Investigating Improving the Speed and Precision of Embedded Real Time Simulation

This research investigates the tradeoffs between speed and accuracy for simulation on compute constrained devices when choosing to simulate in single, double, or mixed precision. This is a concern for embedded real time simulation, where a simulation of a system is embedded into the real system to achieve new behavior, often using a microcontroller as a computing platform. To perform the investigation simulations of different precisions were run on embedded processors to determine the speed of the simulation. Accuracy was measured by comparing the simulation to the analytic solution of the system and calculating the percent error at 50 equally spaced points in the simulation. The results of this research show that single precision will perform simulations several times faster than the equivalent double precision algorithm. The double precision state vector method was also investigated as a mixed precision method, and it offered no speed or accuracy benefits over single precision. Double precision was found to be desirable in cases where extremely high accuracies are required or when the simulation needed to be run for a long time relative to the step size.