di03 and di04: The Utility Servers
Not every workload needs a Threadripper.
Before deploying the heavy iron, I brought up the utility tier. di03 and di04. Smaller machines. Essential roles.
di03: Utility (Capable) #
| Component | Spec |
|---|---|
| CPU | Ryzen 9 5950X (16 cores) |
| RAM | 128GB DDR4 |
| GPU | RTX 4090 (AIO watercooled) |
| VRAM | 24GB |
| Storage | NVMe + NFS from nexus |
| Fans | Green (Simon Says) |
The “capable” utility server. 14B-32B models run comfortably. Heavier batch processing. Backup inference capacity when the big machines are busy.
Self-contained AIO watercooling on the 4090. CPU stays air cooled. Thermal headroom matters for sustained workloads. This machine can run hot for hours without throttling.
di04: Utility (Light) #
| Component | Spec |
|---|---|
| CPU | Ryzen 9 3900X (12 cores) |
| RAM | 96GB DDR4 |
| GPU | RTX 4060 Ti |
| VRAM | 8GB |
| Storage | NVMe + NFS from nexus |
| Fans | Blue (Simon Says) |
The light duty server. 7-8B models. Transcription with Whisper. OCR. Embeddings generation. The small fast stuff that doesn’t need heavy iron but still needs to run locally.
8GB VRAM is limiting, but that’s the point. This machine handles the workloads where 8GB is enough, freeing the bigger machines for bigger work.
The Build #
Both machines: MSI Prestige X570 Creation motherboards. Solid, stable, nothing exotic. Clean NixOS installs following the standardized config.
di03 took an afternoon. di04 took a few hours. The patterns were established. The configs were ready. Deploy, test, done.
No drama. That’s the goal.
Simon Says #
The color scheme matters more than it seems.
- di01: Red (highest priority)
- di02: Yellow
- di03: Green
- di04: Blue (lowest priority)
Walk into the server room. Know instantly which machine is which. When you’re debugging at 2am, visual identification saves time.
di03 glows green. di04 glows blue. No guessing.
Why Utility Tier Matters #
Not every task deserves the flagship.
Transcribing a meeting? di04. Generating embeddings for a document set? di04. Running a 14B model for quick iteration? di03.
The heavy iron (di01, di02) stays available for the work that actually needs it. 70B inference. Training runs. The stuff where VRAM and compute actually matter.
Fleet architecture isn’t about having the biggest machines. It’s about having the right machine for each job.
Centralized Storage in Action #
Both machines mount their important storage from nexus over NFS. Local NVMe for scratch and cache. Everything that matters lives on ZFS.
If di03 dies tomorrow, rebuild it from NixOS config, mount the storage, keep working. The machine is replaceable. The data is not.
This is the pattern. Utility servers proved it works before the heavy iron went online.
Part 9 of the Node0 build series. Next: di02: Production Server, the stable inference workhorse.