Resources · AI

AI workstation spec guide.

The decisions that actually matter when you're sizing a local AI machine.

GPU

VRAM is usually the constraint

  • Model size + quantization sets the floor
  • Batch size and context length push the ceiling
  • Multi-GPU spreads VRAM but adds latency and complexity
Platform

CPU, RAM, and PCIe

  • PCIe lane count matters for multi-GPU and NVMe
  • RAM at least 2x total VRAM is a good baseline
  • Single-thread performance still matters for data prep
Cooling and power

Sustained load, not peak

  • PSU sized for sustained multi-GPU + headroom
  • Case airflow over GPU shrouds
  • Noise budget for the room
Software

Stack notes

  • Linux + CUDA for most production stacks
  • Ollama / llama.cpp / vLLM trade-offs
  • Container vs bare-metal driver paths