Software-based on-chip thermal sensor calibration for DVFS-enabled many-core systems

Due to increase in power density and temperature gradient in modern chips, multiple thermal sensors are deployed on the chip area to provide realtime temperature feedback for fine-grained dynamic thermal management (DTM) techniques. Thermal sensor accuracy is extremely prone to intra-die process variation and aging phenomena, and its report gradually drifts from the nominal value. This necessitates efficient calibration techniques to be applied before the sensor values are used. In addition, in modern many-core systems which are often enabled with dynamic voltage and frequency scaling (DVFS), thermal sensors located on cores are sensitive to the core's current voltage-frequency (VF) level, meaning that dedicated calibration is needed for each VF level. In this paper, we propose a general-purpose software-based auto-calibration strategy for thermal sensors without using any hardware infrastructures for DVFS-enabled many-core systems. We adopt a 2-point calibration method for calculating the calibration constants of each thermal sensor at each VF level. We demonstrate the efficiency of the proposed calibration strategy on a many-core platform, Intel's Single-chip Cloud Computer (SCC), covering all voltage and frequency combinations on the platform.