EPYC 7542 Proxmox Node
This is the current main compute node of my homelab. It replaced the HP Z640 as the serious Proxmox box, mostly because I wanted more PCIe lanes, more memory headroom, IPMI, more local NVMe, and enough GPU space to separate media transcoding from AI experiments.
The short version: Supermicro H12SSL-I, AMD EPYC 7542, ECC LRDIMM, Proxmox, 10G networking, RTX 3090, Quadro P2200, and a bunch of dead RDIMMs.
Hardware
- Motherboard: Supermicro H12SSL-I
- CPU: AMD EPYC 7542
- Cooler: Noctua NH-U14S TR4-SP3
- RAM: 3*64GB DDR4 LRDIMM sticks, currently running the stable set while I sort out the full channel population
- Case: SilverStone RM51 5U rackmount chassis
- PSU: Corsair RM1000x
- Networking: onboard 1G + IPMI + 10G SFP+ uplink to the CRS310
- GPU: RTX 3090 24GB for local LLM / AI experiments
- GPU: Quadro P2200 for Jellyfin transcoding
- Storage: 990 Evo Plus boot drive, ASUS Hyper M.2 card with 2x Samsung 990 PRO / 2x 990 EVO Plus SSDs
Role in the Lab
This machine is meant to be the main VM and GPU host. The old Debian server used to carry too many services in one Docker Compose file, which worked until the service list became too long and every change felt like defusing a bomb. The EPYC node lets me split services into clearer VMs, assign VLANs properly, and pass through GPUs to the workloads that actually need them.
- Infrastructure VMs: DNS, Caddy, Gitea, Jenkins, Terraform controller, monitoring.
- Media VMs: Jellyfin, Shoko, Komga, and related services.
- AI VMs: Ollama and Open WebUI, with the RTX 3090 reserved for local model experiments.
- Kubernetes lab: K3s control/worker nodes for Rancher, AWX, and cluster experiments.
- Sandbox VMs: disposable systems for testing before touching the actual network.
Proxmox Layout
Proxmox is the base hypervisor. The current VM flow is to use a Debian cloud-init template with
qemu-guest-agent, then let Terraform clone and size VMs, and let Ansible take over the actual
configuration. This took a while to make sane because guest-agent state, cloud-init timing, and Proxmox clone
behaviour all have opinions.
Networking is done through a VLAN-aware bridge connected to the 10G path. VMs get placed into the relevant VLAN instead of everything living in one flat network like an accidental LAN party.
GPU Split
The RTX 3090 is better spent on local LLM experiments, Open WebUI/Ollama, and anything that needs 24GB of VRAM. The P2200 is a better fit for Jellyfin transcoding, because media serving should not have to fight an LLM for GPU time just because I asked a model to explain Raft while someone is watching anime.
Status
- Working: Proxmox install, IPMI access, VM templates, Terraform-driven VM creation, 10G uplink.
- Working: Core infra VMs and media/AI VM migration path.
- In progress: final RAM population, GPU passthrough polish, monitoring, and documentation.
- Lesson learned: memory errors on a server board suck. Test sticks properly.