Hotspot Modeling and Mitigation

On-chip hotspots are a negative effect of recent trends in transistor scaling and result in a variety of adverse side- effects, ranging from incorrect circuit operation, bit errors, and reduced device lifespan. In response to these issues, researchers in the computing industry have increasingly focused on techniques to reduce or eliminate such hotspots. These techniques span a variety of technical areas, including Physical Cooling Systems, Circuit and EDA Tools, Software, and Computer Architecture. In spite of these combined efforts, however, hotspots remain an unsolved issue.

We have constructed a toolchain to measure and characterizatizing on Chip Thermal Hotspots.

Existing hotspot mitigation techniques are predominately static, while modern on-chip hotspots are increasingly application dependent, suggesting a need for runtime hotspot mitigation techniques. We propose a distributed hotspot predictor which predicts hotspots based on runtime conditions and prevents them from ever occurring.

Our next stage of the project we are looking at modeling next generation many-core architectures, cooling systems for a future release of HotGauge. In addition, we are studying the impact of thermal hotspots on reliability.

Publications

  • Alexander Hankin (Tufts University), David Werner (Tufts University), Julien Sebot (Intel), Kaushik Vaidyanathan (Google), Maziar Amiraski (Tufts University), Mark Hempstead (Tufts University). HotGauge: A Methodology for Characterizing Advanced Hotspots in Modern and Next Generation Processors. The 2021 IEEE International Symposium on Workload Characterization (IISWC), November 2021. [PDF][Presentation, Slides][Abbreviated-Presentation]
  • Maziar Amiraski*, David Werner*, Alexander Hankin, Julien Sebot, Kaushik Vaidyanathan, Mark Hempstead. (*joint first authors), Boreas: A Cost-Effective Mitigation Method for Advanced Hotspots using Machine Learning and Hardware Telemetry. IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS), April 2023. [PDF]

Funding this project has been supported by Google and Intel.