Leakage-aware Kalman filter for accurate temperature tracking

Due to the effect of technology scaling, leakage power now consists of a significant portion of the total power consumption of a silicon chip. Leakage power also increases exponentially with chip temperature, while temperature itself is a strong function in total power (positive feedback effect). Most of the existing techniques for estimating the runtime chip temperature do not consider the nonlinear leakage effect. This could lead to many problems such as under-estimation of the real chip temperature, improper thermal control actions and eventually, unreliable chip behavior. In this paper we discuss two linearization techniques that can be used to extend the existing thermal tracking approaches and explicitly account for the leakage effect. The first one uses Taylor series expansion to approximate leakage power to the first order (extended Kalman filter). The second one uses concepts from probabilistic matching. Both methods can approximate leakage power with high accuracy while maintaining similar computational efficiency compared to the standard Kalman filter. The experimental results demonstrated that our approaches can reduce the temperature estimation error by 60%, thus significantly improving the thermal-awareness of chip system and enhancing the performance of many dynamic power/thermal management techniques.