high performance computing on graphics processing units: hgpu.org

In this research implementation of a dynamic programming algorithm (Viterbi) has been done on graphics processing unit of NVidia using CUDA model. As graphical processing units are becoming important in supporting central processing units for the acceleration of complex floating point calculations. The complex computation goes on parallel in graphics processing unit as it contains hundreds of processing units as compared to central processing units which have only extended up to tens of cores yet. Because of this capacity of graphics processing units complex computations can be accelerated up to 10-100 times and even more depending upon the program. And more importantly data transfer between central processing unit and graphics processing unit do not create a bottleneck because latest PCI bus can transfer data up to 8 GB/s. Recently two main graphics processing unit’s architectures are becoming popular one is the CUDA (Compute Unified Device Architecture) and the other one is ATI Stream. CUDA is modification of C language with some extensions and limitations. Here a dynamic programming algorithm named "viterbi" has been implemented in CUDA using graphics processing unit and concluded that it has been accelerated from 3 to 6 times as compared to the serial execution on central processing unit. This research can help in future implementation of dynamic programming algorithms using matrix-matrix multiplication approach instead of matrix-vector multiplication approach which can speed up the processing of this kind of algorithms. This research has explored the use of the graphics processing unit in Viterbi algorithm which is used for sequence alignment, keyword spotting system etc.