Thermal- and Power-Aware VLSI Optimization Using Predictive Machine Learning Models for Heterogeneous Systems

Abstract

The increasing model of heterogeneous devices like multi-core CPUs, GPUs, domain-specific
accelerators, and reconfigurable logic in present-day VLSI systems have caused unparalleled
rises in the energy density, heat complexity, and have posed urgent challenges to the energy
efficiency, reliability, and sustainability of performance. Traditional methods of thermal
management are either highly passive or dynamic in nature, and increasingly no longer
capable of dealing with the nonlinear and time-dependent interactions between workload
dynamics, architectural heterogeneity and power-temperature coupling. This paper
presents a multifaceted thermal- and power-aware VLSI optimization model that is driven
by predictive machine learning models to heterogeneous systems. The framework is based
on supervised learning methods to precisely predict short-horizon power usage and spatial
temperatures with an expressive assortment of both run time and design time properties,
such as workload properties, voltage-frequency conditions, utilisation values, and observed
thermal data. These design-time and runtime predictive insights are naturally incorporated
in the processes of design-time and runtime optimization, and, as a result, predictive
thermal-aware floor planning, programmable voltage and frequency scaling and smart task
scheduling of heterogeneous processing cores. The proposed method and technique predicts
thermal violations prior to their happening (compared with traditional performance based
approaches that depend on a threshold), thus reducing sudden performance throttling of the
equipment, as well as thermal stress. The results of a substantial body of experimental analysis
performed on relevant representative heterogeneous system-on-chip benchmarks indicate
that the proposed framework can attain up to 23% (average) power) and 17 performance/
per-watt (compared to the state-of-the-art heuristic-based thermal management schemes).
These findings confirm the effectiveness, scalability and low-overhead of predictive machine
learning-aided optimization that make it a promising solution to next-generation thermally
constrained heterogeneous VLSI systems.