Research On Runtime Power Budget Technology Of Many-core Dark Silicon Chips

In the past three decades of integrated circuit chip industry,as the size of the chip continues to decrease,the performance of the chip continued to rise.However,with the failure of Dennard's scaling law,the power density of the chip increases as the chip size decreases,resulting in the bottleneck of the power wall.In order to continue to increase the operating frequency of the chip,the development of chip has moved towards the direction of multi-core.However,multi-core does not fundamentally eliminate the effects of the increase in static power proportions that lead to the failure of Dennard's scaling law.In recent years,the number of cores of the chip has gradually increased,and the power density of the chip has also increased and is approaching the power wall again.This phenomenon that causes multiple cores of the chip to not work at the highest frequency at the same time is called the dark silicon phenomenon.Therefore,runtime power management for this many-core dark silicon chip has become very important.In the research of this article,we propose a runtime power budget technology for many-core dark silicon chips.First,because the technology should be applicable to many-core chips,a local thermal model is constructed for each core.Based on this local model,a distributed runtime power budget algorithm with low computational time complexity is proposed.The technical method is roughly divided into three parts: first determine the location of the active core;then according to the runtime power budget algorithm,each open core can independently calculate the power budget;finally,match the voltage and frequency of the processor core according to the power model.This technology firstly distributes the active cores evenly in runtime power management to avoid reliability problems caused by the hotspot issues of multi-core chips,and then,the power budget calculated according to the safe temperature threshold is to maximize the performance of many-core dark silicon chips.At the same time,the overall average temperature of the chip is controlled below the safe temperature threshold.Because the core selection algorithm is used in many-core dark silicon chips,the active cores are first evenly distributed,which improves the heat dissipation capacity of each core and provides performance potential for subsequent runtime power budget technology.At the same time,each active core is a distributed model that only needs to communicate with adjacent cores to perform runtime power management,which makes the calculation time complexity of this technology far lower than traditional lumped power budget technology.In the end,the comparative experiment also proved that the technology has a small computational expense,and it also brings higher performance to the entire many-core dark silicon chip system.Although this technology is a distributed algorithm based on a local chip model,it is slightly inferior to the traditional lumped power management method in terms of the final management temperature accuracy,but experiments show that the error of the management temperature is controlled within a safe range,while the performance and speed both are better than the existing power budget technology.