Pi Cluster Boards & Cluster Setups

The "I want a real-looking Kubernetes cluster on my desk" category. Most cluster boards take Pi Compute Modules (CM4 / CM5) or sibling NVIDIA Jetson / RK1 modules and give you shared power, networking, and a single management plane. Honest take: a cluster of Pi CMs is a fantastic learning rig, an okay homelab, and a poor production target — a single mini-PC with VMs is cheaper and faster (see SBC vs. Mini-PC). For Pi-side OSes that suit clustering see Pi OS Distributions; for upstream Kubernetes ecosystem see Kubernetes Distros.

Cluster carrier boards

• ★ Turing Pi 2 (Turing Machines) — Mini-ITX cluster board with 4 SO-DIMM-style module slots. Accepts CM4, NVIDIA Jetson Nano / Orin Nano (via adapter), and Turing RK1 modules (RK3588, 32 GB option). Built-in 1G + 1G + management switch, BMC for IPMI-style remote control, dual SATA. ~$200 board + modules. The most polished cluster carrier in 2026.
• ★ DeskPi Super6c — single-board carrier for 6× CM4 with built-in 1 GbE switch, dual NVMe sockets shared across compute modules. ~$160. The "all CM4, no exotics" pick. CM5 support is rolling out via firmware updates in 2026.
• DeskPi RackMate T1 / T2 — 10" mini server rack (not a carrier itself); pair with a DeskPi 1U to mount Pis / mini-PCs in rack form. The "I want a tiny server rack on my desk" hardware. Pure mounting, no cluster management.
• Pine64 Clusterboard — 7× SOPINE A64 module slots; older, ARM A53; only buy if you have SOPINEs already.
• PicoCluster (PicoCluster.com) — pre-built 5-, 10-, 20-Pi cases with switch + power; more "appliance" feel than DIY.
• Pimoroni Cluster HAT (4× Pi Zero on a Pi 4) — a HAT that makes 4 Pi Zeros + 1 host into a "cluster." More toy than serious cluster. Educational only.
• Mythic Lab MX 1U Pi Rack — niche 1U enclosure for 4-8 Pis.

Module options

For most cluster boards the compute module is the variable.

• ★ Pi CM5 — 4-core A76 @ 2.4 GHz, up to 16 GB; PCIe 2.0 x1; the current Pi compute module. Drop-in for most CM4 carriers (firmware updates required on some). See Pi Models.
• Pi CM4 — 4-core A72; widely available, deeply tested with cluster boards.
• Turing RK1 — RK3588, 8-core (4× A76 + 4× A55) @ 2.4 GHz, up to 32 GB. ~2x faster than CM5 in throughput-heavy workloads. SO-DIMM module from Turing Machines. Popular for k8s clusters that want more RAM than CM5 offers.
• NVIDIA Jetson Orin Nano on Turing Pi 2 — for AI cluster experiments. See AI SBCs.
• Mixtile Core 3588E — RK3588 SO-DIMM compute module; pin-compatible with several carriers.

Software stacks for Pi clusters

The Kubernetes-on-Pi story split into "lightweight distros" by 2026.

• ★ k3s (Rancher / SUSE) — single-binary k8s, ~50 MB; the de-facto Pi/SBC k8s distro. Runs CM4 / CM5 / RK1 fine. Default for new Pi clusters.
• ★ k3sup — bash helper that bootstraps k3s clusters over SSH. The "I have 4 Pis, give me a cluster in 90 seconds" tool.
• MicroK8s (Canonical) — snap-based k8s; opinionated; fine on Pi but heavier than k3s.
• Talos Linux (Sidero) — immutable, API-only host OS; arm64 supported including Pi 4 / Pi 5 / RK1. The "production-quality cluster on tiny hardware" path. See Container Host OSes.
• k0s (Mirantis) — competitor to k3s; smaller community on Pi.
• kubeadm + Ubuntu Server — vanilla path; works but more setup than k3s.
• Docker Swarm — sometimes the right answer for "I have 3 Pis and want orchestration without k8s." Maintenance-mode upstream but still works.
• Nomad (HashiCorp) — single-binary scheduler; great for non-k8s clusters. See Nomad Orchestration.

Cluster GitOps / fleet patterns

• ★ FluxCD or ArgoCD for cluster-state-from-git — see GitOps CD.
• Ansible for non-k8s Pi fleets — see Pi Fleet Management.
• balenaCloud for "I just want OTA images on a fleet" — see Pi Fleet Management.

Storage in a Pi cluster

The hardest problem.

• ★ Longhorn — Rancher's k8s-native distributed block storage; works on arm64; the standard pick for k3s clusters.
• OpenEBS Mayastor / cStor — alternative; more complex.
• Rook + Ceph — possible on Pi but painful; Ceph wants more RAM and IOPS than CMs typically have. Skip unless you're learning Ceph specifically.
• NFS from a NAS — boring, works, fastest path to "PVCs that mount somewhere."
• GlusterFS — old but reliable for a few-node setup.

For clusters where each module has its own NVMe (Turing Pi 2 with NVMe-equipped modules, DeskPi Super6c with shared NVMe), Longhorn with replication count 2-3 across nodes is the clean pattern.

Networking

• MetalLB — bare-metal LoadBalancer for k3s; the "give my Service an IP" tool.
• Tailscale + Tailscale Operator — exposes cluster Services into your tailnet without LoadBalancer fuss. See VPN Mesh.
• Traefik (k3s default) or NGINX Ingress for HTTP routing.
• Cilium as CNI works on arm64 Pi clusters since v1.14, but eBPF features may need newer kernels than Pi OS ships.

Power, cooling, monitoring

• POE+ for cluster boards — Turing Pi 2 supports PoE+ uplink; DeskPi Super6c has DC barrel jack only.
• Active cooling on RK1 / Jetson modules — they thermal-throttle hard. Don't skip the heatsinks.
• node_exporter + Prometheus for cluster monitoring — see Prometheus Stack.
• DCGM for Jetson nodes if running NVIDIA modules.

Honest cluster math (May 2026)

A 4-node Turing Pi 2 with 4× RK1 8GB modules:

• $200 board + 4× $190 modules + $50 PSU + cables ≈ $1,000 for 32 cores / 32 GB total.

A used Beelink S12 Pro with N100 + 16 GB upgrade:

• ~$300 for 4 cores / 16 GB but 2-3× faster per-core, 1× the cables, 1× the power supply, runs the same k3s.

If you want to learn distributed systems, the Pi cluster is fantastic — the failure modes (one node going down, partial network partition) are realistic and you can yank a module to simulate them. If you want to run services, two Beelinks in a Tailscale tailnet is faster, cheaper, and quieter.

Pick this if…

• Most polished cluster carrier: Turing Pi 2 with RK1 modules.
• All-CM4/CM5, NVMe per node: DeskPi Super6c.
• You want a tiny visible rack on your desk: DeskPi RackMate + Turing Pi 2 or 1U Pi shelves.
• Default cluster k8s software: k3s, bootstrapped with k3sup.
• Production-quality cluster OS on Pi: Talos Linux.
• Cluster storage that actually works: Longhorn + per-node NVMe, or NFS from a NAS.
• You want to learn k8s the painful way: Pi cluster.
• You want to actually run workloads: mini-PCs in a Tailscale tailnet — see Homelab and SBC vs. Mini-PC.