- Choose hardware based on your real use case, not marketing.
- Understand the difference between RAM, VRAM, and context.
- Avoid buying the wrong hardware for Ollama, RAG, and local coding.
Quick reference by goal
Learn and experiment: RAM: 16 GB VRAM: 6-8 GB or Apple Silicon with unified memory Models: 3B-8B Q4 Private RAG with documents: RAM: 32 GB recommended VRAM: 8-12 GB Models: 7B/8B Q4-Q5 + local embeddings Reasonable local coding: RAM: 32 GB VRAM: 12-16 GB Models: Qwen/DeepSeek coder 7B-14B Q4-Q5 Agents and long-running tasks: RAM: 64 GB or more VRAM: 16-24 GB or more Models: 14B-32B, controlled context and logs
NVIDIA, AMD, and Apple Silicon
- NVIDIA: usually the most straightforward path for GPU acceleration on Windows/Linux thanks to the CUDA ecosystem.
- AMD: can work very well, but depends more on drivers, ROCm/Vulkan, operating system, and tooling.
- Apple Silicon: unified memory helps a lot; look at total RAM and bandwidth, not just the chip name.
- CPU-only: fine for learning and small models, but don't expect fast agents.
Checking your hardware
# Windows / PowerShell systeminfo | findstr /C:"Total Physical Memory" wmic path win32_VideoController get name,adapterram nvidia-smi # macOS system_profiler SPHardwareDataType system_profiler SPDisplaysDataType # Linux free -h lspci | grep -Ei "vga|3d|display" nvidia-smi