GPU High-Performance Computing Applied Research And Experimental In Computer GO

As game project, computer GO is the product of human brain's higher cognitive functions. Since 2006, the research of computer GO developed with leaps and bounds. And the core is reflected in Monte Carlo tree search. According to the characteristics of the Monte Carlo method, more simulation times in unit time, the faster convergence. In Monte Carlo tree search, the most obvious and necessary aspect in parallel processing is the process of each simulation. Although the computer power of CPU is not simple weak, but the status of 8 to 16 core of laboratory server did not lead the convergence to a good position. Inherent, the GPU is a many-core parallel processer and its processing units are far more than the number of CPU. With the introduction of CUDA, the door of making use of general-purpose GPU computing has been completely opened. So developers can tap the GPU computing power in a friendly environment. This also allows us to make use of the power of GPU computing in computer GO as possible.This paper proposed and implemented a computer GO game engine based on CUDA. This engine make use of the many-core parallel processors of GPU to handle the process of Monte Carlo simulations of tree search operations to speed up the convergence rate of Monte Carlo in order to improving the thinking depth of our computer GO system. Finally, this paper compares the experimental data between the original CPU multi-threaded engine and the CUDA engine. The result shows the effectiveness of this method.