Student Session: Data distribution on a multi-GPU node for TomoBayes CT reconstruction

Computed tomography (CT) is an imaging technique that uses iterative algorithms to reconstruct the interior of large volumes. Graphics Processing Units (GPUs) are currently the preferred technology for computation acceleration. However, the collected data storage can exceed the internal memory of current GPUs and therefore requires costly CPU GPU data transfers. In this paper, we present a strategy of data distribution over several GPUs avoiding this bottleneck. We provide experimental results showing that our memory-saving method accelerates the iterative reconstruction. We achieve a better parallelisation efficiency using 8 GPUs to reconstruct a volume of size 4 GB than reconstructions based on centralised data storage on CPU.