Transient power analysis to estimate the thermal time lag of a microprocessor hot spot

Abstract

The spatial and temporal non-uniform power of the microprocessor components endangers its functionality and reliability. The rapid change in dynamic power affects electric circuitry, leads to electron migration and high leakage currents; all of which contribute to localized hot spots and high thermal stresses. Though, the dynamic thermal management techniques (DTM) keep the chip temperature within the thermal budget but their efficacy depends on the accurate measurement of the emerging hot spot temperature. The time lag between the peak power input to the microprocessor components and the corresponding hot spot on its surface is therefore a critical parameter for the effective application of DTM techniques. This paper analyses the dynamic power of the microprocessor components to estimate the thermal time lag of the hot spot. The real power estimates of an Intel P4 processor are collected for selected benchmarks, analyzed and then supplied to the numerical model for thermal analysis. In all the benchmarks, the component contributing to hot spot had significantly high heat flux compared to the remaining components of the processor. The power data of the hot component showed periodic peaks of high power which raises the hot spot temperature beyond working limits. The time lapse between peak power input to the time when hot spot temperature becomes 120°C, is defined as thermal time lag for our analysis. The knowledge of thermal time lag can be applied by DTM techniques to prevent the power overshoot in high heat flux components. The thermal time lag can also be exploited to boost the clock frequency repeatedly in cyclic fashion to realize high performance over an extended duration. This repeated boosting can be realized by ensuring that the safe temperature limits are not exceeded in each boosting period and with a sufficient cool-down interval between boosts.