Hardware Inventory

The cluster is built from a mix of x86 and ARM hardware, mounted in an open-frame 12U rack together with a patch panel, a PDU, and a UPS. AMD mini PCs run the control plane and Ceph, a Dell R630 provides bulk compute and ZFS storage, a GPU workstation runs AI workloads, and Intel nodes and Raspberry Pis fill out the worker pool.

Server rack

All nodes are on VLAN 20 with static DHCP reservations (terraform/unifi/reservations.tf). CPU and memory figures below are what Kubernetes reports.

Compute Nodes

Control Plane + Workers

These nodes run the control plane and schedule workloads: the control-plane taint is removed in talos/pitower/control-plane/01-cluster.yaml, with systemReserved/kubeReserved set so app load cannot starve the API server. The VIP 10.20.10.0 floats across all three (https://10.20.10.0:6443).

Hostname	Hardware	IP	CPU	RAM	Storage	Schematic
worker-01	AMD Ryzen mini PC	10.20.10.1	16 threads	16 GB	NVMe boot + SATA SSD (Ceph OSD)	`amd`
worker-02	AMD Ryzen mini PC	10.20.10.2	16 threads	32 GB	NVMe boot + SATA SSD (Ceph OSD)	`amd`
worker-03	AMD Ryzen mini PC	10.20.10.3	16 threads	32 GB	NVMe boot + SATA SSD (Ceph OSD)	`amd`

Intel Workers

Hostname	IP	CPU	RAM	Boot disk	Notes	Schematic
worker-04	10.20.10.4	4 cores	8 GB	eMMC	`dedicated=media-home` taint	`intel`
worker-05	10.20.10.5	4 cores	8 GB	Samsung SSD	Also on the untagged LAN (`networking/lan`); no VT-x	`intel`
worker-06	10.20.10.6	4 cores	8 GB	Samsung SSD	Also on the untagged LAN (`networking/lan`); no VT-x	`intel`

worker-07 (Dell R630)

A bare-metal Dell R630 (ex proxmox-01, see the migration record) at 10.20.10.7, on a 10G Intel X710 port.

Property	Value
CPU	48 threads
RAM	~755 GiB usable (ZFS ARC capped at 64 GiB)
Disks	Six 745 GiB SAS SSDs and four ~600 GB 10K SAS HDDs behind the PERC in HBA mode
Talos system disk	md RAID1 across two SSDs (`RAIDArrayConfig`, selected by WWID)
ZFS `fast`	Four SSDs as two mirrors: `/var/mnt/extra`, `/var/mnt/runners`, Garage metadata
ZFS `hdd`	Four HDDs as raidz1: `/var/mnt/media`, Garage blocks
Schematic	`r630` (adds the `zfs` extension)

The OpenEBS classes openebs-hostpath-fast, -runners, and -media only provision here, and the self-hosted home-ops CI runners are pinned to it. It also runs Garage S3. ZFS is administered from a privileged hostPID pod with nsenter, since Talos has no shell; system/zfs-scrub scrubs both pools monthly.

worker-ai-01 (GPU)

A bare-metal ASUS ProArt B850 board with an NVIDIA RTX 3090 Ti at 10.20.10.11, dual-booted with Bazzite.

Property	Value
CPU	AMD, 16 threads
RAM	64 GB
System disk	1 TB Kingston NV3 (shared with Bazzite; Talos EPHEMERAL capped at 200 GiB)
Models disk	2 TB Samsung 9100 Pro, Talos user volume `models` at `/var/mnt/models` (`openebs-hostpath-models`)
Schematic	`nvidia` (open GPU kernel modules + container toolkit)
Taint	`dedicated=gpu:NoSchedule`

The GPU is shared through the NVIDIA DRA driver (system/dra-driver-nvidia-gpu) and can be passed through to a KubeVirt VM with vfio-pci. system/nvidia-power-limit caps it at 300 W. Because it is often booted into Bazzite, it is excluded from Cilium's L2 and BGP announcements.

Raspberry Pi Workers

Hostname	IP	Model	RAM	Boot disk	Schematic
worker-08	10.20.10.8	Raspberry Pi 4 (PoE HAT)	4 GB	128 GB USB SSD	`rpi-poe`
worker-09	10.20.10.9	Raspberry Pi 4 (PoE HAT)	4 GB	128 GB USB SSD	`rpi-poe`
worker-10	10.20.10.10	Raspberry Pi 4 (PoE HAT)	4 GB	128 GB USB SSD	`rpi-poe`

The rpi-poe schematic uses the sbc-raspberrypi overlay and sets PoE HAT fan thresholds. KubeVirt workloads are kept off these arm64 nodes.

Network Equipment

Device	Purpose
UniFi Cloud Gateway Fiber	Router, VLANs, DHCP, BGP peer (ASN 64512) for LoadBalancer and pod routes
TP-Link PoE switch	Powers the Raspberry Pi nodes
UniFi access points	Wi-Fi
Patch panel	Front of the rack

Networks (terraform/unifi/networks.tf):

Network	VLAN	Subnet
Default	untagged	`192.168.0.0/24`
home	10	`10.10.0.0/16`
servers	20	`10.20.0.0/16`
management	50	`10.50.0.0/24`
iot	101	`10.101.0.0/16`

Storage

Device	Purpose
Ceph OSDs	One SATA SSD (~500 GB) on each of worker-01..03, `ceph-block` default StorageClass
worker-07 ZFS	`fast` and `hdd` pools, OpenEBS hostpath classes and Garage S3
worker-ai-01 NVMe	2 TB `models` volume
Synology NAS (`data`, `10.20.10.100`)	NFS for media and bulk data

Power

Device	Purpose
PowerWalker UPS	In the rack
PDU	Power distribution in the rack
Eaton 5S, CyberPower PR1500LCDRT2U	UPSes served by the NUT host (`nut`, Ansible role `nut`) and scraped by `monitoring/nut-exporter`

Network Topology

graph LR
    Internet -->|WAN| GW[UniFi Cloud Gateway Fiber]
    GW -->|VLAN 20 + LAN| SW[Switching]

    SW --> CP[worker-01..03<br/>AMD]
    SW --> IN[worker-04..06<br/>Intel]
    SW --> R630[worker-07<br/>R630, 10G]
    SW -->|PoE| PI[worker-08..10<br/>Pi 4]
    SW --> AI[worker-ai-01<br/>GPU]
    SW --> NAS[Synology NAS]

    CP & IN & R630 & PI -.->|BGP| GW