Lab & AI

The lab behind Rocking JA

Five hypervisor hosts on two platforms, 112 GB of GPU memory across three NVIDIA cards (including a 96 GB private AI server), and about 110 workloads, all built, monitored and backed up by me.

  • 5hypervisor hosts
  • ~264CPU threads
  • ~1.4 TBRAM
  • 112 GBGPU VRAM (lab total)
  • ~42 TiBusable storage
  • ~110workloads

Compute

Proxmox VE 9.2 (2 nodes) · VMware vSphere 8.0 U3 (vCenter, 3 ESXi hosts across 3 datacenters)

Dell PowerEdge R740

2× Xeon Platinum 8276

Cores
56C / 112T
Memory
512 GB

Dell PowerEdge R7515

AMD EPYC 7452

Cores
32C / 64T
Memory
512 GB
GPU
NVIDIA RTX A4000 (16 GB)

Supermicro H13SAE-MF

AMD Ryzen 9 9950X

Cores
16C / 32T
Memory
192 GB DDR5
GPU
2× Quadro RTX 8000 (2× 48 GB)

ASRock Rack X470D4U

AMD Ryzen 9 5950X

Cores
16C / 32T
Memory
128 GB DDR4

Dell PowerEdge R720

2× Xeon E5-2630 v2

Cores
12C / 24T
Memory
96 GB

Workloads

About 110 running

vSphere VMs
60
Proxmox VMs
10
LXC containers
5
Docker / Podman
35

GPUs

112 GB of VRAM across three NVIDIA cards in two servers

Supermicro H13SAE-MF

2× NVIDIA Quadro RTX 8000

VRAM
2× 48 GB = 96 GB
Role
Supreme Intelligence AI server

Dell PowerEdge R7515

NVIDIA RTX A4000

VRAM
16 GB
Host
EPYC 7452, 512 GB

Lab total

112 GB GPU memory

Quadro RTX 8000
96 GB
RTX A4000
16 GB

Private AI server

Supreme Intelligence

A self-hosted AI server with 96 GB of its own GPU memory: two Quadro RTX 8000 cards (48 GB each) passed through to a Fedora 43 VM. (The RTX A4000 is separate, bringing the lab total to 112 GB.)

Language models

llama.cpp across two GPUs

Models are split across both cards, which talk over PCIe, with a 262K-token context window. The default model is Qwen3.8-27B (8-bit, reads images), with Qwen3.8-27B-Uncensored or Gemma-4-31B ready to swap in.

Image generation

FLUX.2-dev, locally

Images are generated on the same GPUs, so prompts and results never leave the lab.

Assistant

Private ChatGPT-style chat

Open WebUI with web search and image input, plus tools for email, weather, radar and Windy maps, stock quotes and returning documents. Ask Grok and Ask Claude hand-offs reach cloud models when needed.

Automation

Agents and hardware care

A scheduled daily morning-brief agent, custom GPU fan control and a live hardware status page.

Storage

  • Synology DS1823xs+ with 8 drives (4× 12 TB and 4× 4 TB) in two RAID 5 pools, about 42 TiB usable
  • NFS to the VMware hosts, plus SMB shares
  • Proxmox Backup Server for VM backups
  • NVMe storage, including a mirrored pair of Samsung 990 PRO drives

Network

  • Netgate 6100 firewall running pfSense Plus
  • Tailscale and OpenVPN for remote access
  • 10 GbE uplinks (Intel X520, Mellanox ConnectX-4 Lx)
  • Separate segments for the lab, the DMZ and storage
  • Public apps behind Cloudflare Tunnel with Zero Trust login

Self-hosted services

  • Gitea
  • Nginx Proxy Manager
  • Grafana
  • VictoriaMetrics
  • Telegraf
  • Emby
  • Dispatcharr
  • Guacamole
  • Isolated web browser
  • Private file drive
  • Camera NVR bridge
  • Nest monitor
  • Power-grid monitor
  • Active Directory
  • Kasm
  • Mail server
  • Game server

Need capacity?

Rent a VM on this hardware

Tell me the vCPU, memory, storage, GPU and OS you need and I’ll follow up with availability and a quote.

VM hosting inquiry