from video, everything we ran used little CPU or mostly one core. The email workflow sat at 4% on average. The 100,000-item job averaged about 104%, which is one core out of two. So for these jobs, a fast single core may matter more than extra vCPUs. We only tested with two cores available, so a strict single-core run is still pending.

Records: RAM matters

We generated records inside n8n, transformed them, filtered them and summarized them by category.

Memory grew by roughly 10 MB for every extra 1,000 records. We also tried Split in Batches, and it ran slower without saving memory, because we had already generated all the records before batching them. It should work better when data arrives in pages, from a paginated API or a database query with a limit.

Video: this is where you want cores

We converted a 30-second clip with FFmpeg called from n8n (the official image doesn’t include it, so we built our own).

200% means both cores fully busy, and FFmpeg got very close. H.265 needed about twice the CPU work of H.264, though we used different quality settings, so this says nothing about compression efficiency.

A full 10-minute video at 720p took 6 minutes 23 seconds and peaked at 761 MB inside the container, less than the short 1080p test. In our tests, longer videos didn’t need much more RAM. They need more CPU time, about 67 CPU seconds per minute of video. If you convert a lot of video, look at CPU first.

Watch the disk

While we ran the large tests with execution history saved, n8n’s SQLite database grew from 13.5 MB to 380.8 MB. We deleted the executions afterwards and the file kept its size. Not saving successful executions barely changed RAM (1,737 MB against 1,641 MB), but it stopped the database from growing. If you run workflows often, decide how much history you need.

What n8n’s documentation says

n8n’s own documentation describes it as not CPU intensive and says memory usually matters more when sizing. Its sizing example, based on n8n Cloud and given for illustration only, lists 320 MB to 2 GB of memory and 512 MB to 4 GB of SSD storage for the database. n8n sizing example ↗ Memory use grows with the amount of JSON and binary data a workflow handles, Code nodes, manual executions and other workflows running at the same time. n8n memory documentation ↗

What we didn’t test

One machine, three runs per test, no concurrency, no strict 1 GB or 2 GB limits, no single-core runs, no PostgreSQL. A shared cloud vCPU may be slower than our i5-8500. The videos are Blender Foundation productions under Creative Commons Attribution licenses. No hosting provider sponsored the tests.

Compare real plans for these sizes

The catalog sizes below are VPS Cost Lab’s editorial starting points; they are not n8n minimums, benchmarks or capacity guarantees, and the catalog has no plans listed below 2 GB. Plans are ordered by published price only; affiliate status never changes the order.

Starter: at least 2 GB RAM, 1 vCPU, 20 GB storage

One person or a small team running scheduled automations, with SQLite or a small PostgreSQL database on the same server. The three plans below are the lowest verified published prices in the catalog that meet every value, one per provider. A plan with more resources than the size asks for can still be the cheapest.

No long commitment: the cheapest plan of this size that can be cancelled monthly is UpCloud · Starter 2 GB / 1 core at $7.00/month.

Filter every catalog plan of this size in the comparison tool ↗

Recommended: at least 4 GB RAM, 2 vCPU, 40 GB storage

n8n, PostgreSQL and a reverse proxy in Docker on one server, with several active workflows, some Code nodes and occasional files. The three plans below are the lowest verified published prices in the catalog that meet every value, one per provider. A plan with more resources than the size asks for can still be the cheapest.

No long commitment: the cheapest plan of this size that can be cancelled monthly is Hostwinds · Unmanaged Linux 4 GB at $18.99/month.

Filter every catalog plan of this size in the comparison tool ↗

Room to grow: at least 8 GB RAM, 4 vCPU, 80 GB storage

Many concurrent executions, larger JSON or binary payloads, or queue mode with Redis and separate workers on the same machine. The three plans below are the lowest verified published prices in the catalog that meet every value, one per provider. A plan with more resources than the size asks for can still be the cheapest.

No long commitment: the cheapest plan of this size that can be cancelled monthly is Hostwinds · Unmanaged Linux 8 GB at $38.99/month.

Filter every catalog plan of this size in the comparison tool ↗

Read the price before you order

The lowest monthly figures often come with a 12- or 24-month term, an introductory rate or an unverified renewal; each row says which. A plan you can cancel monthly can cost more per month and less in total if your project is short. Taxes, backups and extra IPv4 addresses can add to the bill.

Calculate the full cost of a plan ↗ · Compare up to four plans side by side ↗ · See which extras change the total ↗

Running n8n on a self-managed VPS

On a self-managed plan you keep the operating system patched, secure SSH and the firewall, put n8n behind HTTPS, update n8n and its database, and back up the database and n8n’s configuration, not only the server disk. Test a restore before you rely on it. If the server runs out of memory, executions fail; give memory-heavy workflows room or move to a larger size.

About the price tables

The catalog is a dated research snapshot, not a live quote. Region availability, taxes, promotions and renewal terms need confirmation at checkout. Published resource units are kept as the provider states them; GB and GiB are not converted.