A fine granularity parallel algorithm for electromechanical transient stability simulation is proposed based on the graphic processing unit（GPU） and implicit gradient method. The whole power system is divided into the generator system and non-generator system. When solving the generator system buses， first each subsystem composed of connected generator buses is allocated an independent GPU block and uses LU decomposition to solve the linear functions. Then the non-generator system is simulated on another GPU block， supported by LU messages from the generator system. Numerical results show that the accuracy of the parallel algorithm proposed is roughly the same as that of the CPU serial method and BPA software. Moreover， with an increase of the power system scale， the speedup effect is more obvious.