Scholarlink

Efficient energy utilization is crucial for High Performance Computing, especially for exascale supercomputers requiring tens of MW to operate. Increased energy awareness benefits both data center operators and code developers, focusing on balancing performance, energy, and power consumption. This work evaluates the impact of reducing GPU core clock frequency, running four production-grade codes on the latest NVIDIA Grace Hopper Superchip and comparing results with the previous Ampere A100 GPU. Findings indicate that a static clock speed adjustment can achieve over \(10\%\) energy savings with less than \(5\%\) degradation in time performance, offering valuable insights for the future sustainability of high-performance data centers.

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

Experience on Clock Rate Adjustment for Energy-Efficient GPU-Accelerated Real-World Codes

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Giorgio Amati,
  • Matteo Turisini,
  • Andrea Monterubbiano,
  • Mattia Paladino,
  • Elisabetta Boella,
  • Daniele Gregori,
  • Danilo Croce

摘要

Efficient energy utilization is crucial for High Performance Computing, especially for exascale supercomputers requiring tens of MW to operate. Increased energy awareness benefits both data center operators and code developers, focusing on balancing performance, energy, and power consumption. This work evaluates the impact of reducing GPU core clock frequency, running four production-grade codes on the latest NVIDIA Grace Hopper Superchip and comparing results with the previous Ampere A100 GPU. Findings indicate that a static clock speed adjustment can achieve over \(10\%\) energy savings with less than \(5\%\) degradation in time performance, offering valuable insights for the future sustainability of high-performance data centers.