The energy consumption and the compute performance of a fluid dynamic code have been investigated varying parallelization approach, arithmetic precision and clock speed. The code is based on a Lattice Boltzmann approximation, is written in Fortran and was executed on high-end GPUs of Leonardo Booster supercomputer. Tests were conducted on single server nodes (up to 4 GPUs in parallel). Performance metrics like the number of operations per second and energy consumption are reported, to quantify how smart coding approach and system adjustment can contribute to reduction of energy footprint while keeping the scientific throughput almost unaltered or with acceptable level of degradation. Results indicate that this application can be executed with 20% of energy saving and reduced thermal stress, at the cost of 5% more computing time. The paper presents preliminary conclusions, as it is a first step of a larger study dedicated to energy efficiency at scale.

翻译：本文研究了流体动力学代码在不同并行化方法、算术精度和时钟频率下的能耗与计算性能。该代码基于Lattice Boltzmann近似方法，采用Fortran语言编写，并在Leonardo Booster超级计算机的高端GPU上运行。测试在单服务器节点（最多并行使用4个GPU）上进行。通过报告每秒操作数和能耗等性能指标，量化了智能编码方法和系统调整如何在保持科学计算吞吐量基本不变或处于可接受降级水平的前提下，助力减少能源足迹。结果表明，该应用可在计算时间增加5%的代价下实现20%的节能并降低热应力。本文提出了初步结论，此为大规模能效系统性研究的首阶段工作。