GPU Swap: GT 710 → GTX 1080

Part A — Software prep (do now, before powering off)

Clean out the failed nvidia-610 leftovers and the CUDA repo so the new driver installs cleanly. SSH in and run:

• sudo rm -f /etc/apt/sources.list.d/cuda-ubuntu2204-x86_64.list
• sudo apt-get purge -y '~nnvidia' 'cuda*' 2>/dev/null; sudo apt-get autoremove -y
• sudo apt-get update

You're now on nouveau, clean. Then shut down: sudo poweroff.

Part B — Physical swap (power removed)

1. Unplug the power cable from the PSU, then hold the case power button ~5 s to drain residual charge.
2. Ground yourself (touch bare metal of the case); avoid carpet.
3. Open the side panel; locate the small GT 710 in its PCIe x16 slot.
4. Before committing, verify the two physical gates — the whole swap hinges on these:
   • A spare 8-pin (6+2) PCIe power cable from the PSU. The GT 710 uses none; the GTX 1080 needs one (~180 W). If there is no spare PCIe cable, stop — do not use Molex/SATA→8-pin adapters on a 180 W card; you need a PSU with a native PCIe lead.
   • Clearance for a 2-slot, ~267 mm card — nothing (drives, cables) fouling the length or the slot directly below.
5. Remove the GT 710: unscrew the bracket, press the slot's retention latch, pull straight out.
6. Seat the GTX 1080 in the x16 slot (you can reuse the GT 710's slot) — press evenly until the latch clicks; screw the bracket down.
7. Connect the 8-pin PCIe power — push until it clicks.
8. Re-check: card seated, latch engaged, power connected, fan clear. Close the panel; reconnect power.

Part C — First boot + driver (headless, over SSH)

Secure Boot is off, so there is no module-signing / MOK step.

1. Power on, wait ~1–2 min, then SSH in: ssh arcade@tv.amiheines.com. (If it will not come up, plug a monitor into the 1080's DisplayPort/HDMI to watch POST — but it booted headless before, so it should again.)
2. Confirm detection: lspci -nn | grep -i nvidia → expect GP104 [GeForce GTX 1080].
3. Install the current driver (Pascal → 550-series, not the 610 that failed):
   • sudo apt-get update
   • ubuntu-drivers devices — should now recommend nvidia-driver-550 (or newer)
   • sudo apt-get install -y nvidia-driver-550
   • sudo reboot
4. Reconnect via SSH, then verify: nvidia-smi → should list GeForce GTX 1080, ~8192 MiB, and a driver version.
5. Hand ollama the GPU (it auto-detects CUDA once the driver is up):
   • sudo systemctl restart ollama
   • journalctl -u ollama --no-pager | grep -iE "gpu|cuda|vram" | tail — should show the GPU / VRAM, not library=cpu.
6. Prove it: in one SSH session run watch -n1 nvidia-smi; in another ollama run llama3.1:8b "hello" — you should see VRAM used + GPU utilization, and generation much faster than CPU.

Notes

• Model sizing: 8 GB VRAM fits llama3.1:8b (Q4 ~5 GB) fully → fast. qwen2.5:14b (~9 GB) is larger than 8 GB, so ollama offloads some layers to CPU (partial speedup). For full-GPU speed favor 7–8 B models; keep 14 B as the CPU / higher-quality option.
• No integrated graphics on Threadripper, so the 1080 becomes the only GPU — fine for a headless box. Want a local console later? Plug a monitor into the 1080.
• Rollback: anything wrong → power off, reinstall the GT 710, boot; you are back to the known-good state. Nothing else on the box is affected by the swap.
• Do not re-run the ollama install script — it will try to pull the broken CUDA-610 driver again. The distro nvidia-driver-550 is all you need.