← sources

Research paper: A Holistic Study of Board Level Reliability of CoWoS-R Advanced Packaging on OCP Accelerator Modules

source: semantic-scholarpublished 2026-05-26 00:00ingested 2026-07-15 07:57

retained text

Query: high bandwidth memory artificial intelligence accelerator
Authors: Sing-Da Jiang, Jing-An Huang, Chi-Shiang Chiou, Shih-Wei Liu, Tsun-Yen Wu
Citations: 0
2.5D CoWoS advanced packaging has been broadly used in applications of artificial intelligence (AI) and high-performance computing (HPC), facilitating the integration of multiple logic dies and memory stacks. Despite its electrical and bandwidth advantages, CoWoS technology introduces challenges to board-level reliability (BLR), due to its large reticle size and complex composite materials stack-up. A major BLR failure mode is ball grid array (BGA) fatigue, exacerbated by the increasing reticle size, which can lead to early failure and compromise the lifetime of the final product. To deep dive into board level performance and failure mechanism of CoWoS-R advanced packaging, temperature cycling testing (TCT), vibration and shock testing were conducted on 3.3X-reticle CoWoS-R packages on OAM board. TCT was identified as a critical modulator of BLR, inducing severe BGA cracks on both substrate and PCB sides. The impact of temperature difference and dwell time on package lifetime during TCT was characterized in our experiment.Based on component and board form factors, a finite element analysis (FEA) model was developed to simulate thermal cycling stress and identify BGA fatigue hot zone on 3.3X CoWoS-R package. The incremental inelastic energy, derived from the Darveaux solder fatigue model, was used as the fatigue failure index to characterize solder ball fatigue risk under various TCT conditions. Further, BGA edge bond dispensing pattern was investigated, and the material of stiffener was optimized. To calibrate BLR TC experimental results, tensile tests and AXP DIC module analysis were performed to determine PCB in-plane Young’s modulus and coefficient of thermal expansion. We also compared fatigue risks from FEA models using bulk, regional, and trace-mapping PCB material properties to study the impact of material mapping resolution on BLR TC simulation results.

document fetched 2026-08-25 06:44 · canonical https://semanticscholar.org/paper/95fd02d9fdd955c9eba6c0378494ce058af7f71d