-+ 0.00%
-+ 0.00%
-+ 0.00%

According to the latest data, the average number of words called per day in the country has exceeded 500 trillion dollars, and demand in the computing power market has skyrocketed. Intelligent computing power cannot be made up in the short term; it can only benefit stocks. In 2025, Mooncake, a large model reasoning system developed by Academician Zheng Weimin's team, won the Eric Riddle Best Paper Award. The core is to use PD separation technology to revitalize existing computing power and use software scheduling to make full use of existing computing power. According to reports, P is responsible for “reading user questions” and D is responsible for “generating answers.” Under the traditional method, the same batch of computing power takes turns doing two things; PD separates one set of read-only questions and transfers the results to another set of specific answers. At the same hardware scale, the system accepts 75% more user requests, and the long text throughput capacity is increased by more than five times.

Zhitongcaijing·11/17/2026 14:17:22
Listen to the news
According to the latest data, the average number of words called per day in the country has exceeded 500 trillion dollars, and demand in the computing power market has skyrocketed. Intelligent computing power cannot be made up in the short term; it can only benefit stocks. In 2025, Mooncake, a large model reasoning system developed by Academician Zheng Weimin's team, won the Eric Riddle Best Paper Award. The core is to use PD separation technology to revitalize existing computing power and use software scheduling to make full use of existing computing power. According to reports, P is responsible for “reading user questions” and D is responsible for “generating answers.” Under the traditional method, the same batch of computing power takes turns doing two things; PD separates one set of read-only questions and transfers the results to another set of specific answers. At the same hardware scale, the system accepts 75% more user requests, and the long text throughput capacity is increased by more than five times.