Collaborative calibration of on-chip thermal sensors using performance counters

Thermal sensors are currently deployed in processors to collect thermal information for dynamic thermal management (DTM). The calibration cost for thermal sensors can be prohibitively high as the number of on-chip sensors increases. We propose an on-line multi-sensor calibration method which combines potentially inaccurate temperature values obtained from two sources: temperature readings from thermal sensors and temperature estimations using system performance counters. A data fusion strategy based on Bayesian inference, which combines information from these two sources, is demonstrated along with a temperature estimation approach using performance counters. The approaches are verified via simulation for an AMD Athlon 64 processor with 24 on-chip temperature sensors scaled to a 45nm technology node. Our results show that the standard deviation of temperature sensor measurement errors can be reduced from 3 ~ 4°C to ≤ 1°C using the proposed method. Additionally, our MATLAB implementation shows that the new approach runs at least 67x faster than competing approaches based on Kalman filtering making it highly appropriate for run-time use.