Figure 5From: Sapporo2: a versatile direct N-body library Multi-GPU speed-up over using one GPU. For each configuration the total wall-clock time is used to compute the speed-up (y-axis) for a given N (x-axis). The wall-clock time includes the time required for the reduction steps and data transfers. Timings performed on K20m GPUs using Sapporo2 and CUDA 5.5.Back to article page