A Feedback, Runtime Technique for Scaling the Frequency in GPU Architectures

This paper presents a dynamic frequency scaling (DFS) technique, PIDDFS, targeting on real-time applications running on GPU platforms. PIDDFS technique applies a feedback controlling algorithm, Proportional-Integral-Derivative (PID), to scale the frequencies of core domain and DRAM domains based on memory access statistics. The major goal of PIDDFS is minimizing the energy consumption while the memory traffic is intensive or even causes the pipeline to stall. Performance can also be improved via increasing the frequency while the memory traffic is in a starving status. Based on the feedback, closed-loop controlling model with proper leading and lagging phases, PIDDFS can respond timely towards the variations during runtime and ignore the insignificant noise that causes unnecessary frequency adjustments. The proposed technique has been simulated on GPGPU-Sim, a cycle-level simulator of GPU architecture and power savings have been modeled by GPUWattch. According to the benchmark simulation result, a power saving of more than 23% with a performance improvement of 4% is achieved at the same time. Stalled cycles caused by saturation of memory request queues are reduced over 40%.