This paper presents performance and quality analysis for graphics processing unit (GPU)-accelerated, cone beam-computed tomography (CBCT) reconstruction using FDK filtered back projection. The paper also compares the current generation (2006) Barco MXRT-7100 GPU and a next generation (2007) GPU prototype. The results presented illustrate that while GPUs have immense computational power, there is opportunity to tune an implementation to specific hardware and quality requirements, and even to the image acquisition geometry.