GPU workload prediction and management

摘要

The disclosed embodiments provide a system that configures a graphics-processing unit (GPU) in a computer system. During operation, the system predicts an incoming workload to the GPU. Next, the system identifies an operational floor for the GPU based on the incoming workload. Finally, the system uses the operational floor to configure the subsequent execution of the GPU, wherein the operational floor facilitates processing of the incoming workload by the GPU.

主权项

1. A method for configuring a graphics-processing unit (GPU) in a computer system, comprising:
predicting an incoming workload to the GPU; identifying an operational floor for the GPU based on the incoming workload; and using the operational floor to configure the subsequent execution of the GPU, wherein the operational floor facilitates processing of the incoming workload by the GPU, wherein using the operational floor comprises:
obtaining an operating level of the GPU; andmodifying the operating level of the GPU based on at least one of the operational floor, an existing workload of the GPU, a bottleneck in the GPU, a minimum configuration of the GPU, an operational ceiling for the GPU, and a thermal limit for the GPU, wherein modifying the operating level includes preparing the GPU to transition to a new operating level associated with an incoming workload by initially adjusting slow-changing hardware-specific attributes based on the new operating level, but leaving fast-changing hardware-specific attributes at a previous level until the incoming workload arrives at the GPU.