T-FDK algorithm is an FDK-type cone beam CT reconstruction algorithm. Like other 3D reconstruction algorithms, T-FDK is time consuming because of the large amount of data processing involved. One solution to this problem is utilizing PC graphics boards (GPU) for acceleration. The recent dramatic evolution of GPU makes this method come to the practical track. In this paper, we use a new floating point GPU to speedup the 3D T-FDK algorithm that is different from original FDK method in structure. Because floating point pipelines are slower than hardwired 8-bit texture mapping facilities but are more precise numerically, we balance the reconstruction speed and quality by using both of them. Using nVIDIA GeForce 6800 GT, our GPU accelerated T-FDK method gives a speed 27.612 times faster than a software implementation.