Proxmox Homelab

The goal was to replace a standalone gaming PC and several would-be cloud instances with a single machine running Proxmox VE, a Type 1 hypervisor built on KVM/QEMU. Everything below runs concurrently on one Ryzen 9 5900X with 40GB of RAM and two NVMe drives (1TB dedicated, 2TB shared pool).

The hardest problem in this build was giving a VM true graphics performance. Consumer GPUs can't be shared between host and guest, so the RX 6700 XT is handed exclusively to the Windows VM using VFIO:

• IOMMU enabled at the firmware and kernel level, with the GPU and its audio function isolated into their own IOMMU group
• The vfio-pci driver claims the card at boot, preventing the host from ever initializing it
• The VM boots via OVMF (UEFI) with the GPU attached as a PCIe device, so Windows sees real hardware with no virtualized graphics layer
• Allocation: 4 cores, 10GB RAM, and a dedicated 1TB NVMe, keeping Windows I/O completely off the shared storage pool

The result is a "desktop" that performs like bare metal but lives alongside server workloads, and can be snapshotted or rebuilt like any other VM. One fun failure mode this surfaced: Windows Boot Manager once hijacked the UEFI boot order and knocked the whole hypervisor offline, a good lesson in why guest OSes should never share boot priority with the host.

A lightweight Debian VM (2 cores, 2GB RAM, 1TB storage) running the always-on services:

• Jellyfin for self-hosted media streaming to any device
• Automated download services whose traffic is forced through a WireGuard VPN tunnel
• iptables killswitch: rules only permit egress through the WireGuard interface, so if the tunnel drops, traffic stops instead of leaking out the WAN. VPN endpoint rotation is handled by a systemd service
• Tailscale for remote access, running here just like on every other VM and the Proxmox host itself. The media library and admin panels are reachable from my phone or laptop anywhere, with zero ports forwarded on the router

The 2GB allocation was deliberate: these are I/O-bound services, and keeping the footprint small leaves headroom for the other guests.

A 4-core / 4GB VM that mirrors the production environment of my SaaS (a DigitalOcean droplet running a three-container Docker stack: main API, PDF generation engine, and a usage-billing microservice). Having a production-like environment locally means I can test container builds, compose changes, and deployment scripts before they ever touch CI/CD, without paying for a staging droplet.

CPU is intentionally not fully committed. The 5900X's 24 threads leave room for host overhead and future guests, and none of the workloads peak simultaneously.