Cluster device for gpu and internal cache management method thereof

CN116881192BActive Publication Date: 2026-10-27NAT INNOVATION INST OF DEFENSE TECH PLA ACAD OF MILITARY SCI
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202310660072.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-06-06
Publication Date
2026-10-27
Estimated Expiration
2043-06-06

AI Technical Summary

Technical Problem

由于GPU程序在运行中会产生大量的访存请求,会使得L1 cache面临较大的失效率,进而导致流处理器中的线程经常因等待数据而处于暂停的状态,极大制约了流处理器的性能,对GPU的性能造成严重影响

Benefits of technology

本发明的用于GPU的cluster装置及其内部一级缓存管理方法能够在设置cluster装置以降低片上互连网络和GPU的硬件开销的基础上,实现cluster装置内部的各个流处理器的一级缓存共享,能够充分利用GPU中的一级缓存资源,提高流处理器和GPU的性能,提高系统资源利用率。

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116881192B_ABST
    Figure CN116881192B_ABST
Patent Text Reader

Abstract

The application discloses a cluster device for a GPU and an internal L1 cache management method thereof. The cluster device comprises: a plurality of stream processors connected with a crossbar switch; the crossbar switch has a plurality of input ports and a plurality of output ports, the plurality of input ports are connected with the plurality of stream processors respectively, the plurality of output ports are connected with the plurality of stream processors and an on-chip interconnection network of the GPU respectively, and the crossbar switch is used for communication between the plurality of stream processors and between the plurality of stream processors and the on-chip interconnection network; and an L1 index routing module arranged in the crossbar switch, which is used for calculating an index of a corresponding L1 cache according to an address of a memory access request sent by a stream processor received by the crossbar switch, and sending the memory access request to the stream processor containing the corresponding L1 cache through the crossbar switch. The application can realize L1 cache sharing of each stream processor in the cluster device, fully utilize L1 cache resources in the GPU, and improve the performance of the stream processor and the GPU.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • L1 cache sharing method for GPU

    CN113377688A

  • MCM-GPU last-stage cache management method and MCM-GPU

    CN116185900A