FIELD MANUAL v1.0

The Local AI Architect: A Field Manual for Inference on Consumer eGPU Hardware

Stop renting the cloud. Stop fighting broken abstractions.

Lock your hardware into maximum performance.

Get the Manual — $14.99 Peek Inside ↓

PDF + Markdown. Instant delivery. 30-day guarantee.

Cover

If you have tried connecting an external NVIDIA GPU to an AMD mini-PC over Oculink on Linux, you already know the standard cloud-native playbook is completely broken.

Out of the box, aggressive kernel power management, mobile-driver heuristics, and fragile USB-C/SFX power delivery will aggressively downtrain your Gen4 x4 link to a crawl—or drop the bus entirely right in the middle of a heavy prompt evaluation.

What's Inside

Three parts. Zero fluff. Every line is distilled from a real failure mode.

🔌

Part I: The Physical Layer

  • • Demystifying Oculink signal degradation & Gen4 link training
  • • The "German Shepherd Failure Mode"
  • • Kernel cmdline: pcie_aspm=off, pcie_port_pm=off
  • • Fixing the "vanishing GPU" with raw udev rescan rules
⚡

Part II: Power & Bare-Metal Serving

  • • Why transient spikes kill SFX units during prompt evals
  • • Purging tlp / auto-cpufreq, locking CPU governors
  • • Raw llama.cpp server compiled from source
  • • --no-mmap discipline: keep paging off the Oculink cable
🌐

Part III: Operations & Gateway Topology

  • • Minimal Caddy reverse proxy with self-signed TLS
  • • Local DNS (ai.local) termination
  • • systemd lifecycle over heavy monitoring stacks
  • • Situational awareness via journalctl & nvidia-smi

📋 Appendices: The Blueprints

Appendix A — Complete unified file manifests: GRUB, sysctl, modprobe.d, mkinitcpio hooks, and systemd units.

Appendix C — Mechanical isolation & "The Bed Variable" for soft/dynamic surfaces.

Appendix D & E — oculink_watchdog.sh, S0ix mitigations, 5-min troubleshooting tree.

The OCABrace Blueprint — Full BOM, assembly, and CAD/STL for a modular strain-relief bracket.

Who Is This For?

🐧

Linux sysadmins running Arch, CachyOS, or performance-tuned kernels who are done debugging link drops by hand.

🤖

Hobbyists deploying autonomous agents locally and tired of sub-1-token/s latency from cloud round-trips.

🔧

Anyone who has watched nvidia-smi vanish mid-generation because the PCIe bus decided to take a nap.

Frequently Asked

What formats do I get?
PDF (print-ready) and Markdown (for grep, version control, or modification). Both are included in every purchase.
Does this cover Windows or macOS?
No. This is a Linux-native manual. The commands, kernel parameters, and systemd units are all written for the Linux kernel. If you run Windows, the high-level concepts transfer but the specific incantations do not.
Do I need a specific GPU or enclosure?
The manual is written around NVIDIA dGPUs over Oculink/PCIe on AMD mini-PCs (the most common and most frustrating setup). The principles apply broadly, but the specific recipes target this stack.
What if the fix doesn't work for my setup?
There's a 5-minute troubleshooting decision tree in Appendix D, plus the oculink_watchdog.sh monitoring script to help you isolate your specific failure mode. Every fix is labeled with its scope so you know what's safe to try.

Stop Fighting Your Configuration.

Own your silicon. Get the battle-tested field manual and lock your eGPU into rock-solid bare-metal inference performance.

Get the Manual — $14.99

Instant download. 30-day money-back guarantee. No DRM, no expiration.