一种分布式注意力计算系统
By optimizing task partitioning and scheduling based on physical core distance awareness in a distributed attention computing system, the problem of misaligned physical proximity in communication paths in existing technologies is solved, achieving a high-efficiency wafer-level chip computing performance improvement.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING TSINGMICRO INTELLIGENT TECH CO LTD
- Filing Date
- 2026-04-13
- Publication Date
- 2026-07-17
AI Technical Summary
Existing technologies in distributed attention computing systems suffer from misaligned physical proximity of communication paths, failing to fully utilize the ultra-high bandwidth of wafer-level chips, resulting in a performance gap. This is especially true when wafer-level chips are combined with GPU clusters, where communication latency accounts for a high percentage of performance, and the overall performance is lower than that of a single GPU.
By dividing the unit and scheduling unit, the input sequence requests are divided into tasks according to the mapping relationship between the logical blocks and the physical cores of the wafer-level chip, and dynamic scheduling calculations based on physical core distance awareness are performed. The query vector and key value vector group members are mapped to adjacent physical core locations using a preset mapping function to optimize the communication path.
It significantly reduces communication overhead, improves computing utilization and system throughput, realizes high-bandwidth, low-latency on-chip data exchange, makes full use of the communication bandwidth of wafer-level chips, and improves computing performance.
Smart Images
Figure CN122019113B_ABST