Deep Reinforcement Learning-Based Optimal Decoupling Capacitor Design Method for Silicon Interposer-Based 2.5-D/3-D ICs
Deep Reinforcement Learning-Based Optimal Decoupling Capacitor Design Method for Silicon Interposer-Based 2.5-D/3-D ICs
复制标题
DOI：
10.1109/tcpmt.2020.2972019
复制
发表时间：
2020-03-01
影响因子：
2.2
通讯作者：
Sim, Boogyo
中科院分区：
文献类型：
--
作者：
Park, Hyunwook;Kim, Seongguk;Sim, Boogyo
在这篇文章中，我们首先提出了一种基于深度强化学习（RL）的最佳去耦电容（decap）设计方法，用于基于硅内插器的2.5-D/3-D集成电路（IC）。所提出的方法提供了一个最佳的开盖设计，满足目标阻抗的最小面积。使用基于奖励反馈机制的深度RL算法，可以导出最佳开盖设计准则。为了验证所提出的方法被应用到测试电源分配网络（PDN）和自PDN阻抗与全搜索仿真结果进行了比较。我们成功地验证了全搜索模拟，所提出的方法提供了一个解决方案集。传统的方法是基于复杂的分析模型，从电源完整性（PI）领域的专业知识。然而，所提出的方法仅需要PDN结构和开盖的规范，沿着简单的奖励模型，从而实现快速且准确的数据驱动结果。该方法的计算时间仅为几分钟，比全搜索模拟的计算时间大大缩短，后者需要一个多月。此外，所提出的深度RL方法覆盖了多达10（17）-10（18）种情况，与之前不使用深度学习技术的基于RL的方法相比，大约增加了10（12）-10（13）阶。
In this article, we first propose a deep reinforcement learning (RL)-based optimal decoupling capacitor (decap) design method for silicon interposer-based 2.5-D/3-D integrated circuits (ICs). The proposed method provides an optimal decap design that satisfies target impedance with a minimum area. Using deep RL algorithms based on reward feedback mechanisms, an optimal decap design guideline can be derived. For verification, the proposed method was applied to test power distribution networks (PDNs) and self-PDN impedance was compared with full search simulation results. We successfully verified by the full search simulation that the proposed method provides one of the solution sets. Conventional approaches are based on complex analytical models from power integrity (PI) domain expertise. However, the proposed method requires only specifications of the PDN structure and decap, along with a simple reward model, achieving fast and accurate data-driven results. Computing time of the proposed method was a few minutes, significantly reduced than that of the full search simulation, which took more than a month. Furthermore, the proposed deep RL method covered up to 10(17)-10(18) cases, an approximately 10(12)-10(13) order increase compared to the previous RL-based methods that did not utilize deep-learning techniques.