engleski

Performance comparison for Nvidia CUDA and Intel Xeon Phi

NVIDIA graphic processor unit (GPU) using Compute Unified Device Architecture (CUDA) and Intel Xeon Phi are two different parallel computing platforms. The challenge using the platforms lies in achieving the best possible performance. The real value in GPU and Xeon Phi coprocessor programming lies in the ability to run large number of parallel threads. To provide fair comparison of the processors, OpenMP technique, hand-tuned CUDA kernels and shared memory have been used. For the Xeon Phi, the application has been executed using offload mode. Memory capacity represents the main limitation for current generation of Intel Xeon Phi coprocessors, which is also a limitation for GPU. For that reason, offload programming is the most popular mode as the memory capacity of the host can be used to prevail that limitation. In this paper, the performance of NVIDIA CUDA and Intel Xeon Phi using matrix multiplication code and K-means algorithm has been compared. Performance tests measure and compares kernel execution times and end-to-end application execution times for both CUDA and Xeon Phi.

coprocessor, CUDA, GPU, Intel Xeon Phi, K-means, matrix multiplication, multicore processing, parallel programming

nije evidentirano