Mixed Precision Arithmetic Accelerates Turbulent Flow Simulations on Modern Hardware.

Mixed precision arithmetic accelerates compressible turbulent flow simulations. Customisable precision levels within the OPS and OpenSBLI frameworks yield performance gains without compromising accuracy, notably with half-single and single-double configurations. Pure half-precision proved inaccurate, while reduced precision lessened memory usage and communication overhead.

Computational fluid dynamics routinely demands substantial computational resources, particularly when modelling turbulent flows. Researchers are actively investigating methods to alleviate this burden without sacrificing fidelity. A team led by Bálint Siklósia (Pázmány Péter Catholic University) and including Pushpender K. Sharma, David J. Lusher (Japan Aerospace Exploration Agency – JAXA), István Z. Regulya, and Neil D. Sandham (all University of Southampton) detail their investigation into reduced and mixed precision arithmetic for simulating compressible turbulent flow. Their work, entitled ‘Reduced and mixed precision turbulent flow simulations using explicit finite difference schemes’, extends existing computational frameworks – OPS and OpenSBLI – to allow for configurable precision levels, demonstrating performance gains and reduced memory requirements through careful selection of numerical precision. The team validates their approach using the Taylor-vortex benchmark, highlighting the trade-offs between computational efficiency and maintaining acceptable numerical accuracy.

Precision Engineering: Optimising Computational Fluid Dynamics with Customisable Arithmetic

Computational fluid dynamics (CFD) routinely demands significant computational resources, particularly when modelling turbulent flows. Researchers continually seek methods to reduce this burden without compromising the accuracy of simulations. A team led by Bálint Siklósia (Pázmány Péter Catholic University) has developed a novel approach centred on customisable numerical precision, extending the capabilities of established frameworks, OPS and OpenSBLI.

The study investigates the application of reduced and mixed precision arithmetic to enhance the efficiency of high-performance computing (HPC) simulations of compressible turbulent flow. Numerical precision refers to the number of significant digits used to represent numbers in a computer; higher precision yields more accurate results but demands greater computational resources. Reduced precision uses fewer digits, accelerating calculations but potentially introducing errors. Mixed precision strategically combines different levels of precision for different parts of a simulation.

The researchers extended the OPS and OpenSBLI frameworks – software tools used to develop and optimise CFD codes – to allow for customisable precision levels. This enables researchers to fine-tune performance without necessarily sacrificing accuracy.

Numerical experiments, utilising the established Taylor-Green vortex benchmark – a standard test case for evaluating CFD codes – revealed a clear trade-off between precision and performance. A complete transition to half-precision (using 16 bits to represent a number) introduced unacceptable levels of error. However, mixed-precision configurations – combining half, single (32-bit), and double (64-bit) precision – successfully balanced computational speed with solution fidelity.

These configurations achieve performance gains by reducing both memory usage and communication overhead. Reduced precision requires less memory to store data, and less data needs to be exchanged between processors in parallel simulations, particularly benefiting multi-CPU and multi-GPU systems. Modern hardware architectures are optimised for operations on lower-precision data types, further accelerating calculations.

The observed improvements stem from several factors. Reduced memory bandwidth requirements and faster operations on modern hardware contribute to the gains. The use of mixed precision also allows for a reduction in communication overhead, as less data needs to be exchanged between processors in parallel simulations.

This work contributes to the growing body of knowledge on the effective utilisation of reduced and mixed precision arithmetic in scientific computing. The ability to fine-tune precision levels based on the specific requirements of each operation represents a significant advancement, offering a powerful tool for optimising performance without compromising accuracy. Future work should focus on developing adaptive precision strategies that dynamically adjust precision levels based on error estimates and computational demands.

The findings highlight the importance of careful precision selection in numerical simulations. Researchers achieve a balance between computational efficiency and the maintenance of accurate results by strategically allocating precision levels, enabling them to tackle increasingly complex scientific problems. Further investigation into the scalability of these mixed-precision techniques on larger, more complex turbulent flow problems is also warranted.