n8n on an ARM server: it runs fine, but is it the better buy?

n8n on an ARM server works out of the box: the official image is multi-arch, and our Hetzner ARM servers ran n8n 2.46.6 with the same compose file as x86. It is not faster, though. In our webhook load test the 2 vCPU ARM CAX11 peaked at 13.9 executions per second against 15.04 on the x86 CX23, and the 4 vCPU CAX21 at 25.77 against 31.93 on the x86 CX33.

Why this matters now: ARM VPS plans are often sold as the efficient, cheap choice, and the ARM guides we found online quote no measured n8n numbers. We ran the identical test on two ARM and three x86 servers. This page is part of our n8n hosting section.

Tested in the Lab Hetzner CX23, CX33, CX43, CAX11, CAX21 · n8n 2.46.6 · Nov 2026

Does the official n8n image support ARM?

Yes, 64-bit ARM. The tag list of n8nio/n8n on Docker Hub shows builds for linux/amd64 and linux/arm64, and n8n's Dockerfile on GitHub notes that its Node.js base image is pinned "for amd64 and arm64". Docker picks the right build for the server, so the image line is the same on both:

image: docker.n8n.io/n8nio/n8n:2.46.6

On our CAX11 and CAX21 that pulled and started with no changes, and so did postgres:16. We saw no 32-bit armv7 build in the official tag list; the armv7 images on Docker Hub come from third parties. We only tested n8n's built-in nodes. If you use community nodes, test them on ARM before you move.

ARM vs x86, same size, same test

All servers: Hetzner Cloud, Helsinki, Ubuntu 24.04, 5 November 2026. The ARM CAX line uses Ampere cores that report as Neoverse-N1. The workflow is webhook → Code node → response with n8n 2.46.6 and PostgreSQL 16, loaded with ApacheBench from the same server. Executions/s is completed workflow runs per second; p95 is the latency 95% of requests beat.

Bold marks the better result in each pair. ARM won only one row: the CAX11 got n8n up faster than the CX23. Idle RAM is the same on both architectures within a few MiB.

Measured by AI Hosting Lab on 2026-11-05 · 5 servers · method · raw data

At 2 vCPU: close, x86 slightly ahead

The CX23 and CAX11 were within 8% of each other at their best (15.04 vs 13.9). The CX23 has an old Intel Skylake CPU, so this is the pairing where ARM should have won.

At 4 vCPU: x86 clearly ahead

The AMD-based CX33 did 24% more executions per second than the CAX21 (31.93 vs 25.77) and answered a single request in 72 ms against 107 ms. It is also the cheaper server.

Why a fast ARM CPU score did not help

In sysbench, the CAX11 scored 3248.26 events/s on one thread, 3.55 times the CX23's 913.87. On all threads it scored 6549.84, about the same as the 4 vCPU CX33 (6453.15). If CPU benchmarks predicted n8n, the CAX11 would match the CX33. It ran 13.9 executions/s; the CX33 ran 31.93.

What tracked n8n throughput in our data was the number of vCPUs, up to about four. In regular mode one n8n main process takes the webhooks and runs the executions (n8n queue mode docs), and Node.js runs JavaScript on one event loop per process (Node.js docs). Each execution also goes through a task runner process and writes to PostgreSQL. On 2 vCPU those processes compete for two cores however fast each one is. That is our reading; we did not measure CPU per process. Sysbench's prime-number loop also measures something narrower than a Node.js web server, so a high score does not carry over. The same pattern in Ollama, where ARM did win at 2 vCPU, is in Ollama on an ARM server.

When ARM is the better buy

• Your stack is already ARM, or you build and test arm64 images anyway. n8n will not be the problem.
• You mix n8n with a small local model on 2 vCPU. Our ARM CAX11 beat the CX23 on Llama 3.2 1B and 3B generation, so a small n8n + Ollama box is a fair case for ARM.
• Prices change. On a provider where ARM costs clearly less than x86 of the same size, the small throughput gap at 2 vCPU could be worth it. Other ARM clouds are not yet tested in our Lab; testing is scheduled.

When to stay on x86

• n8n is the main job of the server. At Hetzner the x86 plan was faster at both sizes and cheaper: 0.61 € a month less at 2 vCPU, 2.42 € less at 4 vCPU.
• You use community nodes or tools that you have not checked for arm64 builds.
• You want the most webhook headroom per euro. Our pick for that is in best VPS for n8n.

The wider comparison of both Hetzner lines, beyond n8n, is in Hetzner CX vs CAX.

Limits of this test

• One run per server, one light workflow, regular mode only. We have not tested queue mode on ARM or x86 yet.
• We have not tested the larger CAX31 and CAX41 for n8n.
• Load came from the same server, so network and TLS time are not included.

Next step: see the full method and every concurrency level in our n8n webhook performance report.