Table 6.
Runtime statistics in hours (h) for MCMC sampling the estimated minimum number of samples (no burn-in, no-thinning) of various models, using a single HCP MGH (552 volumes) and single RLS-pilot (134 volumes) dataset.
| Model | Number of samples | HCP MGH - GPU | HCP MGH - CPU | RLS-pilot - GPU | RLS-pilot - CPU |
|---|---|---|---|---|---|
| Ball&Stick_in1 | 11,000 | 1 h | 36 h | 0.25 h | 4.5 h |
| Ball&Stick_in2 | 15,000 | 2 h | 48 h | 0.5 h | 15 h |
| Ball&Stick_in3 | 25,000 | 6 h | 72 h | 1.5 h | 45 h |
| NODDI | 15,000 | 4 h | 80 h | 1 h | 20 h |
| Tensor | 13,000 | 1 h | 22 h | 0.5 h | 4 h |
| CHARMED_in1 | 17,000 | 6 h | 54 h | n/a | n/a |
| CHARMED_in2 | 25,000 | 24 h | 139 h | n/a | n/a |
| CHARMED_in3 | 25,000 | 47 h | 200 h | n/a | n/a |
Statistics are for a whole brain mask of 410,000 voxels for HCP MGH and 204,993 voxels for RLS-pilot. Results are computed using an AMD Fury X GPU and an Intel Xeon e3 18 core CPU.