By n8n's own figures, n8n needs little memory and even less processor power. Its cloud runs the Starter plan on 640 MiB of memory, and the software alone uses about 180 MiB on average.
On your own server, add room for the operating system, the database and the files your workflows handle. Our reading: 2 GB is the floor, 4 GB suits most small teams, and 8 GB fits file-heavy or queue-mode set-ups.
These figures come from n8n's documentation and from a load test that n8n ran on AWS servers.
Who this is for
- you are about to rent a server for n8n and want to pick its size
- your n8n server stops with out-of-memory errors or slows down under load
- you are weighing queue mode and want to know what each part needs
Not for you if you use n8n Cloud: n8n sizes the server for you, and your plan sets the memory.
Key numbers
- 640 MiB memory on n8n Cloud's Starter plan, the same as on its trial Official page,
- 180 MiB what the n8n software alone uses on average, by n8n's Cloud documentation Official page,
- 4 GB memory of the smaller AWS server in n8n's own load test Official page,
- 220 per second workflow runs a single n8n instance can handle at most, by n8n's documentation Official page,
What n8n's documentation says
n8n gives two sets of sizing figures. The first is a table for companies that build n8n into their own product. The company calls it illustrative and says it is based on n8n Cloud.
Plan around memory: n8n says its memory needs usually come before its processor needs.
| Part | Size in n8n's table | Supported |
|---|---|---|
| Memory | 320 MB to 2 GB | any server |
| Database | 512 MB to 4 GB, on SSD | SQLite or PostgreSQL |
| Processor | at least 10 CPU cycles, more as needed (n8n's wording) | any public or private cloud |
The same page says an idle n8n Cloud instance needs about 100 MB. What raises memory use is the kind of workflow and the data it handles.
The second set is the memory and processor share of each n8n Cloud plan. These are the limits n8n runs its own paying customers on.
Even the plan for 50,000 executions a month runs on 1,280 MiB, about 1.3 GB.
| n8n Cloud plan | Memory | Processor share |
|---|---|---|
| Trial | 640 MiB | 10 millicores, burstable |
| Starter | 640 MiB | 10 millicores, burstable |
| Pro-1 (10,000 executions) | 640 MiB | 20 millicores, burstable |
| Pro-2 (50,000 executions) | 1,280 MiB | 80 millicores, burstable |
| Enterprise | 4,096 MiB | 80 millicores, burstable |
A millicore is a thousandth of a processor core. Burstable means the instance can use more when the machine has power to spare. A MiB is slightly more than a megabyte: 640 MiB is about 0.67 GB.
What uses up the memory
n8n puts no limit on how much data each step can fetch and process. When a run needs more memory than the server has, n8n stops it or restarts. Its documentation names these causes:
- the amount of data a workflow handles, as JSON
- files such as images and PDFs, which n8n keeps in memory unless you change its storage mode
- the number of steps (nodes) in a workflow
- the Code node, which makes copies of the data before and after it runs
- runs you start by hand in the editor, because n8n keeps an extra copy for the screen
- other workflows running at the same time
AI steps follow the same rules. When the model runs at an API provider, n8n handles the text going in and out, not the model itself. Long chat histories and large documents sent to a model still count as data in n8n.
What n8n suggests when memory runs short
- process data in smaller batches, for example 200 rows a run instead of 10,000
- use the Code node less
- avoid manual runs on large amounts of data
- split big workflows into sub-workflows that return only a small result
- store files on disk with
N8N_DEFAULT_BINARY_DATA_MODE=filesystem; in queue mode, usedatabaseinstead - or give n8n more memory: a bigger server, or a higher Node.js limit through
NODE_OPTIONS=--max-old-space-size=...
On a server you run yourself, the out-of-memory message in the logs reads "JavaScript heap out of memory". The Docker image of n8n restarts on its own after such a crash; an npm install may need a manual restart.
The processor: why a small one is enough
n8n says it is not CPU intensive, so even small cloud servers suit most uses. Its memory needs usually come first.
The processor matters when many runs arrive at once. In regular mode, n8n puts no limit on parallel production runs, and too many can make it slow or unresponsive.
Set N8N_CONCURRENCY_PRODUCTION_LIMIT, for example to 20, and n8n queues the extra runs in the order they arrived. The limit covers runs started by a webhook or trigger, not runs you start by hand.
n8n's own load test, on a 4 GB server and a 32 GB one
n8n tested its software on two AWS servers in a blog post dated 24 September 2025. The smaller c5.large had 4 GB of memory, which n8n describes as 1 vCPU with 2 threads.
The larger c5.4xlarge had 16 vCPUs and 32 GB. Each test raised the load from 3 to 200 virtual users. The post does not name the n8n version, and calls regular mode "single mode".
