Saturday 05 April 2025
The quest for efficiency in high-performance computing has led scientists to explore new ways to reduce energy consumption and increase performance. One approach is mixed-precision computing, which involves using different levels of precision for different parts of a calculation. This can be particularly effective in computational fluid dynamics (CFD) applications, where complex simulations are used to model real-world phenomena.
A team of researchers has developed a methodology for enabling mixed-precision computing in CFD codes, and tested it on two representative case studies: Nekbone and Neko. Nekbone is an open-source code that solves the Navier-Stokes equations using a finite-element method, while Neko is a modern, portable, and scalable framework for high-fidelity CFD.
The researchers used a combination of code inspection, advanced tools, and targeted implementation to enable mixed-precision computing in both codes. They found that by using single precision for certain calculations and double precision for others, they could achieve significant performance improvements while maintaining the required accuracy.
In Nekbone, the team was able to reduce the time-to-solution by up to 40.7% and the energy consumption by 47.1%. In Neko, they achieved a 30% reduction in execution time and a 24% reduction in energy consumption. These improvements were achieved without sacrificing accuracy, which is critical in CFD applications where small errors can have significant consequences.
The researchers also developed a Gather-Scatter library that allows for efficient communication between processors in parallel computations. This library was used to implement the mixed-precision strategy and helped to achieve the observed performance gains.
The team’s work has important implications for the field of high-performance computing, particularly in the context of exascale computing. As simulations become increasingly complex and demanding, the need for efficient and scalable algorithms will only continue to grow. By exploring new approaches like mixed-precision computing, researchers can help to reduce energy consumption and increase performance, ultimately enabling more accurate and detailed simulations.
The use of mixed-precision computing is not limited to CFD applications. The same approach could be applied to other fields where complex calculations are required, such as weather forecasting or materials science. As the demand for high-performance computing continues to grow, it is likely that we will see increasingly sophisticated techniques being developed to achieve better performance and efficiency.
The development of mixed-precision computing is an important step towards making high-performance computing more sustainable and efficient.
Cite this article: “Unleashing the Power of Mixed-Precision Computing: A Game-Changer for High-Performance Computational Fluid Dynamics”, The Science Archive, 2025.
High-Performance Computing, Mixed-Precision Computing, Computational Fluid Dynamics, Cfd Codes, Navier-Stokes Equations, Finite-Element Method, Energy Consumption, Accuracy, Gather-Scatter Library, Exascale Computing.
