Posts

A hands-on curriculum for anyone who wants to stop guessing at performance and start understanding the machine. Modules # Module Status 1 The Silicon Sandbox – Understanding Hardware Thermal Limits Read now 2 The Kernel’s Kitchen – Scheduling, Governors, and Power States Read now 3 The Memory Hierarchy – Cache, Swap, and Allocation Strategy Coming soon 4 The I/O Labyrinth – Disk, Network, and Bus Bottlenecks Coming soon 5 The Tuning Toolkit – Profiling, Benchmarking, and Validation Coming soon Last updated: June 16, 2026

If you’ve tried to build a compact, bare-metal local AI inference node using a modern AMD mini-PC and an external NVIDIA dGPU via Oculink, you’ve probably stared at a screen full of AER (Advanced Error Reporting) logs right before your system dropped the link down to a crawl. Standard consumer hardware handles PCIe power management and driver heuristics like a mobile laptop. The moment your LLM server drops to idle, the kernel downshifts your lane speeds. The moment a heavy prompt eval hits, transient power spikes or minor cable torque kills the connection entirely. ...

Previous Post in Series: The Silicon Sandbox – Understanding Hardware Thermal Limits Taming the Beast – Linux Kernel Optimization & Power Management The Hook: Taking control away from conservative default OS behaviors to maximize throughput. In the first part of this series, “The Silicon Sandbox,” we ventured into the physical battlefield of thermal limits, understanding how raw physics dictates the performance ceiling of your hardware. We saw how easily a compact, passively cooled system can suffocate under its own heat, throttling performance long before you even realize it. ...

I built something weird and useful: a biometric hardware monitor that maps your system’s physical strain into an organic, beating heart on your desktop. It’s called Anima System Heartbeat 🫀, and it turns CPU/GPU load into a real-time visual and auditory pulse. Because why stare at boring graphs when your machine can literally have a heartbeat? What It Does The project ships two flavors: Desktop Widget (kimi_desktop_pulse.py): A borderless, floating heart that lives on your Linux desktop. Drag it wherever you want. Right-click to kill it. It beats faster when your system is stressed, slower when it’s idle. Web Dashboard (kimi_heartbeat.py): A lightweight Flask-based local dashboard with full metric breakdowns and a stylized CSS pulse. The Nerdy Bits Unified Stress Tracking: It samples both CPU usage and NVIDIA GPU utilization, then uses whichever is higher (max(cpu, gpu)) to drive the pulse engine. This means GPU-heavy workloads (rendering, training, gaming) register just as hard as CPU-bound tasks. Heart Rate Variability (HRV): I built in Gaussian distribution variance so the intervals between beats fluctuate naturally. No robotic metronome — it feels alive. Thermal Color Shifting: The heart dynamically shifts gradients based on load intensity: Cyan — Idle, cool, chill Amber — Active, warming up Crimson — Peak stress, your system is sweating Native Audio Streams: Raw 16-bit signed PCM thumps piped straight into aplay asynchronously. No PulseAudio/PipeWire deadlocks, no audio server drama. Just clean, immediate thumps. Why I Built This I spend a lot of time staring at htop, nvidia-smi, and temperature graphs. They’re functional but sterile. Anima Heartbeat gives me ambient awareness of what my machine is doing without eating screen real estate or cognitive bandwidth. It’s a glanceable biometric for silicon. ...

This is the first post on archdaemon.xyz. I built this site to have a single place for my technical work — thermal engineering notes, Linux optimizations, and whatever else I’m tinkering with. The domain name?

An automated, idempotent system tuning utility engineered specifically for Small Form Factor (SFF) Mini PCs running modern Linux distributions paired with external graphics infrastructure (Oculink / dedicated PCIe links). This engine eliminates aggressive power-state downshifting, bypasses mobile driver thermal throttling, stabilizes LUKS storage hooks, and tunes virtual memory boundaries to guarantee zero-latency throughput for heavy developer compute, deep learning inference, and high-frame-rate rendering. The Core Problem (Why This Exists) Standard Linux distributions are aggressively optimized for mobile power conservation. When running a desktop discrete graphics card through a dedicated Oculink interface, the kernel’s default power governors constantly drop the link down to PCIe Gen 1 speeds during idle periods. ...

Your Computer Is Dying of Heat Before You Even Touch It Every system you will ever tune is already losing a war against physics. Silicon doesn’t care about your benchmarks, your uptime, or your carefully crafted dotfiles. It cares about one thing: staying below its thermal ceiling. Cross that line and the hardware either throttles itself into uselessness or cooks itself into an early grave. Small-form-factor mini PCs are the worst offenders. The same chips that live comfortably in tower cases with 240mm AIOs get crammed into boxes the size of a paperback book, breathing through pinhole vents and praying the ambient room isn’t summer-hot. If you’re running an SFF machine — a NUC, a DeskMini, a custom mini-ITX build, one of the new wave of handhelds — you’re working inside a thermal pressure cooker. Understanding that environment isn’t optional. It’s the foundation everything else rests on. ...

Previous Post in Series: Taming the Beast – Linux Kernel Optimization & Power Management Unleashing the Pipe – Oculink, eGPUs, and PCIe Stability Demystifying external graphics bottlenecks and maintaining a stable Gen 4 connection without signal degradation. For years, external graphics cards (eGPUs) on Linux were a compromise whispered about in niche hardware forums. You bought an expensive enclosure, wrestled with Thunderbolt 3 daemon configurations, and ultimately accepted a massive performance penalty. Thunderbolt’s 40 Gbps cap—which translates to roughly 32 Gbps of actual data throughput after display protocol overhead—forced an external GPU to squeeze through the equivalent of a microscopic PCIe Gen 3 x4 pipeline. ...