Research On Distributed Thermal Management Technology Based On Deep Learning For 3D IC

With the development of integrated circuit technology,the characteristics size of the integrated circuit continues to decrease and the number of core of the integrated circuit also keeps increase,meanwhile the common integrated circuit develops to threedimensional integrated circuits.These factors increase power density of integrated circuit.The problem of thermal reliability of the integrated circuit has become a large problem in the industry.Traditional lumped thermal modeling and thermal management can address this problem when the number of core is small,but when the number of core is large,the size of the thermal model will also change large and the cost of thermal management will change large,which will seriously affect the performance of integrated circuits.Distributed thermal management technology can alleviate this problem and reduce the size of model.Traditional distributed thermal management technology only set up a hot for each core,but the temperature of different position of the core can be difference as high as 4?.That is to say,the temperature of the hot spot location will be well managed and meet the requirements.While the temperature of the other position is likely to not be well managed,which results in the decrease of reliability of integrated circuits.This paper presents a distributed thermal management technology based on deep learning.In this way,we set up 64 hot for each core to ensure that the maximum temperature on the core can be relatively stable near the target temperature that we set up.This thermal model corresponds to a pattern recognition problem.Therefore,this paper uses deep learning technology to fit this thermal model.In this paper,convolutional neural network and recurrent neural network,two commonly networks in deep learning,are used and briefly introduced.Then,these thermal models are implemented for distributed thermal management,in which,extended kalman filter is used to estimate the state value of each core.Power regulator is also used to adjust the power value through grading.An iterative algorithm is proposed to solve the problem of information exchange between thermal management models.Experimental results show that the presented distributed thermal modeling method based on deep learning can effectively reduce the size of the thermal model.The thermal model using the convolution neural network constructed and recurrent neural network constructed is of high accuracy.At the same time,this article propose the technology based on the deep learning 3D IC distributed thermal management method has good control effect and high control precision.It can provide accurate control strategy.