The most computationally intensive part of molecular dynamics simulations is the calculation of the forces between neighboring particles. Therefore, it is necessary to optimize these as much as possible to achieve acceptable runtimes, both through algorithmic and hardware-specific means. An example that fits both categories is parallelizing the workload to a high degree. On a low level, this means utilizing vector instructions to fully exploit the processing power of modern CPUs.

As one approach to this, in this thesis we implement modes for reduced precision into the molecular dynamics library AutoPas which currently uses double precision exclusively. Through these contributions, certain elements or even all parts of the force calculations are executed in single precision in order to increase parallelization and speed them up. The first part of the work explains the necessary changes to achieve this. It is followed by a demonstration that the adjustments do not impact the desired accuracy of the results. Lastly, we present in our findings that mixed precision did not yield a convincing performance gain, yet single precision delivered a significant uplift up to 1.77 as well as savings in power over the previous double precision implementation.