Homelab Server Build: Complete Component Selection Guide

# Homelab Server Build: Complete Component Selection Guide

## What is a Homelab?

A homelab is a personal server setup at home used for learning, self-hosting services, experimentation, and development. Unlike cloud services, you own and control the hardware. The barrier to entry has dropped dramatically — affordable mini-PCs, used enterprise gear, and ARM-based SBCs make entry-level homelabs accessible for under $100.

## Step 1: Define Your Use Case

Before buying anything, answer these questions:

1. **What services?** (media server, VPN, NAS, dev environment, Kubernetes cluster, game server, local LLM)
2. **How many containers/VMs?** (1-3 vs 10+ changes requirements dramatically)
3. **Budget range?** ($50 Pi vs $500 mini-PC vs $2000 enterprise tower)
4. **Power tolerance?** A server drawing 200W 24/7 costs ~$175/year at $0.12/kWh

---

## CPU Selection

### Low-Power (ARM / Atom-class)
Best for: NAS, Pi-hole, VPN, light containers, always-on services.

| Option | TDP | Cores | Notes |
|--------|-----|-------|-------|
| Raspberry Pi 5 | 5W | 4 (ARM Cortex-A76) | Best ecosystem, active community |
| Orange Pi 5 | 5-8W | 8 (RK3588) | NPU included, excellent perf/watt |
| Beelink EQ12 (Intel N100) | 6-10W | 4 | x86, runs Proxmox natively, ~$150 |
| Beelink SER6 (Ryzen 7840U) | 28-54W | 8 | iGPU good for local inference |

### Mid-Range (Workstation / Small Server)
Best for: Home lab with 5-10 VMs, ML experiments, Kubernetes learning cluster.

| Option | TDP | Cores | Notes |
|--------|-----|-------|-------|
| Intel Core i5-12600 | 65W | 10 | Strong single-thread, good value |
| AMD Ryzen 9 7900 | 65W | 12 | ECC support, low TDP for power |
| Used Intel Xeon E5-2690 v4 | 105W | 14 | Cheap used (~$30), but power-hungry |

### Enterprise Used Gear
Used Dell PowerEdge R730, HP ProLiant DL380 G9 — loud, power-hungry (300-500W idle), but IPMI remote management and ECC RAM are invaluable for serious storage setups.

---

## RAM

- **Minimum 16GB** for basic NAS / Proxmox host running 2-3 VMs
- **32GB** comfortable for 5-8 VMs with moderate workloads
- **64GB+** for Kubernetes cluster, multiple databases, or ML workloads

**ECC RAM:** Strongly recommended if running NAS or ZFS filesystem. Bit flips on non-ECC RAM can silently corrupt the entire ZFS pool. Most consumer platforms don't support ECC — use Xeon or EPYC platforms if ECC is required.

---

## Storage Tiers

### Tier 1: Boot Drive
NVMe SSD for the OS. 128GB is more than enough. Recommended: Samsung 970 Evo Plus, WD SN770, Kingston SNV2S. Budget option: any DRAM-less NVMe is fine for OS.

### Tier 2: Fast Storage (VMs, databases)
SATA SSDs in RAID-1 or ZFS mirror. 2× 1TB gives ~$80 at current prices and provides redundancy.

### Tier 3: Bulk Storage (media, backups)
3.5" NAS-rated HDDs. **WD Red Plus** or **Seagate IronWolf** (both CMR). Never use desktop drives 24/7 — they aren't rated for it.

> ⚠️ **SMR Warning:** WD Red (without "Plus") uses SMR technology. SMR drives perform catastrophically under ZFS resilver/rebuild loads. Always verify CMR before buying.

### The 3-2-1 Backup Rule
3 copies of data, 2 different media types, 1 offsite. Even in a homelab, losing irreplaceable data hurts.

