適用以矽智產為基礎之系統晶片設計的高階功率模型研究

On High-Level Power Modeling Approaches for IP-Based SOC Designs

薛文燦(Wen-Tsan Hsieh)

指導教授 ： 劉建男

國立中央大學/資訊電機學院/電機工程學系/博士(2007年)

摘要

在預估複雜數位電路的功率消耗時，常用功率模型的方式來預估，這個方法因為在使用上不需要電路內部詳細的資訊，所以在非常高的設計層次中，便能預估到電路的功率消耗。目前來說，以功率模型為基礎的功率估測方式，似乎是最適合用在系統晶片的設計環境當中，可以根據不同的輸入狀況，快速地提供此功能區塊對應的功率消耗大小。但是一個系統之中通常包含了各種性質不同的功能區塊，如何針對各種區塊的性質差異，建立適合的高階功率模型，也是一個需要深入探討的課題。在這篇論文中，我們針對系統晶片中的主要的電路類型來提出其相對應的高階功率模型的建構方法。在這裡，我們將主要的電路類型區分成四種：組合邏輯電路區塊，序向邏輯電路區塊，嵌入式記憶體電路區塊，處理器電路區塊。
在組合邏輯電路及序向邏輯電路中，我們著力研究在利用類神經網路來建立一種嶄新的功率消耗模型，利用類神經網路來學習輸入資訊及相對應功率消耗的值，進而應用在高階功率估測上。不同於查表法，透過特徵值萃取處理程序後，我們的模組並不會與電路大小有那麼直接的關係，且因為類神經網路的特性，使得這個簡單的模型還可以擁有很好的準確度。從我們的實驗數據裡可看出，在廣泛的輸入變化範圍內，此功率模型依然保有相當的準確度。
在系統晶片中，嵌入式記憶體常被用廣泛地被使用。為了能更正確地估測出整個設計的功率消耗，一個準確的記憶體功率模型是必需的。但是目前電子自動設計工具所使用的傳統記憶體功率模型都過於簡單，導致功率預估的結果不夠精準。在這篇論文裡，我們便針對這方面提出了兩個方法來試著解決此問題。一個是改進的記憶體功率模型，其主要差別在於將非控制輸入信號的變化也納入考量，以便使功率模型更接近實際電路的功率消耗行為模式。另一方面，為了此方法與現今的電子自動設計工具能一起使用，我們也發展了一套冗餘模組方法來克服現今應用軟體上的一些限制。由於我們也將記憶體的大小也列入模型的考量，所以我們提出的方法可以很有效的與記憶體產生器結合，根據不同大小的記憶體能自動調整我們的功率模型，而無需再對不同大小的記憶體再重新建立另一個功率模型。更凸顯出我們方法的實用性。
在處理器的電路區塊中，我們也提出了一套應用在指令層級的功率模型建構方法可應用在管線式超大指令字處理器。在我們的方法中，不論是各個指令的基本功率消耗還是兩道指令間所多出來的額外功率消耗都能被考慮進來。由於我們也考慮管線的效應，這使得我們在處理器上的功率預估能更接近實際的功率消耗行為。整個功率預估的流程主要可以分成兩階段：能量萃取階段及功率重組階段。從我們的實驗數據裡可得知，此功率模型可以有相當的準確度。

關鍵字

高階功率模型 ； 高階功率預估 ； 系統晶片功率預估

並列摘要

Power dissipation of complex digital circuits is often estimated by using power modeling approaches. The usage of this kind of approaches does not require any detailed information of circuits. Therefore, it can be used to estimate the power dissipation of the circuits in very high abstract level. In current approaches, power-model based power estimation methods seem to be more suitable for SOC designs to quickly provide the corresponding power consumption for a given set of input vectors. However, there are still many difficulties to apply existing power models to SOC designs. In addition, a typical system often consists of many blocks with quite different properties, which require different approaches to build their power models. Therefore, in this dissertation, we focus on developing the corresponding high-level power modeling methodologies for each block in SOC designs. These IP blocks can be roughly classified into random logic blocks with combinational and sequential circuits, embedded memory blocks and processor blocks.
For random logic blocks with combinational and sequential circuits, we focus on developing a novel power modeling approach using neural network to estimate the power dissipation on high level. In our approach, the multi-layer feed-forward neural network is used for modeling combinational circuits and multi-layer recurrent neural network is used for sequential circuits. Experimental results on ISCAS’85, ISCAS’89 benchmark circuits and some real designs show that the developed power models achieve good accuracy 4.72% and 4.19% for both combinational and sequential circuits, respectively.
Embedded memories have been used extensively in modern SOC designs. In order to estimate the power consumption of the entire design correctly, an accurate memory power model is needed. However, the memory power model commonly used in commercial EDA tools is too simple to estimate the power consumption accurately. In this dissertation, we develop two methods to improve the accuracy of memory power estimation. Our enhanced memory power model can consider not only the operation mode of memory access, but also the address switching effects with scaling capability. The developed approach is very useful to be combined with the memory compiler to generate accurate power model for any specified memory size without extra characterization costs. Then the developed dummy modular approach can link our enhanced memory power model into the existing power estimation flow smoothly. The experimental results have shown that the average error of our memory power model is only less than 5%.
For micro-processor circuits, we develop a new instruction-level energy modeling approach for pipelined Very Large Instruction Word (VLIW) DSPs. The developed approach can take care of both the base energy cost of each instruction and the additional energy cost of consecutive instructions in each pipeline stage. Therefore, the power estimation can be much closer to the real pipelined behavior since the pipeline issue had been considered. The overall power estimation procedure can be separated into two phases: energy extraction phase and model re-construction phase. The experimental results have shown that the average error of our approach is less than 3% compared to gate-level power simulation.

並列關鍵字

high level power model ； high level power estimation ； soc power estimation ； IP power model

參考文獻

[2] A. Salz, M. Horowitz, “IRSIM: An Incremental MOS Switch-Level Simulator,” in Proc. of ACM/IEEE on Design Automation Conference, pp. 173-178, 1989.

[4] R. Burch, F. N. Najm, P. Yang and T. N. Trick, “A Monte Carlo Approach for Power Estimation,” IEEE Trans. on VLSI System, vol. 1, pp. 63-71, Mar. 1993.

[5] F. N. Najm, “A Survey of Power Estimation Techniques in VLSI Circuits,” IEEE trans. on VLSI Systems, vol. 2, pp. 446-455, Dec. 1994.

[6] P. Landman, “High-Level Power Estimation,” in Proc. of International Symp. on Low Power Electronics and Design, pp. 29-35, 1996.

[7] E. Macii, M. Pedram, F. Somenzi, “High-Level Power Modeling, Estimation, and Optimization,” IEEE Trans. on Computer-Aided Design, vol. 17, pp. 1061-1079, Aug. 1998.