Efficient Peak Temperature Computation with Leakage Power Consideration

隨著現代多核技術在晶片上尋求更多的計算能力, 過熱問題在智慧手機、個人電腦和高端伺服器的晶片設計上是一個關鍵。因此，發展出各式各樣的能源管理演算法在不維背熱能的限制下來達成計算的要求，所有的能源管理機制不僅僅依賴對現下溫度測量，也需考量後面時間點上的溫度預測。然而，多數既存溫度預測演算法假設晶片上的漏電流功率消耗為常數，即並沒有意識在進階多核晶片中越來越重要的漏電流功率對溫度的影響。這篇論文探討溫度和漏電流功率消耗之間在電路上的回饋迴圈機制，並在多核平台上發展出對溫度敏銳的功率模型並提供精確快速的溫度預測。論文所提設的預測策略的好壞將從預測誤差和計算時間等方面來看，由基準測試 (benchmark) 來評估。實驗結果表示出我們的策略比相競爭的類似解法更加精準，能擁有更短的計算時間使得我們的策略適合即時的能源管理演算法。With the modern multi-core technology to explore more computing power in a chip, overheating is now a critical problem of chip designs for smart phones, personal computers, and high-end servers. Therefore, various power management algorithms are developed to meet the computing requirements without violating the thermal constraints, and all of the power management algorithms rely on not only the measurement of the current temperature but also the temperature estimation for the following time points. However, most of the existing temperature estimation algorithms assume that the leakage power consumption of a chip is a constant and are not aware of the impact of temperature on the leakage power consumption which is more and more significant in the advanced multi-core chips. This paper explores the feedback loop between the temperature and the leakage power consumption of a circuit and develops a temperature-aware power model to provide the accurate temperature estimation for multi-core platforms. The performance of the proposed estimation scheme is then evaluated with benchmark suites in terms of the estimation error and the computing time. The experimental results show that our scheme is more accurate than all of the competing solutions, and the short computing time also makes our scheme suitable for real-time power management algorithms