Task calculation method and device

By establishing a data transmission path between the GPU's resource scheduling module and the cache area, and handing over low-parallelism sub-computing tasks to PRA for processing, the problem of poor computing performance of the GPU when processing low-parallelism tasks is solved, and the overall computing efficiency and UCU utilization are improved.

CN120780438AActive Publication Date: 2025-10-14INNOSILICON MICROELECTRONICS (WUHAN) CO LTD
9 Cites 0 Cited by

Patent Information

Application Number
CN202511221784.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-08-29
Publication Date
2025-10-14
Estimated Expiration
2045-08-29

AI Technical Summary

Technical Problem

In the existing technology, GPU has poor computing performance when processing computing tasks with low parallelism, resulting in low UCU utilization and a long time consumption in the task issuance and resource allocation process.

Method used

Sub-computing tasks with high parallelism are assigned to UCU for calculation, while sub-computing tasks with low parallelism are assigned to PRA itself for calculation. By establishing a data transmission path between the resource scheduling module and the cache area, the computing resources and cache area data of PRA are utilized to optimize task scheduling and dependency management.

Benefits of technology

It improves the computing performance of the overall computing task group, reduces the time spent on task dispatching, and improves the computing efficiency of the GPU, especially reducing latency in large language models and edge AI scenarios.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention relates to the technical field of chips, and provides a task calculation method and device. The method comprises the steps that a resource scheduling module executes a first sub-calculation task and issues a second sub-calculation task to a calculation module so that the calculation module can execute the second sub-calculation task; wherein the degree of parallelism of the first sub-calculation task is lower than that of the second sub-calculation task. According to the method, the sub-calculation tasks with high degree of parallelism are submitted to the UCU for calculation, the sub-calculation tasks with low degree of parallelism are submitted to the PRA for calculation, so that the sub-calculation tasks with low degree of parallelism do not need to occupy the UCU, the calculation capabilities of the UCU and the PRA are fully utilized, and the calculation performance of the whole calculation task group is improved.
Need to check novelty before this filing date? Find Prior Art