---

## Networking

| Speed | Use Case | Cost |
|-------|----------|------|
| 1 GbE | Standard home use, ~125 MB/s max | Built-in on most boards |
| 2.5 GbE | NAS serving 4K media to 2-3 clients | Intel I226-V NIC ~$30 |
| 10 GbE | High-speed inter-VM traffic, Ceph clusters | SFP+ ~$50, RJ45 10G ~$80 |

**Managed switch** recommended even for small labs: VLANs separate IoT / trusted / management / storage traffic, reducing attack surface and broadcast storms.

---

## Power Supply & UPS

Match PSU wattage to actual load + 20% headroom. Oversized PSUs run inefficiently at low load (80+ Platinum/Gold efficiency curves matter).

**UPS (Uninterruptible Power Supply) is non-negotiable for any NAS.** A sudden power cut mid-write can corrupt ZFS pools. APC Back-UPS 900VA (~$100) covers a typical homelab NAS for 15-20 minutes — enough for graceful shutdown.

---

## Recommended Software Stack

| Role | Software | Notes |
|------|---------|-------|
| Hypervisor | Proxmox VE (free) | KVM + LXC, excellent web UI |
| Container orchestration | Docker + Portainer | Easy management, large ecosystem |
| NAS OS | TrueNAS Scale or OpenMediaVault | ZFS, SMB/NFS, plugin support |
| Firewall/Router | OPNsense or pfSense | Full network control |
| Monitoring | Grafana + Prometheus + node_exporter | Full observability stack |
| DNS / Ad-block | Pi-hole or AdGuard Home | Network-level ad blocking |
| Reverse proxy | Nginx Proxy Manager or Caddy | HTTPS for internal services |

---

## Power Efficiency & Compute Optimization (by @nikolatesla)

Power draw directly impacts your homelab's total cost of ownership. Here's how to measure and optimize.

### Idle Power Profiling

```bash
# Check CPU frequency scaling governor
cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor

# Set to powersave for idle workloads
echo powersave | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor

# Check current draw via IPMI (enterprise servers)
ipmitool dcmi power reading
```

Use a Kill-A-Watt meter to measure actual wall power draw before and after optimizations.

### Wake-on-LAN for Non-24/7 Hosts

If a machine doesn't need to run constantly, sleep it and wake remotely:

```bash
# Enable WoL on network interface
ethtool -s eth0 wol g

# Wake remotely (from another machine on the LAN)
wakeonlan AA:BB:CC:DD:EE:FF
```

Saving 60W × 8,760 hours/year = ~526 kWh = ~$63/year at $0.12/kWh.

### ARM vs x86 for Local AI Inference

For LLM inference workloads (llama.cpp, Ollama), ARM SoCs with NPU now compete seriously with x86:

| Platform | NPU Capability | Llama-3 8B (Q4, tok/s) |
|----------|---------------|------------------------|
| Raspberry Pi 5 (no NPU) | CPU only | ~2 |
| Orange Pi 5 (RK3588) | 6 TOPS NPU | ~8 |
| Beelink SER6 (Ryzen 7840U) | 780M iGPU | ~20-25 |
| NVIDIA Jetson Orin Nano 8GB | 40 TOPS | ~30-40 |
| Mac Mini M4 Pro (36GB) | 38 TOPS Neural Engine | ~80+ |

For dedicated local LLM serving, a Beelink SER6 or Mac Mini M4 delivers the best performance-per-watt. Jetson Orin is ideal for embedded/edge deployment.

### Cluster vs Monolith Trade-off

A single beefy server (e.g., Ryzen 9 7900 + 128GB RAM) often outperforms a cluster of cheaper nodes for typical homelab use cases:

- No network overhead between services
- Simpler maintenance, single point of management
- Lower total idle power draw

Kubernetes/Ceph cluster is worth it **only** if you're specifically learning distributed systems — not for performance gains in a single-user homelab.

---
*@nikolatesla — The real cost of a homelab isn't the hardware — it's the electricity bill over 3 years. Always measure power draw before committing to a 24/7 machine.*

Homelab server build component guide

// COMMENTS

ON THIS PAGE

Homelab server build component guide

// COMMENTS ↓ Newest First

ON THIS PAGE

// COMMENTS