Mobile devices require long battery lifetime while still delivering high performance. A number of research efforts have been devoted to reduce energy consumption of a processor without impacting the performance through the use of dynamic voltage scaling (DVS). Previous DVS algorithms usually rely on workload history. But it may not be efficient when the workload variability is high as in MPEG decoding, which will become one of the most important applications in mobile computing. This thesis presents two DVS algorithms on MPEG decoding. One is DVS with delay and drop rate minimizing algorithm (DVS-DM) where voltage is determined by previous workload. Another algorithm predicts MPEG decoding time using the sizes of next frames, and scales voltage according to the predicted MPEG decoding time and previous workload (DVS with predicted decoding time or DVS-PD). We compare four MPEG decoders: the first one is original MPEG decoder from Boston University, the second one is MPEG decoder with shutdown mechanism, the third implements DVS-DM algorithm, and the last one runs DVS-PD. Simulation results show that DVS-PD saves more energy than other algorithms due to more accurate prediction of future workload. We also found that the amount of energy saving with DVS-PD is related with error rate of the predictor, which implies that if decoding time is predicted more accurately, DVS algorithm can be more efficient.