Humanoid robots require comprehensive dependability under the restrictions of power and size. The two constraints complicate the development of dependable humanoid robots. This study addresses this issue by considering the fault tolerance of the computer unit of humanoid robots. In the future, various tasks ranging from daily chores to safety-related tasks will be carried out by individual humanoid robots. If the importance of the tasks is different, the required dependability will also vary accordingly. Therefore, in humanoid robots, fault tolerance that dynamically changes based on the importance of the tasks is desirable because fault-tolerant designs involving hardware redundancy are power intensive. This paper presents a task-based dynamic fault tolerance scheme and a duplex computer system for realizing the scheme. The duplex system is characterized by performing safety failover when the standby computer unit is in a cold standby state. This paper describes the hardware implementation of the safety failover subsystem.