The Machine
This is a consumer-grade gaming PC repurposed as an AI server. No server hardware needed. Everything was already in the living room, connected to a TV.
CPU
AMD Ryzen 9 7950X3D
RAM
96 GB DDR5
GPU
AMD RX 7900 XTX 24GB
Root SSD
Disk 0: 2 TB (Predator GM9)
Data SSD
Disk 1: 4 TB (Fanxiang S790)
OS
Ubuntu 24.04 (Partition 5) + Windows (C:)
Storage Layout
Disk 0 (Predator SSD GM9 2TB) is dual-boot: Windows + Ubuntu on the same physical drive. Disk 1 (Fanxiang S790 4TB) is pure data.
Disk 0: Predator SSD GM9 2TB (System + Ubuntu)
| Partition | Size | Type | Purpose |
|---|---|---|---|
| 1 | 300 MB | System (EFI) | Windows boot |
| 2 | 16 MB | Reserved | Microsoft reserved |
| 3 | 1.6 TB | C: (NTFS) | Windows (creative powerhouse) |
| 4 | 841 MB | Recovery | Windows recovery |
| 5 | 300 GB | Linux (ext4) | Ubuntu (AI server, 24/7) |
Disk 1: Fanxiang S790 4TB (Data)
| Partition | Size | Filesystem | Mounted At | Purpose |
|---|---|---|---|---|
| D: | ~2 TB | NTFS | /media/wm/2TB-Data | Ollama models, Restic repo, Timeshift |
| E: | ~1.9 TB | NTFS | OS-BK | General storage, Macrium Reflect backups |
💡 Pro tip
Ollama models are 72 GB. Keeping them on the 4TB SSD via symlink (
/home/wm/.ollama/models → /media/wm/2TB-Data/ollama-models) dropped root partition usage from 73% to 47%.
Dual Boot
Same physical machine (Disk 0), two OS personalities. GRUB handles boot selection.
Ubuntu side (Partition 5, 300 GB)
- Role: Automated AI server, 24/7
- All AI services run here (Ollama, Open WebUI, Nextcloud, Jellyfin, AdGuard)
- All work done remotely via WeChat — rarely used as desktop in person
- Restic + Timeshift for backup
- HER (Ubuntu agent) — the primary workhorse
Windows side (C: drive, 1.6 TB)
- Role: Creative powerhouse, booted as needed
- Adobe Creative Suite (subscription)
- GOG (gaming library)
- Macrium Reflect (full disk imaging including Ubuntu Partition 5)
- LM Studio, Ollama (local AI inference)
- No Microsoft Office (deliberate choice — avoids school/work account sync, zero safeguarding complications)
- HER (Windows agent) — creative agent when booted
⚠️ Timeshift + NTFS conflict
Timeshift wants exclusive mount control during snapshots. If your data partition is mounted at boot via fstab, Timeshift will fail. Solution: use
noauto in fstab + a systemd oneshot service to mount after local-fs.target.
Why AMD GPU?
NVIDIA is the default choice for AI, but AMD ROCm works well for inference:
- 24 GB VRAM — runs 30B models comfortably (18 GB model + overhead)
- ROCm 6.2 — Ollama supports it natively
- Free — already owned the card
- Linux-first — ROCm is better supported on Linux than Windows
One quirk: the 7900 XTX reports as gfx1100 but needs HSA_OVERRIDE_GFX_VERSION=11.0.0 to work with ROCm's compatibility layer.
Total Cost
| Item | Cost | Note |
|---|---|---|
| PC hardware | $0 | Already owned |
| Ubuntu | $0 | Free |
| Ollama | $0 | Open source |
| Open WebUI | $0 | Open source (MIT) |
| Nextcloud | $0 | Open source (AGPL) |
| Jellyfin | $0 | Open source (GPL) |
| AdGuard Home | $0 | Open source |
| Restic + Timeshift | $0 | Both open source |
| Total | $0 | Existing hardware only |