The speed of image reconstruction is an important indicator of the three-dimensional Computed Tomography (CT) imaging systems. GPU has become a focus of the research in CT reconstruction due to its characteristic of parallel computing. In this paper, the acceleration of backprojection, which is the most time consuming step of reconstruction, was implemented on multi-GPU hardware and OpenCL platform. The optimization techniques of parallelization strategy and memory management were explored and studied. The experimental results showed that the method obtained a considerable speedup, and the calculation times of total backprojection using 2 GPUs that were 1.67 times faster than using a single GPU. The method may be taken as the useful reference for researchers working on GPU applications and high-performance applications.