---
title: "AI hardware guide"
description: "The hardware under modern AI — GPUs, TPUs, and interconnects, plus what to actually buy for a home training setup."
canonical_url: "https://fanout.sh/ai/hardware"
md_url: "https://fanout.sh/ai/hardware.md"
access: "public"
---

# AI hardware guide

The hardware under modern AI — GPUs, TPUs, and interconnects, plus what to actually buy for a home training setup.

## Topics

### GPU vs CPU for ML

**Why GPUs?** Neural network training is massively parallel matrix multiplication. GPUs have thousands of cores optimized for this.

**Key metrics:**
- **VRAM** — most important. Determines max model size you can load
- **TFLOPS** — raw compute speed (FP16/BF16 matters most for ML)
- **Memory bandwidth** — how fast data moves to/from GPU memory
- **Tensor Cores** — specialized hardware for matrix ops (NVIDIA)

**Rule of thumb:** A 7B parameter model in FP16 needs ~14GB VRAM. In 4-bit quantization, ~4GB.

---
This representation contains public Fanout content only. Protected Pro lessons, account data, billing, checkout, and pricing are not included.

Browse the public content map: https://fanout.sh/sitemap.md