Figure 4

Performance difference between fourth and sixth order kernels. Shown is the time required to solve \(N^{2}\) force computations using different configurations. The number of source particles is equal to the number of sink particles indicated on the x-axis. The y-axis indicates the required wall-clock time to execute the gravity computation and to perform the data transfers. The fourth-order configuration using double-single precision is indicated by the dotted line with open square symbols. The fourth order configuration using double precision is indicated by the dotted line with filled square symbols. The sixth order configuration using double-single precision is indicated by the solid line with open circles and the sixth order with double precision is indicated by the solid line with filled circles. The straight solid line without symbols indicates the \(N^{2}\) scaling. Timings performed on a K20m GPU using CUDA 5.5.