Figure 4.

Estimated combined computation times for reading the fMRI data, computing the connectivity graph and writing the thresholded (and thus sparse) connectivity matrix to compressed (left) and uncompressed (right) RData files are depicted for different numbers of subjects on a single GPU. Since computation time depends linearly on the number of subjects, computation time for larger numbers of subjects are estimated using average per-subject computing times measured from 200 subjects. By employing multiple GPUs the runtime can be reduced linearly; for example, using four GPUs instead of one for computing 500 subject's connectivity graph would reduce the computation time from about 36 days down to 9.