Dynamic Thermal Management in Many-Core Systems Leveraged by Abstract Modeling

Abstract

For years, transistor size reduction led to a linear decrease in power dissipation. This is the Dennard scaling law, which ended in 2000's technology nodes because the supply voltage no longer scales. The consequence is the increase of the power density in integrated circuits, leading to high temperatures, accelerating aging effects. Dynamic thermal management (DTM) is a technique adopted at runtime to act in the system using the components' current temperature to minimize hotspots and peak temperatures. This paper aims to present a DTM for NoC-based many-cores, using an abstract model, to estimate the temperature at the processing element level and task migration as the main actuation mechanism. Results show up to 12% of temperature reduction and almost 10 o C of peak temperature reduction, highlighting the approach's effectiveness, compared to the pattering and spiral mappings.

PODTherm-GP: A Physics-Based Data-Driven Approach for Effective Architecture-Level Thermal Simulation of Multi-Core CPUs

“…High temperature gradients and hot spots not only impair performance of processors but also degrade their reliability [2], [3]. Thermal management and thermal-aware design exploration of the high-performance processors [4], [5] have been the effective approaches to minimize these thermal issues to improve the performance and reliability of the processors. For instance, as found in [6], the average performance is further improved by 8.9% for the heterogeneous multi-core processor through an adaptive thermal management framework, compared to ARM's DVFS (Dynamic Voltage Frequency Scaling)-based intelligent power allocation.…”