Prefill
The compute-bound phase where the AI reads your prompt
Fundamentals
In plain English
The AI reading and absorbing the entire test question at once.
Technically
The first phase of LLM inference where the entire prompt token sequence is evaluated in parallel. Because large matrix multiplications occur simultaneously across all input tokens, arithmetic intensity is high (>100 FLOPs/byte), making prefill bound by GPU compute TFLOPs.