Multi-Scale Thermal Modeling of 3D-Heterogeneous Integrated Processing-Near-Memory Chip for Edge Large Language Model Inference

In this paper, we demonstrated a multi-scale thermal modeling work for 3D-heterogeneous integrated processing-near-memory (PNM) chips. To achieve this, We extracted thermal conductivity using a compact model, reducing simulation time by 95 % combining with FEM. Furthermore, we explored solutions for running LLAMA2 on 3D DRAM PNM chips, and proposed a bank-level refresh scheme to lower temperatures and reduce refresh power by 34 %. We proposed a distributed DRAM solution to further optimize the maximum operating temperature.