Formalin: Architectural Support for Power & Performance Aware GPU

Muhammad Husni Santriaji, Henry Hoffmann

Abstract: This paper presents a control strategy for managing the performance and power consumption of a graphics processing unit (GPU). GPUs are an attractive target for control because they have ample sources of feedback and support multiple actuators. The key challenge is that GPUs are increasingly tasked with running general purpose computations and it is extremely difficult to build a single model that captures all possible workloads' responses to available actuators. To overcome this challenge we propose a hierarchical control structure. During GPU application execution, a global controller sets performance goals for individual application components (or kernels). A local controller then ensures that those goals are met for the kernel while respecting the power budget. The local controller relies on an integral control strategy that can explore actuator response online. We implement this strategy and observe that it meets performance and power goals with low error while saving energy compared to commonly used heuristic strategies.