February 28thDeepSeek The University of Beijing and the University of Tsinghua are on the linepaperThe official launch of the new technology programme, DualPath, focused on addressing the historical data reading bottlenecks encountered by the AI Large Model in carrying out complex multi-wheel missions。
ACCORDING TO THE PAPER, NOW THE AI SYSTEM HANDLES THE SUPER-LONG CONTEXT WITH TWO COMPUTING MODULES FOR "PROCESSING INPUT INFORMATION " AND " GENERATING TEXT ANSWERS " , WHERE DATA CHANNEL RESOURCES ARE MISMATCHED。
In response to this questionThe new DualPath breaks the normal single-line transmission limit, allowing historical data to enter the `generation of answers' module first through an idle channel and then to be forwarded to the `processing input' module in an instant, using a high-speed network within the cluster。
ACCORDING TO OFFICIAL EVALUATION DATA, WHEN DEALING WITH REAL AND COMPLEX AI TASKS, IT IS NOT POSSIBLE TO FIND A WAY TO DO THISThe technology increases the amount of offline treatment throughput by a maximum of 1.87 times, and online service throughput by an average of 1.96 times。
At the moment, large models are moving rapidly in the direction of "Agent" with the ability to plan autonomously, and AI needs to look at the context of tens of thousands of words, leading to a shift in systemic capacity constraints from "incompetent" to "too slow" data transmission。
The validation of the DualPath technology, on the other hand, demonstrated that, by optimizing the bottom-level data flow, a large amount of idle resources could be generated without increasing the cost of hardware procurement。
DeepSeek, you know what's still waiting for. V4 still not present