Scholarlink

Multi-GPU servers often exhibit uneven characteristics. For instance, the data transfer bandwidth between four NVIDIA V100 GPUs can vary due to the NVLink connecting these devices to a specific CPU in servers with IBM POWER 9 processors, which means that the communication bandwidth between other devices is comparably slower. To address this issue, the Multi-GPU Generalized Matrix Multiplication (GEMM) algorithm has been adapted for platforms with uneven data transfer bandwidths. The performance profile of these adaptations was analyzed and methods for optimizing performance were introduced. In addition, a model for selecting optimal parameters was developed to enhance the efficiency of such systems.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

GEMM Algorithm for Multi-GPU Platforms with Regular Uneven Data Transfer Links

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yea Rem Choi,
  • Sergey Malkovsky,
  • Vladimir Stegailov

摘要

Multi-GPU servers often exhibit uneven characteristics. For instance, the data transfer bandwidth between four NVIDIA V100 GPUs can vary due to the NVLink connecting these devices to a specific CPU in servers with IBM POWER 9 processors, which means that the communication bandwidth between other devices is comparably slower. To address this issue, the Multi-GPU Generalized Matrix Multiplication (GEMM) algorithm has been adapted for platforms with uneven data transfer bandwidths. The performance profile of these adaptations was analyzed and methods for optimizing performance were introduced. In addition, a model for selecting optimal parameters was developed to enhance the efficiency of such systems.