Queue mode raised what both servers could take; large files needed the big server as well.
| Test | Server | Regular mode | Queue mode |
|---|---|---|---|
| One workflow on a webhook | c5.large, 4 GB | held up to 100 users; at 200, answers took up to 12 s and 1% failed | 72 requests a second, under 3 s, no failures at 200 users |
| One workflow on a webhook | c5.4xlarge, 32 GB | 16.2 requests a second | 162 requests a second, under 1.2 s, no failures |
| 10 workflows at once | c5.large, 4 GB | at 50 users, over 14 s and 11% failed; at 200 users, 34 s and 38% failed | 74 requests a second from 3 to 200 users, no failures |
| 10 workflows at once | c5.4xlarge, 32 GB | at most 23 requests a second, with 31% failed | 162 requests a second, no failures, about 5.8 s at the heaviest load |
| Large file uploads | c5.large, 4 GB | 3 requests a second at 3 users; 74% failed at 200 | 87% failed at 200 users |
| Large file uploads | c5.4xlarge, 32 GB | 4.6 requests a second, 11% failed | 5.2 requests a second, no failures |
Read these as limits, not daily loads. A load of 200 users at the same moment is far beyond what most small teams send. The rows that matter for a small server are the last two: large files fail first.
n8n's documentation adds a second figure: a single n8n instance can handle up to 220 workflow runs a second. Its test for that figure used a 4 GB server and a two-step workflow.
The rest of the server: database, Redis and disk
n8n keeps workflows, credentials and past runs in its database. SQLite is the default, and PostgreSQL is the other choice; queue mode needs PostgreSQL.
By its own documentation, PostgreSQL sets aside 128 MB for its shared buffers by default. Redis, which queue mode uses, takes about 3 MB when empty, by the Redis documentation.
Disk use grows with saved runs. Unless you change it, n8n deletes finished runs after 14 days, or once there are more than 10,000. Once a workflow works, you can stop saving its successful runs.
On SQLite, the space of deleted runs is reused but not given back; set DB_SQLITE_VACUUM_ON_STARTUP to free it. The sizing table above gives 512 MB to 4 GB of SSD for the database, and most plans we track include 20 GB or more.
Queue mode: when you need it, and what each part does
In queue mode, one main n8n instance takes the triggers and webhooks. Redis holds the queue, and workers run the workflows. Each worker runs up to 10 jobs at once by default.
- Triggers and webhooks (schedules, forms, app events): connects to Main n8n instance (webhook).
- Main n8n instance (takes each trigger, serves the editor): connects to Redis (run ID).
- Redis (the queue of waiting runs): connects to Workers (next job).
- Workers (run the workflows, 10 at a time each): connects both ways with PostgreSQL (reads and writes).
- PostgreSQL (workflows, credentials, results)
Queue mode pays off when many runs arrive together, as n8n's test shows, or when one heavy workflow must not slow the rest. It also asks more of you: PostgreSQL, Redis and the workers all need running and updating.
- it needs PostgreSQL: n8n does not support queue mode on SQLite
- files cannot be kept on disk; use the
databasemode instead - the Simple Memory node for AI chats does not work in active workflows, because a chat's next message may reach another worker
- n8n suggests a worker concurrency of 5 or more, since low values with many workers can run out of database connections
- on n8n Cloud, queue mode comes only with Enterprise plans
Redis has no password by default, so keep it off the internet. Our guide to securing self-hosted n8n has the other steps.
n8n's own test ran queue mode on the 4 GB server. With PostgreSQL and heavier workflows on the same machine, we would start at 8 GB.
Which server size to pick
This table is our reading of the figures above, not a measurement. According to n8n, the nature of your workflows sets the memory you need, so watch the server's memory use after the first week.
Most plans we track start at 4 GB, which covers the first two rows.
| Your use | Memory | Processor | Why |
|---|---|---|---|
| Trying n8n with a few light workflows | 2 GB | 1 core | n8n Cloud runs its Starter plan on 640 MiB; the rest is for the system and the database |
| A small team: daily workflows, AI steps through a model API | 4 GB | 2 cores | room for PostgreSQL, bursts and runs started in the editor |
| Files, PDFs, many runs at once, or queue mode on one server | 8 GB | 2 to 4 cores | files sit in memory unless stored on disk; queue mode adds Redis and workers |
| A local AI model on the same server | 8 GB plus the model's own size | 4 cores or more | the model loads into memory next to n8n |
Our guide to how much memory an AI model needs gives the model sizes. On a server without a graphics card, the model sits in ordinary memory instead.
