Scholarlink

We present our experience porting the Fock operator application, which is a key operation of highly accurate computational methods in Electronic Structure Calculations, implemented in the BigDFT code, to SYCL, to enable the code for computer systems accelerated by Intel GPUs. We investigate the performance of the new SYCL code on Intel CPUs and GPUs and compare the results to the existing OpenMP CPU implementation and the CUDA implementation on Nvidia hardware. We show that the SYCL implementation is highly competitive, outperforming the alternatives. We additionally investigate the intra- and inter-node strong scaling properties of the SYCL code to multiple Intel GPUs with a focus on the communication overhead imposed by the SYCL implementation and derive readiness for larger computer 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

Introducing SYCL to Accelerate a Fock Operator Calculation Library of the BigDFT Electronic Structure Code

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Christoph Bauinger,
  • Luigi Genovese

摘要

We present our experience porting the Fock operator application, which is a key operation of highly accurate computational methods in Electronic Structure Calculations, implemented in the BigDFT code, to SYCL, to enable the code for computer systems accelerated by Intel GPUs. We investigate the performance of the new SYCL code on Intel CPUs and GPUs and compare the results to the existing OpenMP CPU implementation and the CUDA implementation on Nvidia hardware. We show that the SYCL implementation is highly competitive, outperforming the alternatives. We additionally investigate the intra- and inter-node strong scaling properties of the SYCL code to multiple Intel GPUs with a focus on the communication overhead imposed by the SYCL implementation and derive readiness for larger computer systems.