Computational Fluid Dynamics (CFD) has established itself as the third fundamental approach in fluid dynamics, complementing theory and experimentation. Although methods such as Direct Numerical Simulation (DNS) provide the highest level of accuracy by resolving all turbulent scales, their computational cost remains prohibitive for practical applications. In incompressible flow solvers, the solution of the Poisson equation for pressure represents the dominant computational kernel in terms of cost. At the same time, the evolution of High-Performance Computing (HPC) systems, driven by the demands of Artificial Intelligence and the introduction of GPUs (Graphics Processing Units), has created a significant performance gap between double precision (FP64) and lower precisions (FP32, FP16), with the latter offering higher throughput and better energy efficiency. This thesis aims to exploit this hardware trend to enhance the efficiency of CFD solvers through a mixed-precision strategy. The investigated method solves the Poisson equation in single precision (FP32), taking advantage of its higher computational speed and lower memory footprint, while keeping the rest of the solver in double precision (FP64) to preserve numerical robustness and overall accuracy. The work quantifies the substantial reductions in time-to-solution achievable with this algorithmic innovation, contributing to making high-fidelity turbulence simulations more sustainable and computationally efficient.
La fluidodinamica computazionale (Computational Fluid Dynamics, CFD) si è affermata come terzo approccio fondamentale nello studio della fluidodinamica, affiancando teoria e sperimentazione. Sebbene strumenti come la Direct Numerical Simulation (DNS) forniscano la massima accuratezza risolvendo tutte le scale di turbolenza, il loro costo computazionale risulta proibitivo per applicazioni pratiche. Nei risolutori per flussi incomprimibili, la risoluzione dell’equazione di Poisson per la pressione rappresenta il kernel computazionale dominante in termini di costo. Parallelamente, l’evoluzione dei sistemi High-Performance Computing (HPC), trainata dalle esigenze dell’Intelligenza Artificiale e dall’introduzione delle GPUs (Graphics Processing Units), ha creato una significativa disparità prestazionale tra precisione doppia (FP64) e precisioni inferiori (FP32, FP16), favorendo queste ultime in termini di throughput ed efficienza energetica. Questa tesi si propone di sfruttare questa tendenza hardware per ottimizzare l’efficienza dei risolutori CFD attraverso una strategia a precisione mista. Il metodo investigato risolve l’equazione di Poisson in singola precisione (FP32), sfruttandone la maggiore velocità di calcolo e ridotta occupazione di memoria, mantenendo il resto del risolutore in doppia precisione (FP64) per garantire robustezza numerica e accuratezza finale. Il lavoro quantifica i significativi guadagni in time-to-solution ottenibili con questa innovazione algoritmica, contribuendo a rendere le simulazioni di turbolenza ad alta fedeltà più sostenibili ed efficienti.
Risoluzione dell’equazione di Poisson per la pressione con precisione numerica variabile: effetti sulla risoluzione delle equazioni di Navier-Stokes
GHMID, OMAR
2024/2025
Abstract
Computational Fluid Dynamics (CFD) has established itself as the third fundamental approach in fluid dynamics, complementing theory and experimentation. Although methods such as Direct Numerical Simulation (DNS) provide the highest level of accuracy by resolving all turbulent scales, their computational cost remains prohibitive for practical applications. In incompressible flow solvers, the solution of the Poisson equation for pressure represents the dominant computational kernel in terms of cost. At the same time, the evolution of High-Performance Computing (HPC) systems, driven by the demands of Artificial Intelligence and the introduction of GPUs (Graphics Processing Units), has created a significant performance gap between double precision (FP64) and lower precisions (FP32, FP16), with the latter offering higher throughput and better energy efficiency. This thesis aims to exploit this hardware trend to enhance the efficiency of CFD solvers through a mixed-precision strategy. The investigated method solves the Poisson equation in single precision (FP32), taking advantage of its higher computational speed and lower memory footprint, while keeping the rest of the solver in double precision (FP64) to preserve numerical robustness and overall accuracy. The work quantifies the substantial reductions in time-to-solution achievable with this algorithmic innovation, contributing to making high-fidelity turbulence simulations more sustainable and computationally efficient.| File | Dimensione | Formato | |
|---|---|---|---|
|
ghmid_omar.pdf
accesso aperto
Dimensione
1.99 MB
Formato
Adobe PDF
|
1.99 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/94641