The iterative reconstruction algorithms for X-ray CT image reconstruction suffer from their high computational cost. Recently Nvidia releases common unified device architecture (CUDA), allowing developers to access to the processing power of Nvidia graphical processing units (GPUs), in order to perform
general purpose computations. The use of the GPU, as an alternative computation platform, allows decreasing processing times, for parallel algorithms. This paper aims to demonstrate the feasibility of such an implementation for the iterative image reconstruction. The ordered subsets convex (OSC) algorithm, an iterative reconstruction algorithm for transmission tomography, has been developed with CUDA. The performances have been evaluated and compared with another implementation using a single CPU node. The result shows that speed-ups of two orders of magnitude, with a negligible impact on image accuracy, have been observed.