A. Shitara, Y. Nishikawa, M. Yoshimi, and H. Amano (Japan)
CUDA, GPU, GPGPU, Self-organizing Map
In this paper, we introduce an implementation of algorithm for self-organizing maps(SOM) using GPUs and discuss its evaluation. We used CUDA provided by NVIDIA Corporation for parallel programming, profiling, and data flow optimization so as to exploit inherent data-level parallelism of the algorithm. By using three NVIDIA’s graphic cards for evaluation, we investigated the relationships among the number of processor elements, amount of memory device and performance. As the result of performance evaluation with various parameter combinations, we found that implementation on GTX280 achieved 150 times higher performance of Intel Core 2 Quad 2.40 GHz when parameters of map size, dimension of vectors and learning size were 1372×1372, 128 and 128, respectively.
Important Links:
Go Back
