An approximation for fast digital hologram generation is implemented on a central processing unit (CPU), a graphics processing unit (GPU), and a multi-GPU computational platform. The computational performance of the method on each platform is measured and compared. The computational speed on the GPU platform is much faster than on a CPU, and the algorithm could be further accelerated on a multi-GPU platform. In addition, the accuracy of the algorithm for single- and double-precision arithmetic is evaluated. The quality of the reconstruction from the algorithm using single-precision arithmetic is comparable with the quality from the double-precision arithmetic, and thus the implementation using single-precision arithmetic on a multi-GPU platform can be used for holographic video displays. © 2009 Optical Society of America.