Hardware where it matters
The router owns its physical network card. The NAS owns the SAS controller and its disks so ZFS sees the storage directly. The AI VM receives a V100, P100, and M40; the media VM receives the second P100.
Hardware & homelab
My setup.
The desktop I build on, and the homelab I learn in.
01 / My PC
The machine on my desk.
A look at my personal setup.
Personal workstation
My PC
Windows 11
64-bit
- Case
- Corsair 7000D AIRFLOWFull-tower ATX
- Processor
- Intel Core i7-13700K13th generation · 16 cores / 24 threads
- Graphics
- Gigabyte GeForce RTX 4090 OC24 GB GDDR6X · PCIe 4.0 ×16
- Memory
- 64 GB G.Skill DDR52 × 32 GB · Configured at 6,400 MT/sModule model: F5-6400J3239G32G
- Motherboard
- ASUS PRIME Z790-P WIFIIntel Z790 · LGA1700 · DDR5PCIe 5.0 ×16 expansion slot
- Primary · C:
- 1 TB WD_BLACK SN850M.2 NVMe · PCIe 4.0 ×4Model: WDS100T1X0E-00AFY0
- SSD · L:
- 500 GB Samsung 970 EVOM.2 NVMe · PCIe 3.0 ×4Model: MZ-V7E500
- HDD · D:
- 4 TB Seagate IronWolf ProSATA 6 Gb/s · 7,200 RPMModel: ST4000NT001-3M2101
02 / Homelab
The homelab.
One rack server. A lot to learn.
My lab brings together the things I work on: virtual machines, storage, networking, self-hosted applications, and local AI. Some services are part of daily life; others are experiments I’m still building.
Documented October 7, 2026 · A snapshot of the setup01 / Hardware
Inside the rack.
Silicon Mechanics rack server, built around a Supermicro X10DRG-HT GPU motherboard.
- Host
- Proxmox VE 9.2IPMI with onboard ASPEED graphics
- Processors
- 2 × Intel Xeon E5-2690 v324 cores / 48 threads
- Memory
- 128 GB DDR4 ECC registered · 2,133 MT/s8 × 16 GB · 8 of 16 slots populated
- Power
- Dual 2,000 W redundant power supplies
- Boot & VM storage
- 4 × 500 GB WD VelociRaptor · 10,000 RPM SATATwo ZFS mirrors · approximately 862 GB usable
- NAS storage
- 21 drives · 9 mirrored pairs and 3 hot sparesApproximately 20.7 TB usable · Broadcom / LSI SAS2308 HBA
- Networking
- Intel X540 dual-port 10GBASE-T + Intel 82599 dual-port 10 Gb SFP+
- GPUs
- NVIDIA Tesla V100 · 16 GB2 × Tesla P100 · 16 GB each
Tesla M40 · 24 GB
02 / Virtual machines
What runs here.
Eleven defined VMs, ten running at the time of this snapshot. Expand a machine for its role and dedicated hardware.
aiLocal AI and developer tools8 vCPU · 16 GiB
Runs local language models through llama.cpp and llama-swap, OpenCode, ComfyUI, SearXNG, Tinyauth, and PostgreSQL. Services use systemd user units and rootless Podman Quadlets.
Tesla V100 16 GB, Tesla P100 16 GB, and Tesla M40 24 GB. The M40 is installed but idle.
k3s-1Kubernetes learning node2 vCPU · 4 GiB
Provisioned · k3s planned
Ubuntu Server VM provisioned for a future three-node k3s cluster. The operating system is installed; k3s is not installed yet.
Virtual disk and network interface
k3s-2Kubernetes learning node2 vCPU · 4 GiB
Provisioned · k3s planned
Second provisioned Ubuntu Server VM for the planned k3s cluster with embedded etcd. A Talos Linux cluster is a later learning project.
Virtual disk and network interface
k3s-3Kubernetes learning node2 vCPU · 4 GiB
Provisioned · k3s planned
Third provisioned Ubuntu Server VM for the planned k3s cluster. No Kubernetes workloads are running yet.
