Figure 6

Results of block time-step simulations with a sixth order hermite code. The figure presents four sub-plots. For each of the plot the x-axis indicates the number of particles used. The four lines indicate the number of GPUs and the accuracy that is used. The solid blue line with square symbols uses double-single (DS) precision using a single GPU. The dashed yellow line with square symbols uses DS and four GPUs. The solid red line with round symbols uses full double precision (DP) using a single GPU. The dashed green line with round symbols uses DP and four GPUs. The top left sub-plot presents the absolute execution time of the simulations. The top right plot shows the speed-up when using four GPUs instead of one GPU. The bottom left plot indicates the average number of particles that were being integrated per time-step (\(N_{\text{active}}\)). The bottom right plot marks the energy error at the end of the simulation. NOTE that the results for one and four GPUs when using DP are the same for the bottom two plots. Timings performed on K20m GPUs using CUDA 5.5.