Hardware Basics
GPUs, RAM, cooling and overclocks — the parts that decide how fast and how big your models can be.
Your rig is what limits the size and speed of the brains you can train. Every component has trade-offs.
The four pillars
- GPU — drives training speed (tokens/sec) and model size (VRAM). The single most important upgrade.
- RAM — your dataset cache and active working set. Comes in four rarity tiers.
- Cooling — keeps the rig stable under overclock. Skimp here and you get crashes.
- HDD — checkpoint storage. Tight space = fewer saves to roll back to.
Overclocking
Push tokens/sec at the cost of stability and heat. Each component has its own headroom. A balanced rig beats a maxed rig every time.
Rarity tiers (RAM example)
- Common — base stats, easy to find, cheap.
- Uncommon — minor bonus, no downside.
- Rare — strong stats, often with a quirky downside.
- Exotic — best stats, unrepairable, lasts 2x longer than standard hardware.
Save your Cu for one good upgrade rather than four mediocre ones.