Efficient random number generation for fermi class gpus
Nirodha Abeywardana · Human Biology · 2012
Abstract EFFICIENT RANDOM NUMBER GENERATION FOR FERMI CLASS GPUs by NIRODHA ABEYWARDANA JAN 2012 Advisor: Dr. Loren Schwiebert Major: Computer Science Degree: Master of Science High quality pseudorandom number generators are very important in computational science applications such as Monte Carlo simulations in order to achieve quality results. As large-scale Monte Carlo computation consumes large amounts of computational power, there has been much research on modern Graphics Processing Units to improve efficiency. Parallel portions of computationally intensive algorithms can be programmed on GPUs using Compute Unified Device Architecture (CUDA) on NVIDIA GPUs. Applicability of existing random number generators on Monte Carlo simulations that runs on the GPU is limited as the transfer of generated random numbers from CPU to GPU is costly. We propose E-MTGP, which is Mersenne Twister based random number generator that runs faster on the GPU.We evaluate the performance of E-MTGP and show how we can use this for both type of applications that run on GPU and CPU.