My homelab is a constantly evolving platform for learning, self-hosting, and running serious workloads — from VMware vSphere clusters and AI inference to network automation and cloud integration. It lives in a dedicated server room, built around a StarTech 25U open-frame rack, and extends into the cloud via AWS and Cloudflare.
Table of Contents
At a Glance
| Compute Nodes | Total RAM | Usable Storage | Network Uplinks |
|---|---|---|---|
| 4 (3× NX-1365-G4 + HP Z840) | 1,280 GB | 95.9 TB | 25 GbE |
I would estimate the replacement value sits somewhere around £15–20k. This has been accumulated over the years through a mix of enterprise surplus finds and deliberate upgrades.
My hardware choices are driven by a consistent set of design principles — some are non-negotiable, others are strong preferences:
- vSphere compatibility — HCL-listed hardware wherever possible, so certifications mean something
- IPMI / remote management — everything must be operable headlessly
- Rack-mount form factor — clean, manageable, and expandable
- Low-power 24/7 operation — the lab runs continuously, so efficiency matters
- Noise isn’t really a factor due to location
- Heat Output isn’t a huge factor
- GPU is an essential in at least one node.
Physical Hardware
Overview
GPU/ Management Cluster
This is made up of a single HP Z840 running ESXi 8.0U3. It is an amazing workhorse and has been a key part of my lab (running multiple functions for over 6 years) It currently runs vSphere and I often use it for running Holodeck on top of vSphere.
| Component | Description |
|---|---|
| HP | Z840 |
| CPU | 2x CPU E5-2673 v3 @ 2.40GHz 30MB Cache, 5.0GTs, 105W |
| RAM | 256GB |
| NIC 0 | Supermicro Intel AOC-STGN-i2S 10Gb Dual Port |
| Storage 0 | Intel Optane DC P4800X used for Local NVMe Datastore |
| Storage 1 | Intel 2TB NVMe used for Local NVMe Datastore |
| GPU0 | NVIDIA GeForce GT 730 — display adapter only (the Z840 would not boot without a GPU installed; keeps the A10 free for compute) |
| GPU1 | NVIDIA Ampere A10 (24 GB) — GPU compute: AI inference, LLM serving, and ML workloads |
| Boot Volume | 2x1TB Raid 1 |
Compute Cluster
Nutanix NX – 3x NX-1365-G4
This is a significant resource for running a lot of my workloads however, due to electric costs, it does not run all the time. It is made up of 3x Nodes in a Supermicro Big-Twin with an identical configuration as below
This is running vSphere 8 with VSAN ESA being used as storage
| Component | Description |
|---|---|
| Model | NXS2U4NL12G400 |
| CPU | 2x XEON E5-2640 V3 |
| RAM | 256GB |
| NIC | Intel Ethernet Controller XXV710 for 25GbE SFP28 |
| Boot Disk | 64 GB SATADOM |
| SSD 1 | Samsung PM863 960GB |
| SSD 2 | Samsung PM863 960GB |
| SSD 3 | Samsung PM863 960GB |
Storage
Primary Storage
Quanta D51PH-1ULH (66.3TiB usable)
Quanta server running TrueNAS scale. This was originally configured with striped mirrors. I have changed this to a RAIDZ2 config, giving me a usable 66.3 TB This is presented as SMB/NFS and iSCSI to multiple environments. More details are here
Secondary Storage
Synology DS1512+ 4x8TB SHR1 ( 21.8TiB Usable)
This is in my office and used for “offsite backups”
| Hostname | Model | NVMe Capacity (TB) | SSD Capacity (TB) | HDD Capacity (TB) |
|---|---|---|---|---|
| HP | HP Z840 | 2.6 | ||
| vSAN ESA | 0 | 7.42 | ||
| Quanta D51PH-1ULH | 0 | 0 | 66.3 | |
| Synology DS1512+ | 21.8 | |||
| Total TB (Usable) | 2.6 | 7.42 | 95.9 |
Network
Physical Network
| Model | Description |
|---|---|
| WatchGuard M400 | Core Firewall |
| Mikrotik CRS504-4XQ-IN | Core Fibre Switch |
| Mikrotik CSS610-8P-2S+IN | POE Switch |
| Mikrotik CRS305-1G-4S+ | Fibre to Ethernet Switch |
Wireless Network
| Ap Model | Description |
|---|---|
| Ubiquiti Unifi AC-Pro | Access Point Office |
| Ubiquiti UniFi 6 Pro | Access Point House |
WAN
| Wan Providers | Description |
|---|---|
| Zen FTTP (Via CityFibre) | 900/900 with a /29 IPV4 |