Performance comparison of CPU and GPU on a discrete heterogeneous architecture

Today Graphics Processing Units (GPUs) in scientific computing have led the computing system to achieve tera-scale computing power to the laptops and peta-scale computing power to the clusters by combining multicore Central Processing Units (CPUs) and many core GPUs which can be called a heterogeneous computer architecture. This paper describes briefly an evolutionary journey of GPUs. For performance comparison, parameters considered are latency and throughput. So based on the execution time of a GPU and CPU for a given task, written with Compute Unified Device Architecture (CUDA) C language, the two parameters are measured with increasing size of workload. When the task size is increased GPU is found to be approximately 51% faster than the multithreaded CPU when GPU achieves 100% occupancy. Throughput of GPU is found to be 2.1 times higher than that of CPU for large task size. The GPU used is NVIDIA's GeForce GT630M with CPU of Intel's i-5 3210M 3 rd generation processor.