| With the development of IC technology,the Dennard scaling law begins to break down,the power density of chips starts to increase as IC technology advances,with power density increases beyond the power wall,some cores of multi-core chips have to be turned off to ensure the thermal safety of chips,resulting in the so called dark silicon problem.For IC chips with dark silicon problem,we need to decide the number and distribution of cores that should be turned on,and calculate reasonable power budget of these cores for design of run time use.For these needs,this article makes a research on the technology of power budgeting,designs power budget methods for both steady and transient state and verify them through experiments.First,for latter design of algorithm and experiments,we build thermal models of multi-core chips and extract models from HotSpot.Then,for providing reliable information for power budgeting,this article creatively designs a fast coupling leakage-aware transient thermal analysis method,which is one of the highlights and core algorithm of this research.We linearize the model of leakage power,as well as the whole thermal model based on Taylor expansion,reduce the original model using a sampling based model order reduction method,and specially design a update technology of projection matrix based on the incremental singular value decomposition.This method achieves tens to hundreds of times speedup against traditional method and serves as a solid basis of power budgeting with the ability of quickly estimating leakage power and temperature,Finally,we propose power budgeting methods of both steady and transient state for dark silicon chip.We analyze the thermal constraints and performance of dark silicon,simplify the optimization problem and design an efficient algorithm of estimating power budget and recommending number of light-core for steady state.With the help of the former fast leakage-aware coupling analysis method and the model order reduction method,we also propose a new transient power budget method for dark silicon.Through a series of experiments,we verify and analysis the efficiency and advantages of the proposed methods,which totally outperform the state-of-art. |