Out-of-core cone beam reconstruction using multiple GPUs

This paper presents a graphics processing unit (GPU) based method capable of accelerating cone-beam reconstruction of large volume data, which cannot be entirely stored in video memory. Our method accelerates the Feldkamp, Davis and Kress (FDK) algorithm in a multi-GPU environment. We present how the entire volume can be efficiently decomposed into small portions to reduce the usage of video memory on each graphics card. Experimental results are also presented to understand the reconstruction throughput on an nVIDIA Tesla S1070 server. It takes approximately three minutes to reconstruct a 2048³-voxel volume from 720 2048²-pixel projections. The effective bandwidth of video memory reaches 137 GB/s per GPU, demonstrating a higher utilization of texture caches.