The Energy-Efficient 10-Chiplet AI Hyperscale NPU on Large-Scale Advanced Package

In this paper, we presented an AI hyperscale processing unit (HPU), integrating a pair of neural processing unit (NPU) and 8 high bandwidth memory (HBM) chiplets above a large scale advanced package, redistribution layer (RDL) interposer. We construct the advanced chiplet package platform (CPP) for AI HPU to ensure stable and reliable function. The CPP encompasses NPU-HBM channel design for high speed signaling, fast & accurate power distribution network (PDN) design and analysis, and thermal integrity analysis for efficient cooling structure.