SCC Thermal Sensor Characterization and Calibration

Many-cores systems on chip provide the highest performance scaling potential due to the massive parallelism, but they suffer from thermal issues due to the high power densities. Furthermore, workload and process variation requires performance run-time adaptation based on feedback controllers, as open-loop control is not sufficiently robust and accurate. The Single-Chip Cloud Computer (SCC) is an experimental processor created by Intel Labs and, as most multi-core prototypes, it integrates thermal sensors to track the thermal behavior of the die. Unfortunately these sensors are not calibrated, preventing the development of thermal management solutions. In this paper we first extensively characterize the SCC thermal sensors and propose a system level technique to calibrate them. Our approach is based on the combination of stress patterns and least-square fitting to extract the thermal sensor characterization directly from the SCC device. Compared to other strategies this method requires only the knowledge of the ambient temperature under the minimum chip power consumption.