28 September 2016 Real-time electroholography using a multiple-graphics processing unit cluster system with a single spatial light modulator and the InfiniBand network
Author Affiliations +
Abstract
Parallel calculations of large-pixel-count computer-generated holograms (CGHs) are suitable for multiple-graphics processing unit (multi-GPU) cluster systems. However, it is not easy for a multi-GPU cluster system to accomplish fast CGH calculations when CGH transfers between PCs are required. In these cases, the CGH transfer between the PCs becomes a bottleneck. Usually, this problem occurs only in multi-GPU cluster systems with a single spatial light modulator. To overcome this problem, we propose a simple method using the InfiniBand network. The computational speed of the proposed method using 13 GPUs (NVIDIA GeForce GTX TITAN X) was more than 3000 times faster than that of a CPU (Intel Core i7 4770) when the number of three-dimensional (3-D) object points exceeded 20,480. In practice, we achieved ∼40 tera floating point operations per second (TFLOPS) when the number of 3-D object points exceeded 40,960. Our proposed method was able to reconstruct a real-time movie of a 3-D object comprising 95,949 points.
© 2016 Society of Photo-Optical Instrumentation Engineers (SPIE)
Hiroaki Niwase, Naoki Takada, Hiromitsu Araki, Yuki Maeda, Masato Fujiwara, Hirotaka Nakayama, Takashi Kakue, Tomoyoshi Shimobaba, Tomoyoshi Ito, "Real-time electroholography using a multiple-graphics processing unit cluster system with a single spatial light modulator and the InfiniBand network," Optical Engineering 55(9), 093108 (28 September 2016). https://doi.org/10.1117/1.OE.55.9.093108 . Submission:
JOURNAL ARTICLE
6 PAGES


SHARE
Back to Top