Glossary: AI System Design
The time a model spends reading the prompt before it writes. It lowers real GPU capacity below the output-speed figure.
UtilizationTTFT (time to first token)