Michiel Piscaer

machinelearning

Summary: I’m building a low-power ARM homelab running Kubernetes on an IPv6-only routed network. I’ll experiment with Open-Xchange, OpenStack, VXLAN, and AI frameworks like PyTorch and TensorFlow.

#homelab #arm #kubernetes #openstack #openxchange #ipv6 #bgp #ospf #vxlan #mikrotik

I’m planning to build a low-power ARM-based homelab that can optionally run Android. The main environment will be a Kubernetes (k8s) cluster operating on an IPv6-only backend network. The underlay network will be fully routed using OSPF or BGP, providing redundancy without relying on LACP.

Later, I plan to deploy Open-Xchange on the Kubernetes cluster, followed by OpenStack-Helm. The provider network will also be routed through the underlay, likely using VXLAN as the overlay protocol. I also want to experiment with AI frameworks such as PyTorch and TensorFlow.

Hardware Overview

The switching layer will be based on two MikroTik desktop switches, forming the core of the redundant network.

Candidate Switch

A switch I’m considering is the MikroTik CRS310-8G+2S+IN, offering:

  • 8 × 2.5GbE ports
  • 2 × SFP+ ports
  • Routing capabilities
  • Support for OSPF and BGP

The primary requirements are having enough ports for all nodes and ensuring sufficient routing capacity for the underlay network.

Compute Nodes

I’ve evaluated two potential ARM-based node options:

Below is a summary of both.

Armsom Sige7-Pro Max

  • CPU: 4 × Cortex-A76 @ 2.4GHz + 4 × Cortex-A55 @ 1.8GHz (8nm)
  • RAM: 32GB
  • GPU: Mali-G610
  • NPU: 6 TOPS
  • AI Support: PyTorch and TensorFlow (limited SBC-class acceleration)

Cost: – $252 (~€210) for the SBC – Additional cost for cooling, power supply, and storage – Approx. €900 total for a three-node setup

MiniS Forum MS-R1

  • CPU: CIX CP8180 (12 cores / 12 threads @ 2.6GHz)
  • TDP: 28W
  • NPU: 28.8 TOPS (45 TOPS total AI capacity)
  • Memory & Storage: 32GB RAM, 1TB SSD included

Cost: – €659 (or €599 during the Black Friday deal) – Approx. €1.977,- for three nodes

Conclusion

The Armsom Sige7-Pro Max is significantly more affordable and offers decent performance for its size, but it is still a single-board computer and noticeably slower overall.

The MiniS Forum MS-R1, on the other hand, provides much stronger CPU performance, better storage capability, and more powerful AI acceleration—but at a much higher price.

In summary:

  • Armsom Sige7-Pro Max:

    • ✔ Budget-friendly
    • ✔ Low power
    • ✘ Limited performance

  • MiniS Forum MS-R1:

    • ✔ Strong performance
    • ✔ Ready-to-use hardware
    • ✘ Considerably more expensive

#homelab #arm #kubernetes #openstack #openxchange #ipv6 #bgp #machinelearning #tensorflow #pytorch