Context Window vs Generation

Processing the input prompt (prefill) is fast because all tokens are processed in parallel. But generation (decode) is slow because tokens are produced one at a time. What is the key architectural difference? Flag format: CONGRESS{difference_in_snake_case}