Scalable Image Registration and 3D Reconstruction at Microscopic Resolution - High-Throughput Image Reconstruction and Analysis - page 203

Figure 8.14 Single node performance under different node configurations for the most work-
intensive image from the mammary set.
The effects of adding a second GPU to the single node configuration are pre-
sented in Figure 8.13. Gains from GPU parallelism are very diverse, starting with
30% to 50% for small window sizes, continuing with 60% for large window sizes,
and finally ending with near-optimal scalability for medium sizes. These gains are
proportional to computational workload, revealing that the GPU is a more scalable
processor when it can exploit arithmetic intensity and GFLOPS are not limited by
data shortages due to insufficient bandwidth between video memory and GPU.
Multiple Node Performance
Scalability results for CPU are shown in Figure 8.15. Gains are fairly consistent
for all images and all parameter sets, since although CPU executions are slower
Figure 8.15 (a) CPU scalability for multiple nodes with the placenta set. (b) CPU scalability for the
mammary set.