ai-rig-upgrade

Source

  • Type: local-file
  • Path: /home/topher/.openclaw/workspace-crash-bot/projects/ai-rig-upgrade.md
  • Bytes: 4332
  • Updated: 2026-05-02T21:15:15.519Z

Content

# AI Rig Upgrade
 
**Status:** Active — GPU shopping
**Created:** 2026-04-20
**Updated:** 2026-04-20
**Tags:** `gpu`, `ollama`, `local-ai`, `aurora-r3`, `upgrade`
 
## Summary
 
Upgrade the Dell Aurora R3 (i7-2600K, 32GB DDR3, Thermaltake 600W) with a used GPU to enable fast local AI inference for OpenClaw, ComBadge, Tricorder, and other AI-enabled projects.
 
**Goal:** 10-20+ tokens/sec on 7B-13B models. Fast enough to iterate on projects without fighting the tool.
 
## Decision Tree
 
### Why GPU (not Mac mini or new build)
- Already has a working Linux box — no new machine to manage
- OpenClaw is already set up on it
- Budget comfort zone: ~$200-250 speculative buy
- Keeps everything in one place (no multi-box management overhead)
- Enables ComBadge + Tricorder development (both need local AI backend)
 
### Why not Mac mini
- Non-upgradeable, expensive ($700-1000+)
- Less "hack factor" for this crew
- Would still need ZFS storage solution
 
### Why not two-box (ZFS + AI separate)
- Split became $300+ before GPU — too expensive
- ZFS is nice-to-have, not a burning need (6TB works fine)
- Park ZFS for later, focus on AI now
 
## Current System
 
| Component | Detail |
|-----------|--------|
| **Motherboard** | Dell Aurora R3 (standard ATX, aftermarket mobo in cheap case) |
| **CPU** | Intel i7-2600K (Sandy Bridge, 6-core @ 3.40GHz) |
| **RAM** | 32GB DDR3 @ 1600MHz |
| **PSU** | Thermaltake 600W (new, 2×8-pin PCIe connectors free) |
| **Current GPU** | Quadro K600 (1GB, display only — to be removed) |
| **PCIe** | PCIe 2.0 x16 slot (free), no ReBAR support on this platform |
 
## GPU Candidates
 
> Target: used card, 8-pin PCIe power, HDMI or DP output, within $200-250 budget
 
| GPU | VRAM | TDP | Power Conn. | Used Price | Priority |
|-----|------|-----|-------------|-----------|----------|
| **RTX 3060 12GB** | 12GB | 170W | 1×8-pin | $180-230 | ⭐ Primary target |
| GTX 1660 Super | 6GB | 125W | 1×8-pin | $100-140 | Budget fallback |
| RTX 3060 Ti | 8GB | 200W | 1×8-pin | $170-220 | Alternative |
| RTX 2060 Super | 8GB | 175W | 1×8-pin | $130-170 | Older gen fallback |
| RTX 2070 | 8GB | 185W | 1×8-pin | $150-200 | If found cheap |
| RTX 4060 Ti 16GB | 16GB | 160W | 1×8-pin | $330-400 | If budget allows |
 
### Why RTX 3060 12GB
- 12GB VRAM — handles 7B Q5 and 13B Q4 models comfortably
- 170W TDP — fits within 600W PSU headroom
- Single 8-pin — Thermaltake has 2 of these free
- PCIe 2.0 compatible — no ReBAR needed
- Standard dual-fan or blower — fits in standard case
- HDMI + 3×DisplayPort — multiple display options
 
### Avoid
- RTX 4070+ (needs 2×8-pin or 12-pin, too power-hungry for 600W)
- Cards without 8-pin PCIe connectors
- Single-fan thermal designs (will throttle in enclosed case)
 
## eBay Search Terms
 
```
RTX 3060 12GB
RTX 3060 ti 8GB
RTX 2060 super
RTX 2070
```
 
**Filters:**
- Seller rating 50+
- Multi-fan or blower style (not single-fan)
- "Works" or "tested" in description
- Check compatibility with older PCIe generation
 
## Expected Performance
 
With RTX 3060 12GB:
| Model | Quantization | Expected Speed |
|-------|-------------|---------------|
| 7B | Q4 | ~30-40 tok/sec |
| 7B | Q8 | ~40+ tok/sec |
| 13B | Q4 | ~15-25 tok/sec |
| 13B | Q5 | ~10-15 tok/sec |
 
## Next Steps
 
- [x] Confirm hardware: Aurora R3 mobo, i7-2600K, 32GB DDR3, 600W Thermaltake
- [x] Remove Quadro K600
- [x] Identify PCIe slot and power connectors
- [ ] Source RTX 3060 12GB (~$200-230 on eBay)
- [ ] Install GPU, install drivers
- [ ] Configure Ollama with GPU support
- [ ] Test inference speed with reference model
- [ ] Verify OpenClaw integration
 
## Related Projects
 
- [[com-badge.md]] — ComBadge wearable (needs local AI)
- [[tricorder.md]] — Tricorder handheld (needs local AI)
- [[ai-desktop-companion.md]] — StackChan desktop robot
- [[zfs-casaos.md]] — ZFS/storage (parked for now, not blocking)
 
## Notes
 
- PCIe 2.0 x16 is bandwidth-limited vs 3.0/4.0, but inference is compute-bound not bandwidth-bound — won't be the bottleneck for these model sizes
- No ReBAR support means some newer optimizations won't work, but doesn't block standard inference
- 600W PSU calculation: ~95W CPU + ~50W system + 170W GPU = ~315W under full load, plenty of headroom
 
---
 
*Last updated: 2026-04-20 — resolved to RTX 3060 12GB, eBay shopping phase*
 

Notes

Referenced By