We have developed a GPU optimized fast 3D MRI simulator for pulse sequence developments. We compared simulation and experimental results for gradient echo images of water phantoms that contains an air-filled cylinder and air-filled sphere. The agreements between simulation and experiments were good if the calculation matrix was more than two times that of original images. Because the processing speed of our simulator varied from 2.2 to 3.1 TFLOPS, we concluded that our simulator is useful for development of MRI pulse sequences.