Evaluating Computational Performance of Backpropagation Learning on Graphics Hardware

Although volunteer computing with a huge number of high-performance game consoles connected to the Internet is promising to achieve large-scale data mining, the programming models of such game consoles for data mining tasks are restricted. As the game consoles have high-performance graphics hardware for state-of-the-art video games, a key to exploit their computation power for data mining is how effectively the data mining is mapped to the hardware as graphics processes. In this paper, therefore, a popular data mining tool called the backpropagation learning neural network is implemented as an application running on graphics hardware. Since the recent graphics hardware has many vector processing units and high memory bandwidth, it is promising to accelerate the backpropagation learning task involving a lot of data-parallel computations. The evaluation results have demonstrated the great potential of our prototype implementation for massive backpropagation learning tasks. The graphics hardware can efficiently work especially if the task is implemented so as to use data-parallel instructions supported by the hardware.