Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/60989
Title: | "三維多核心系統晶片之熱流模型,分析與溫度管理" Thermal modeling, analysis and management in 3D ICs |
Authors: | Hitoshi Mizunuma 水沼仁志 |
Advisor: | 楊佳玲(Chia-Lin Yang) |
Co-Advisor: | 盧奕璋(Yi-Chang Lu) |
Keyword: | 熱流模型,溫度管理,多核心系統,三維晶片, Thermal modeling,Thermal management,Multi-Processor System-on-Chip,3-Dimensional Integration, |
Publication Year : | 2013 |
Degree: | 博士 |
Abstract: | 3D stacking with TSV technology is expected to change the IC design paradigm. It not only allows us to integrate heterogeneous components but also overcomes the limitations with respect to delay, bandwidth, and energy consumption of the interconnects. Hence, the whole microelectronic systems previously requiring multiple discrete components such as processors (cores), memories, sensors, and RF devices on the printed circuit board could be now built in a form of the single systems-on-a-chip. The high operating temperature and large thermal imbalance are one of the major design constraints for aggressively stacked 3D ICs, especially for multi-core systems. It is difficult to remove heat from inside of ICs as cores are located at different layers and have significantly different heating/cooling rates in comparison to conventional 2D ICs. 3D multi-core systems also suffer from large temporal and spatial temperature fluctuation due to workload variations by time and among cores. IC designers proposed a number of design-time and run-time thermal management techniques for 3D multi-core systems. However, IC-level thermal management alone is not enough to keep 3D ICs away from thermal emergency as the trend of ever increasing number of cores and stacked layers pushes the power density to the limit of conventional fan-based air-cooled heatsink. Microchannel liquid cooling is an attractive solution for massive heat removal. The underlying concept is to embed the microchannels into the silicon, which acts as a local thermal ground. Shorter vertical distance over which the heat travels around improves heat conduction within solids. In the meant time, the use of channel-flow produces a new type of heat transportation in fluid-solid boundary: heat convection. Furthermore, the heat injected to fluid is moved away through the channels as a result of another type of heat transportation within fluid: heat advection. The extra heat dissipation paths complicate the heat propagation pattern within the microchannel-cooled 3D ICs. The heat generated by a core in the conventional air-cooled 3D ICs propagates only at almost vertical direction because of the back-side attached heatsink that plays dominate role in heat removal. On the other hand, the heat generated by a core in microchannel-cooled 3D ICs propagates vertically via solids as well as horizontally through the fluid. The new cooling technology brings us two challenges. The first challenge is the development of a fast and accurate thermal model to facilitate temperatureaware 3D-IC design. Thermal modeling for microchannel cooling is computationally costly because it requires numerical simulation to solve the conjugate heat transfer (conduction among solids, advection in fluid, and convection between them). The second challenge is the development of thermal management techniques suitable for the microchannel-cooled 3D-ICs. Existing OS-based thermal-aware task scheduler for conventional air-cooled 3D ICs is aware of the vertical heat propagation pattern, thus capable of vertically balancing power and temperature profiles by simply allocating higher-power tasks to the cores closer to the heatsink. However, those schedulers are apparently not effective on microchannel-cooled 3D ICs. In this thesis, we handle the thermal modeling, analysis and management of 3D ICs. The targeted cooling solution for 3D ICs is an integrated microchannel heatsink. The thermal management scheme to be studied is an OS-based on-line task allocation. This thesis is divided into two parts. In the first part, we propose a modeling methodology for microchannel-cooled 3D ICs. Our novel approach, which is our key contribution in this thesis, is that we decouple heat advection from heat conduction/convection, and model the heat advection component using a new kernel function that is extracted off-line and reusable. The advantage of our methodology is that we can re-integrate the kernel function into the conventional heat conduction model that can be solved as efficient as the conventional resistance network model, yet greatly improves the modeling accuracy. Next, to achieve further speedup, we propose Channel-merging technique: multiple microchannels sharing the same core are modeled as a virtual single channel with much fewer grids, which is our second contribution. It is based on the idea that grid density finer than power granularity does not add more useful information from IC design point of view. Validation results show that the modeling error is less than 5 % with more than 3,300x speedup. Finally, we demonstrate the value of our model in thermal-aware placement that reduces the peak temperature by more than 20◦C, which is our third key contribution. In the second part, we present a fast and near-optimal thermal propagation aware task allocation algorithm for microchannel-cooled 3D many-core systems. Our key contribution behind is to derive analytical-model-based guidelines separately in vertical and horizontal direction, and incorporate them into the task allocation algorithm. The most important guideline derived in this study is that the flow descendant order always returns lower hotspot than the flow ascendant order when a certain thermo-hydrodynamical condition is satisfied. The advantage of our idea is that we could evaluate the above condition off-line to know the better allocation policy without knowing the real power consumption of each task. The run-time complexity of the task allocation is as low as O(NlogN) since the on-line task allocation process is automated by those guidelines. The temperature profile obtained is close to optima because those guidelines minimize the vertical and horizontal heat propagations. We compare the obtained temperature profile and the on-line execution time of the proposed scheme against the state-of-the-art near-optimal O(N2logN) algorithm. The experimental results show that the solution obtained by our scheme is very close to the referencing scheme: 0.8◦C and 1.5◦C higher hotspot and spatial thermal imbalance, with 30x - 90x less execution time for 4x4x2 and larger core configurations. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/60989 |
Fulltext Rights: | 有償授權 |
Appears in Collections: | 資訊工程學系 |
Files in This Item:
File | Size | Format | |
---|---|---|---|
ntu-102-1.pdf Restricted Access | 1.69 MB | Adobe PDF |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.