High temperature greatly affects modern processors’ performance, reliability, and user experience. Hence, good thermal design and management policies are needed to maintain low temperatures. A fast and accurate temperature model is essential for the evaluation of these different designs or proactive cooling techniques. State-of-the-art finite element method (FEM) software used in temperature modeling has high accuracy, but its simulation speed is slow. The objective of this work is to develop a fast processor temperature predictor with high accuracy using the learning-based model. In this thesis, steady-state and transient temperature models were both developed and explored. For the steady-state model, a convolutional neural network (CNN) using supervised learning and a physics-informed...[ Read more ]

High temperature greatly affects modern processors’ performance, reliability, and user experience. Hence, good thermal design and management policies are needed to maintain low temperatures. A fast and accurate temperature model is essential for the evaluation of these different designs or proactive cooling techniques. State-of-the-art finite element method (FEM) software used in temperature modeling has high accuracy, but its simulation speed is slow. The objective of this work is to develop a fast processor temperature predictor with high accuracy using the learning-based model. In this thesis, steady-state and transient temperature models were both developed and explored. For the steady-state model, a convolutional neural network (CNN) using supervised learning and a physics-informed neural network (PINN) using unsupervised learning were implemented. Although PINN does not require training data, its training time required is much longer and even longer than the training time plus the training data simulation time for the CNN. In addition, with the same network complexity, the accuracy of PINN is lower than CNN. For the transient model, a recursive non-linear autoregressive network with exogenous inputs (NARX) and a variable time step temperature prediction network (VTSN) were constructed. The recursive NARX was aimed to study multi-step temperature estimation whereas the VTSN was aimed to study variable time step size temperature prediction. The mean absolute error and inference speed of NARX and VTSN are about 1.84 °C, 0.67 °C, and 9.5 ms, 8 ms respectively. Compared to the FEM-based software, there is about 1,263X or 1,250X-2,500X speedup with using NARX or VTSN.