Mixed-Precision Efficiency Improvements for Multi-Physics Flow Simulations

Accelerated architectures are becoming increasingly prevalent in contemporary high performance computing architectures, and are planned for inclusion in exascale extreme scale) architectures in the coming years. Some of the salient characteristics of these architectures include limited memory capacities and enhanced capacity to perform arithmetic at reduced precision. The proposed work aims to build a theoretical framework to help quantify the impact of finite precision errors for high fidelity, multi-physics flow simulations. In addition, efficient mixed precision implementations of low-dissipation numerical discretization of the flow equations will be performed to assess the trade-offs between computational throughput/memory and solution accuracy. This will be done using several canonical validation benchmarks of turbine blade heat transfer, high speed aeroacoustics, and lean, premixed flames. Increasing throughput or reducing memory on accelerated architectures can help reduce the computational cost in dollars) associated with high fidelity simulations, which is presently one of primary obstacles to widespread adoption of such simulation techniques.