Most of the conventional parallel game tree search methods try to prevent increased total computation in environments with dezens of processors, resulting in low parallelism, which cannot utilize a large number of processors effectively. In this paper, we propose a method to realize large speed-up in environments with hundreds of processors by executing speculative tasks, which may revealed to be unnecessary afterwords. These speculative tasks are controlled so as not to disturb mandatory tasks. Evaluation through implementing the proposed method shows high speed-up with appropriate granularity and priority settings. It also shows better performance than program with conventional method.