Thermal Profiling using Edge Awareness in Heterogeneous Chiplets

Abstract

As heterogeneous chiplet-based architectures become increasingly prevalent in modern computing systems, managing thermal profiles across disparate components presents critical challenges to ensure reliability, performance, and longevity. These chiplets often integrate diverse technologies—such as CPUs, GPUs, AI accelerators, and memory modules—each exhibiting distinct power densities and thermal behaviors. Traditional thermal management approaches relying on centralized monitoring and control are often inadequate for these complex, spatially distributed systems. This paper proposes an edge-aware thermal profiling framework that leverages distributed sensing and localized data processing to enable fine-grained, real-time thermal monitoring across heterogeneous chiplets. The proposed methodology employs a network of embedded thermal sensors strategically placed near chiplet boundaries (“edges”) to capture spatial temperature gradients with high resolution. By integrating lightweight edge computing units, the framework performs on-chip thermal data aggregation, anomaly detection, and predictive modeling locally, reducing latency and communication overhead with centralized controllers. A thermal propagation model based on finite element analysis is incorporated to simulate heat diffusion patterns and validate sensor readings. The framework is evaluated on a heterogeneous chiplet platform comprising CPU, GPU, and AI accelerator chiplets fabricated in 7 nm and 14 nm process nodes.