Fast and accurate thermal modeling and simulation of manycore processors and workloads

Today's microprocessors require careful analysis of their thermal behavior both at design time as well as at runtime. However, accurate prediction of thermal behavior is possible only through simulations due to the complexity and the high dynamicity of their operation. Unfortunately, accurate simulations of such complex systems are computationally intensive, therefore time consuming, while on the other hand, simplified models often cause mispredictions leading to overdesign and lowered performance. In our work, we propose a new discrete-time method of assessment of the run-time temperature of a processor based on an approximation of its instructions-per-cycle (IPC) by a finite Fourier series expansion. Our method is AC-based, and shows significant increase of accuracy in comparison to well-known DC (average)-based models. Furthermore, we provide a new Dynamic Voltage and Frequency Scaling (DVFS) model based on our estimations.