high performance computing on graphics processing units: hgpu.org

Parameter optimization is a common task in various fields such as computational biology. In these scientific fields, optimization can be, e.g. based on ordinary differential equations with the computational task getting increasingly computation-intensive for increasing complexity of ODE and the parameters to determine. Hence, this raises requirements for an efficient treatment on high-performance computing architectures. In HPC, it is essential, besides an efficient implementation, to choose the best fitting architecture for each problem. Latest research has shown that FPGAs are a better accelerating device than GPUs or multi-core CPUs for specific issues. Therefore, this thesis deals with the implementation and assessment of ODE solvers optimized for FPGAs. Since FPGAs are relatively new in HPC, the thesis first explains the essential components of FPGAs and how to program them. Subsequently, the thesis expounds on the FPGA implementation of an efficient ODE solver and how it integrates into an automatic code generation to support different ODE systems. Furthermore, it illustrates the effect of the different optimizations on hardware utilization and execution time. The results are finally compared to those of CPUs and GPUs. The comparison reveals that for small problem sizes, the FPGA performs better than the CPU and almost as good as the GPU. For larger problems, both other architectures outperform the FPGA. The results give a first tendency when which architecture fits best. Consequently, this thesis builds a good foundation for further research about the usability of FPGAs for ODE solvers. This thesis did not take energy efficiency into account.