Virtual disk and network interface
mediaMedia, files, and personal cloud8 vCPU · 16 GiB
Hosts the Podman media stack, including Jellyfin, Seerr, Navidrome, Nextcloud, FileBrowser Quantum, and Tdarr. A shared NAS path supports hardlinks; container state stays on the VM disk. An HLS proxy supports shared playback in VRChat.
Dedicated Tesla P100 16 GB for media processing
omada-controllerWi-Fi management4 vCPU · 4 GiB
Manages three TP-Link Omada access points: two EAP720 units and one EAP725-Outdoor. Wireless clients use a separate network from server workloads.
Virtual disk and network interface
opnsenseRouter and firewall4 vCPU · 8 GiB
Routes IPv4 and IPv6 traffic between home devices, VLAN 10 (VMLAN) for virtual machines, and VLAN 20 (WLAN) for wireless clients. Provides firewall rules, DHCP with network-boot options, and IPv6 prefix delegation.
Intel 82599 dual-port 10 Gb SFP+ card passed through directly
pxeNetwork boot and deployment4 vCPU · 8 GiB
Serves an iPXE installation menu for Windows 11, Arch, Debian, Ubuntu, and Rocky Linux. Boot entries are generated from JSON definitions, and install media is served from a read-only NAS folder.
Virtual disk and network interface
tankZFS storage server8 vCPU · 32 GiB
Manages 18 data drives in nine mirrored pairs, with three hot spares and approximately 20.7 TB usable capacity. Shares storage over NFS and SMB, with monthly ZFS scrubs.
Broadcom / LSI SAS2308 HBA and all 21 attached drives passed through directly
testScratch environment4 vCPU · 8 GiB
Stopped
A stopped VM reserved for experiments. It does not start automatically with the host.
Virtual hardware
webWeb hosting and reverse proxy2 vCPU · 2 GiB
Runs nginx with HTTP/2 and HTTP/3, TLS termination, and reverse proxying. Configuration comes from Git; deployment validates it before reloading. Tinyauth provides sign-in gating for administrative applications.
Virtual disk and network interface
03 / Architecture
How it fits together.
Separate networks, a shared host
OPNsense separates home devices, server VMs, and wireless clients. VLAN 10 (VMLAN) carries the virtual-machine network, while VLAN 20 (WLAN) carries wireless clients. A VLAN-aware Proxmox bridge connects the virtual workloads. The network supports both IPv4 and IPv6, with OPNsense handling routing, firewall rules, and IPv6 prefix delegation. nginx on the web VM provides the entry point for web applications.
Storage shared across services
The NAS exposes its dataset through NFS and SMB. Media applications share a consistent folder layout so imports can use hardlinks. The deployment VM reads only its own installation-media folder.
Local AI, with mixed hardware
llama.cpp and llama-swap serve local models, with the V100 as the everyday card and the P100 available for split models. ComfyUI handles image generation. The driver and CUDA versions are deliberately pinned to support the older GPU generations.
04 / Operations
Keeping it running.
- Configuration in Git. nginx configuration, container definitions, deployment scripts, and runbooks live together.
- Updates with boundaries. Containers update weekly; GPU drivers, model tooling, and selected major application versions are maintained deliberately.
- Database backups first. PostgreSQL is dumped daily and before updates. An unsuccessful pre-update backup stops the update.
- Recovery is documented. Scripts restore containers after interrupted processes or power loss, and bulk file renames have reversal logs.
- Still learning. The three Ubuntu Kubernetes nodes are provisioned. Installing k3s, then exploring Talos, is planned work.
05 / Around the rack
The rest of the setup.
TP-Link Omada access points
Two EAP720 units and one EAP725-Outdoor, managed by the Omada controller VM.
Unmanaged PoE switch
Connects the access points.
Unmanaged SFP+ switch
Provides SFP+ connectivity for the lab. I'm looking to upgrade to a managed Layer 3 switch with more ports.
Intel Arc B580 streaming PC
A separate Linux computer running Sunshine for game streaming, connected privately to the desktop.