# AMD EPYC 7402P Proxmox Host

> Originally built 2022-07-14 — actively evolved

Status: active · Kind: snapshot · Last verified: 2026-07-20 · Date: 2022-07-14

> Status: active
> Last verified: 2026-07-20 (against the current physical-host inventory)
> Scope: public hardware configuration and design intent

*Originally built 2022-07-14 — actively evolved*

Say hello to my little friend.

---

## System Overview

This system is a **single-socket AMD EPYC virtualization host** running Proxmox VE.
It serves as a primary compute node for infrastructure services, GPU-accelerated
workloads, and general-purpose virtualization.

The design emphasizes predictable performance, clean I/O topology, and long-term
serviceability over maximum density or theoretical peak benchmarks.

> This system is intentionally **single-socket**. No dual-socket configurations are
> used anywhere in the lab to avoid NUMA complexity and cross-socket latency.
> Design principles: see **[Lab Philosophy](/homelab/philosophy)**.

---

## Hardware Configuration

| Component        | Manufacturer | Model / Details |
|------------------|--------------|-----------------|
| Motherboard      | Supermicro   | H12SSL-C (SP3, PCIe 4.0) |
| CPU              | AMD          | EPYC 7402P (24C / 48T, Rome, P-series) |
| Heatsink         | Supermicro   | SNK-P0064AP4 |
| Memory           | Micron       | 256 GB DDR4-3200 RDIMM (8 × 32 GB) |
| Chassis          | Rosewill     | RSV-L4000U (4U rackmount) |
| Drive Cage       | Icy Dock     | ToughArmor MB516SP-B |
| Boot Pool        | Mixed        | Mirrored SATA SSDs (Samsung 850 PRO 256 GB + Crucial MX200 250 GB) |
| Data Pool        | Seagate      | Nytro 1551 SATA SSDs (ZFS mirrors) |
| GPUs             | NVIDIA       | 2 × GeForce RTX 3090 (passed through to VM) |
| NIC              | Intel        | X520 (82599ES) Dual-port 10 Gb SFP+ |
| Fans             | Noctua       | NF-A12x25 PWM |
| Power Supply     | Corsair      | HX1200 |
| OS               | Proxmox      | Proxmox VE |

---

## CPU & Platform Notes

The **EPYC 7402P** is a P-series processor intended exclusively for **single-socket**
deployments. This provides:

- 24 cores / 48 threads
- Full 8-channel memory bandwidth
- Full PCIe lane availability
- No inter-socket NUMA overhead

This makes it particularly well-suited for virtualization and mixed workloads where
predictable latency and clean topology matter more than raw socket count.

---

## Memory Configuration

- Total installed: **256 GB DDR4-3200**
- Configuration: **8 × 32 GB RDIMMs**
- Topology: One DIMM per channel (1DPC)

This layout maximizes memory bandwidth while maintaining stability under sustained load.

---

## Storage Layout

### Boot Pool (`rpool`)
- ZFS mirror
- Mixed SATA SSDs:
  - Samsung 850 PRO 256 GB
  - Crucial MX200 250 GB

This pool hosts the Proxmox OS and system datasets.

### Fast Storage Pool (`fast`)
- ZFS mirrored vdevs
- Enterprise SATA SSDs (Seagate Nytro)

Used for VM disks where consistent I/O latency matters.

---

## GPU Configuration

The system contains **two NVIDIA GeForce RTX 3090 GPUs**, both of which are:

- Isolated via IOMMU
- Passed through directly to a dedicated VM
- Used for AI / ML workloads

No GPU sharing or mediated devices are used — each GPU has exclusive access within
the guest for maximum performance and driver stability.

---

## Networking

- **Intel X520 dual-port 10 Gb SFP+ NIC**
  - Based on the Intel 82599ES controller
  - Used for high-speed storage and backbone connectivity
- Onboard 1 Gb interfaces are retained for management and fallback access

---

## Cooling & Power

Cooling and power delivery were selected for **sustained load and reliability**, not
minimum acoustics.

- Noctua **NF-A12x25 PWM** fans provide front-to-back airflow across drive cages
- Corsair **HX1200** PSU supplies ample headroom for dual GPUs under load

Thermal behavior remains stable under continuous CPU and GPU workloads.

---

## Design Notes

Key design goals for this system:

- Single-socket architecture to avoid NUMA complexity
- High memory bandwidth with simple DIMM topology
- Clean PCIe layout for GPUs and high-speed NICs
- Hardware that can remain in service for many years with incremental upgrades

This machine reflects a preference for **operational clarity** and **predictable behavior**
over chasing maximum density or short-term benchmarks.

---

## Related Documentation

- [AI Inference VM](/homelab/hardware/ai-inference-vm) — GPU-accelerated inference platform running on this host

---

## Photos

*Updated photos showing the dual-GPU configuration and current layout will be added
in a future revision.*