Signal processing methods and hardware-structure for on-line characterization of thermal gradients in many-core processors

In this paper we propose a hardware-structure based on signal processing methods for on-line extraction of on-chip spatial and temporal thermal gradients in many-core chips. The proposed method uses digital filters to compute the magnitude of high-frequency variations in thermal fields and correlate that to thermal gradients. The functionality of the system is demonstrated using full-chip thermal simulation of a predictive 1024 core systems in 11 nm nodes. The power and area of the structure is evaluated using post-layout simulations in predictive 45 nm technology.