热的
热阻
堆栈(抽象数据类型)
接头(建筑物)
材料科学
带宽(计算)
炸薯条
缩放比例
计算机科学
电子工程
电气工程
光电子学
结构工程
工程类
电信
热力学
物理
数学
程序设计语言
几何学
作者
Taehwan Kim,Jonggyu Lee,Youngdeuk Kim,Hwanjoo Park,Hee-Jung Hwang,Jaechoon Kim,Hangi Jung,Dong‐Wook Kim
标识
DOI:10.1109/ectc51909.2023.00133
摘要
As the demands for implementing High Performance Computing (HPC) increase rapidly, the bandwidth and capacity required for High Bandwidth Memory (HBM) are expected to increase by two to three times per generation. Owing to these increased requirements, the power of the next HBM is expected to exceed 30W, and the number of stacks of the HBM leads to a high-level stack demand of 12 and beyond, which also increases the physical thermal resistance of the HBM. Therefore, it is inevitable to strive to improve the thermal performance of HBM. Herein, we demonstrate thermally enhanced joint of HBM with state-of-the-art Thermocompression (TC) bonding. We found that the joint thermal resistance with TC was 35% less than that with Mass reflow (MR) bonding because the joint thickness reduction is more effective to reduce the resistance and that is possible with TC. In addition, even when the joint thermal resistance is almost zero by applying the Hybrid Copper Bonding (HCB), thermal challenge is still exist in the case of high-level stack and high power condition. Therefore, we added analysis on how to operate the HBM below the refresh limit temperature with solution and conditions such as power reduction, on-chip optimization and temperature sensor location.
科研通智能强力驱动
Strongly Powered by AbleSci AI