應用於通用圖形處理器上具熱感知及位置相關之三維佈局規劃演算法

Abstract

隨著製程的演進，在縮小尺寸上遇到了瓶頸。因此三維堆疊技術被提出並使用於延續摩爾定律(Moore’s Law)。因為三維堆疊令單位面積功率增加，產生了散熱不易的現象，導致晶片過熱，這使得熱議題(thermal issues)無法被避免。 熱模型(thermal model)被廣泛應用在熱感知(thermal-aware)佈局規劃(floorplan)。經由熱模型，我們可以知道熱的分佈情形；然而精確的熱模型因為執行時間過長，不適合應用在佈局規劃上。然而簡化的熱模型雖然有快速的執行能力，但喪失了精確度。因此我們在這篇論文裡提出一個兼顧快速、且可提供相當程度精確性的熱模型。除此之外，我們提出兩種技術去預防潛在性熱點(hotspot)的產生–斥力(repulsion force)法以及過熱保護區(over-heat prevention zone)機制。經由這些方法，我們使佈局規劃裡的最高溫得到改善，以及令溫度分佈更為均勻。 最後，我們把這個演算法應用在通用圖形處理器上(GPGPU)。藉由大量的核心數去縮短執行時間。

As the development of integration technology, shrinking chip size has encountered bottlenecks. Three-dimensional (3D) integration is then proposed to continue sustaining Moore’s Law. However, due to the significant increasing of the stacking power, it is hard to dissipate heat by using 3D stacking structure. Therefore, the thermal issues become inevitable and has to be handled carefully. In order to solve thermal issues, the thermal model is widely used in thermal-aware floorplan to mimic the temperature distribution. But the runtime of accurate thermal model is too long to be used in floorplan. On the other hand, the simplified thermal model is fast and suitable for floorplan, but it loses accuracy. As a result, in this thesis, we propose a fast thermal model, which also provides a considerably precise temperature estimation. Besides, we also propose two techniques to prevent potential hotspots, named repulsion force (RF) and over-heat prevention zone (OHPZ). By these methods, we not only reduce the maximum temperature but let temperature distribution be more uniform. Finally, we implement this algorithm on GPGPU (General Purpose computing on Graphic Processor Unit). Runtime is reduced by using a great amount of cores.