As computing platforms evolve into heterogeneous, 3D-integrated many-cores, increasing performance per unit area (or volume) comes unavoidably with growing power density, which becomes heat, leading to degradation, acceleration of chip aging and increase in cooling costs. Thermal dissipation is difficult to control at the micro-scale, where typically spatial and temporal power gradients are orders-of-magnitude higher than at the macro-scale.
MultiThermMan will move beyond the unsustainable worst-case design practices adopted in traditional thermal planning and reactive thermal management. We propose to integrate thermal-aware platform design, thermal control with workload management and shaping in a distributed, multi-scale strategy. This would enable to dynamically adjust the operating mode and active cooling control of each component in complex computing platforms to achieve the highest performance compatible with temperature constraints. The development of a synergistic performance, power and thermal management strategy requires major breakthroughs in several areas, namely architectures, run-time systems, resource management middleware, code optimization tools and programming models. To meet this challenge, MultiThermMan will bring together concepts and techniques from several disciplines: computer architecture and circuits, control theory, combinatorial and continuous optimization. statistical model-building and artificial intelligence. Results will be demonstrated on physical and virtual prototypes, proving practical applicability and relevance for industrial applications.
Field of science
- /natural sciences/computer and information sciences/artificial intelligence
Call for proposal
See other projects for this call