Data released on November 17 shows that domestic average daily token calls have exceeded 500 trillion, driving a surge in demand for computing power. In the short term, intelligent computing supply is insufficient, making it necessary to improve efficiency by tapping into existing capacity. In 2025, Mooncake, a large-model inference system developed by Academician Zheng Weimin's team, won the Eric Riedel Best Paper Award. Its core innovation is PD separation technology: "P" handles reading user queries, while "D" handles generating answers. Under the traditional approach, the same pool of computing power alternates between the two tasks. PD separation, by contrast, assigns one group to read queries exclusively and pass the results to another group dedicated to answering, thereby revitalizing existing computing resources. At the same hardware scale, the system handles 75% more user requests, and long-text throughput capacity increases more than fivefold.
