Abstract: A GPU performs dynamic power level management by switching between pre-defined power levels having distinct clock and voltage levels. The dynamic power level management includes identifying a first performance metric associated with processing workloads at the for a consecutive number of measurement cycles. In some embodiments, the consecutive number of measurement cycles includes a current measurement cycle and at least one previous measurement cycle. Based on a determination that the consecutive number of measurement cycles exceeds a minimum hysteresis number, an estimated optimization is determined to be applied to the GPU for a future measurement cycle. A power level setting at the GPU for the future measurement cycle is adjusted based on the estimated optimization. By considering performance metrics including, for example, different processing workloads and hardware configurations, the GPU is able to dynamically adapt its power settings to the particular workload that it is currently processing.

Abstract: A GPU performs dynamic power level management by switching between pre-defined power levels having distinct clock and voltage levels. The dynamic power level management includes identifying a first performance metric associated with processing workloads at the for a consecutive number of measurement cycles. In some embodiments, the consecutive number of measurement cycles includes a current measurement cycle and at least one previous measurement cycle. Based on a determination that the consecutive number of measurement cycles exceeds a minimum hysteresis number, an estimated optimization is determined to be applied to the GPU for a future measurement cycle. A power level setting at the GPU for the future measurement cycle is adjusted based on the estimated optimization. By considering performance metrics including, for example, different processing workloads and hardware configurations, the GPU is able to dynamically adapt its power settings to the particular workload that it is currently processing.