Strength of Diversity: Exploiting Cheap Heterogeneous Noisy Sensors for Accurate Full-Chip Thermal Estimation and Prediction

Thermal sensor characteristics and placement directly impacts the effectiveness and accuracy of full-chip thermal characterization necessary for dynamic thermal management (DTM) and reliable on-chip operation of multi/many-core chips. Temperature sensor characteristics widely vary in their area, power, and accuracy; the number of deployable sensors is constrained by the on-chip area/power constraints. However, recent approaches have considered only one type of sensor without leveraging the diversities among different sensor types. In this paper, we exploit the flexibility and trade-off in area/power and error characteristics of varied thermal sensors to perform a heterogeneous sensor allocation and placement (HSAP) to precisely recover the full-chip thermal map with high-fidelity. Unlike traditional sensor allocation and placement techniques that consider only one sensor type, HSAP finds the best combination of the heterogeneous sensors along with their placement for a given sensor area and power budget such that the full-chip thermal characterization error is minimized. Experimental results with multicore Alpha processor show significant improvements compared to the state-of-the-art in terms of reconstruction accuracy for the same sensor area and power budget. In particular, our HSAP approach achieves superior accuracy (around 10-100x error reduction with three types of sensors in comparison to a single type without any additional overhead) and execution speedup of over 20× for full-chip thermal monitoring over a state-of-the-art technique.