Learning Peak Temperature in 3DICs by Deep Differentiable Forest

Machine learning is a very attractive idea for thermal problems from 3D integrated circuit. This paper study the peak temperature problem under various power configurations. We give the detailed model and algorithm of deep differentiable forest. A complex 3DIC device with hybrid bonding and heterogeneous technology nodes integration is used to test this model. Commercial finite element tool needs million grids to get high accuracy solution. Compare to FEM tool, our model is hundreds of times faster and the relative error is less than 1%. We also compare its performance with GBDT and random forest model. The experimental results show differentiable forest is a model with higher accuracy in predicting peak temperature.