Molecular dynamics simulations are a central tool in computational science but are increasingly constrained by performance and energy consumption on modern high-performance computing systems. Mixed-precision arithmetic has emerged as a promising technique to improve computational efficiency by selectively reducing numerical precision, yet its applicability to long-running molecular dynamics simulations remains challenging due to strict accuracy and stability requirements.This thesis investigates the integration and systematic evaluation of mixed- and reduced-precision computation within the AutoPas molecular dynamics framework and its reference application md-flexible. Mixed-precision support is reintroduced into the current AutoPas architecture, resolving incompatibilities introduced by architectural evolution and enabling consistent execution in double precision, mixed precision, and single precision modes. A comprehensive experimental study evaluates the impact of numerical precision on macroscopic accuracy, performance, energy consumption, and auto-tuning behavior. Accuracy is assessed using potential energy and virial for both small diagnostic systems and large-scale simulations consisting of approximately 512,000 particles. Performance and energy efficiency are analyzed under controlled conditions across different workload regimes. The results show that mixed- and reduced-precision configurations closely reproduce the macroscopic behavior of double-precision simulations, with no observable error accumulation as system size increases. Performance benefits from reduced precision emerge for large, compute-dominated workloads, while small systems are dominated by precision-related overheads. Energy efficiency is found to be strongly platform-dependent and does not directly follow runtime performance. AutoPas’ algorithmic tuning behavior remains largely stable across precision modes. These findings demonstrate that mixed-precision techniques can be applied to molecular dynamics simulations within a flexible framework without compromising numerical correctness, provided that workload characteristics and hardware properties are carefully considered.     «

Molecular dynamics simulations are a central tool in computational science but are increasingly constrained by performance and energy consumption on modern high-performance computing systems. Mixed-precision arithmetic has emerged as a promising technique to improve computational efficiency by selectively reducing numerical precision, yet its applicability to long-running molecular dynamics simulations remains challenging due to strict accuracy and stability requirements.This thesis investigates...     »