Timing results for the hybrid algorithm (one-half NVIDIA K80+CPU) illustrated in Fig 5.

The wall time (W), the time required to assemble the system (A), the time required for the linear solves (L) and the overhead due to offloading to the GPU (O) are shown. Note that for the GPU implementation the time required by the linear solve (which is done on the CPU) always dominates the total runtime. Thus, we have W = L + O. The number of slices that yield the optimal run time are shown in bold. All measured times are in units of seconds. In addition, the standard deviation determined from 20 repetitions of the simulation is shown next to the wall time.