# GPU Prefill vs. Single-Token Decode: Why Prefill Is Faster

## The Core Difference: Compute vs. Memory Bound

**Prefill** (processing the prompt) processes many tokens **in parallel**, while **decode** (generating each new token) processes **one token at a time**. This leads to fundamentally different performance characteristics.

## 1. Arithmetic Intensity

- **Prefill**: All prompt tokens flow through the model simultaneously. Matrix multiplications have large dimensions (e.g., `[batch, seq_len, d_model] × [d_model, d_model]`), so each weight loaded from