| Plan | Price | New-customer offer | Link |
|---|---|---|---|
| Hostinger KVM 1 1 vCPU, 4 GB | $6.49/month 24-month term, paid up front; Black Friday first-term price, then $11.99 checked 25 Oct 2026 | Black Friday prices on VPS plans and AI automation apps, for the first subscription term only checked 30 Oct 2026 | Visit |
| DigitalOcean Basic Droplet, 2 vCPUs, 4 GiB 2 vCPU, 4 GB | $24/month billed per second (60-second minimum), capped at this monthly price · VAT not stated checked 25 Oct 2026 | New accounts get a $5 credit that lasts 90 days and covers every product except SaaS add-ons checked 29 Oct 2026 | Visit |
| Kamatera Cloud server Type B (General Purpose), 2 cores, 4 GB 2 vCPU, 4 GB | $39/month monthly, no long-term contract · VAT not stated checked 25 Oct 2026 | A 30-day trial of one cloud server in a configuration worth up to $100 checked 29 Oct 2026 | Visit |
| UpCloud Starter, 2 CPU cores, 4 GB 2 vCPU, 4 GB | €12/month billed by the starting hour, at most 28 days a month; no minimum term checked 25 Oct 2026 | No offer | Visit |
| Liquid Web Cloud VPS (General), 2 vCPUs, 4 GB 2 vCPU, 4 GB | $17/month monthly · VAT not stated checked 25 Oct 2026 | No offer | Visit |
Some of the lowest prices need a long term paid in advance; the term and any renewal price are under each price.
| Plan | Price | New-customer offer | Link |
|---|---|---|---|
| Hostinger KVM 2 2 vCPU, 8 GB | $8.79/month 24-month term, paid up front; Black Friday first-term price, then $14.99 checked 25 Oct 2026 | Black Friday prices on VPS plans and AI automation apps, for the first subscription term only checked 30 Oct 2026 | Visit |
| Contabo Cloud VPS 4 4 vCPU, 8 GB | €4.40/month 24-month term, paid up front; the price applies to the first 24 months and to the EU region in Germany; other regions cost more each month checked 25 Oct 2026 | No offer | Visit |
| DigitalOcean Basic Droplet, 4 vCPUs, 8 GiB 4 vCPU, 8 GB | $48/month billed per second (60-second minimum), capped at this monthly price · VAT not stated checked 25 Oct 2026 | New accounts get a $5 credit that lasts 90 days and covers every product except SaaS add-ons checked 29 Oct 2026 | Visit |
Hostinger's 8 GB plan has 2 cores; Contabo's and DigitalOcean's have 4.
If you would rather not size a server at all, n8n Cloud sets the memory by plan. Its Starter plan includes 2,500 executions a month.
| Plan | Price | New-customer offer | Link |
|---|---|---|---|
| n8n Cloud Starter | €20/month yearly plan, paid up front for 12 months and not refundable; works out at this price per month checked 25 Oct 2026 | Free trial of n8n Cloud with Pro-level features and no card needed checked 29 Oct 2026 | Visit |
| n8n Cloud Pro | €50/month yearly plan, paid up front for 12 months and not refundable; works out at this price per month (the 10,000-executions-a-month option) checked 26 Oct 2026 | Free trial of n8n Cloud with Pro-level features and no card needed checked 29 Oct 2026 | Visit |
Prices exclude taxes. A free trial without a card is capped at 1,000 executions.
Mistakes to avoid
Keeping files in memory
Instead: Set N8N_DEFAULT_BINARY_DATA_MODE to filesystem, or to database in queue mode.
Testing large data runs in the editor
Instead: Manual runs keep an extra copy for the screen; test on a small sample.
Saving every successful run
Instead: Once a workflow works, save only failed runs and keep pruning on.
Queue mode on SQLite
Instead: n8n does not support it; move to PostgreSQL first.
How we checked
We read n8n's pages on sizing, scaling and n8n Cloud's plan limits on 27 October 2026. The load test figures come from n8n's own blog post of 24 September 2025.
Our own measurements will be added here, with their dates, when our test bench runs them.
Other questions
What are the minimum requirements for n8n?
n8n gives no hard minimum. Its illustrative table lists 320 MB to 2 GB of memory, and its cloud runs the Starter plan on 640 MiB. On your own server, we would start at 2 GB.
How much RAM does n8n use at idle?
About 100 MB for an idle n8n Cloud instance, by n8n's sizing page. Its Cloud documentation puts the software alone at about 180 MiB on average.
Is 1 GB of RAM enough for n8n?
For a few light workflows it can be, since n8n Cloud's Starter plan has 640 MiB. With PostgreSQL, files or AI steps on the same server, 1 GB leaves little room.
Does n8n need a GPU?
No. n8n runs on the processor and uses little of it. A GPU helps only if you run an AI model on the same server, and small models can run without one.
When should I use queue mode?
When many runs arrive at the same time, or heavy workflows slow the others. In n8n's own test with 10 workflows, queue mode took a 4 GB server from 38% failed requests to none.
Sources and test notes
- n8n: measure performance, read
- n8n: fix memory issues, read
- n8n: prerequisites for an OEM deployment (sizing table), read
- n8n Cloud: manage your data (memory on each plan), read
- n8n: manage execution data, read
- n8n: handle binary data, read
- n8n: control concurrency, read
- n8n: enable queue mode, read
- n8n: Simple Memory node, read
- n8n Cloud: understand concurrency, read
- n8n blog: the n8n scalability benchmark (24 Sep 2025), read
- PostgreSQL: resource consumption (shared_buffers), read
- Redis: FAQ (memory footprint), read
Every page read on 27 October 2026.