Final Report Summary - MULTITHERMAN (Multiscale Thermal Management of Computing Systems)
The ERC-Advanced MULTITHERMAN project was funded to conduct leading-edge research in optimizing the energy and thermal efficiency of next generation computing systems ranging from small Internet-of-Things processors to supercomputers. The MULTITHERMAN Laboratory, lead by Prof. Luca Benini, has been established in 2013, leveraging a strongly multidisciplinary cooperation of 4 faculty members in the fields of Artificial Intelligence, Operation Research, Control Theory and System Identification. The project, concluded in 2018, has achieved several major outcomes:
First, new design methods and architectures have been developed in the field of near-threshold computing which have led to the design and open sourcing of PULP, the Parallel Ultra-Low Power platform (www.pulp-platform.org). PULP is one of the most successful open-source hardware project worldwide, and it is used for research and commercial products by large companies such as Google, as well as EU fabless start-ups, such as Greenwaves Technologies.
Second, MULTITHERMAN-sponsored researchers have proven the feasibility of advanced model-predictive control for thermal management of high-performance computing systems. The MULTITHERMAN research team has developed the first scalable model predictive controller for Intel Many-core Processors. In addition, researchers of the MULTITHERMAN laboratory made significant contributions to the thermal analysis and modeling of three-dimensional memory-on-logic systems, and demonstrated highly accurate modeling results on a world-first 3D-IC prototype developed by STMicroelectronics and CEA-LETI.
Third, MULTITHERMAN has achieved major milestones in the area of energy-efficient High-performance computing systems, where a strategic cooperation has been established with CINECA, the only Tier-0 computing facility in Italy. MULTITHERMAN researchers have built a complete characterization and energy analysis infrastructure for Eurotech’s EURORA supercomputer, which achieved top ranking in the GREEN500 list of most energy-efficient supercomputer in 2013, and they co-designed (in cooperation with E4) the D.A.V.I.D.E supercomputer, which achieved the 18th position in GREEN500 in 2018.
First, new design methods and architectures have been developed in the field of near-threshold computing which have led to the design and open sourcing of PULP, the Parallel Ultra-Low Power platform (www.pulp-platform.org). PULP is one of the most successful open-source hardware project worldwide, and it is used for research and commercial products by large companies such as Google, as well as EU fabless start-ups, such as Greenwaves Technologies.
Second, MULTITHERMAN-sponsored researchers have proven the feasibility of advanced model-predictive control for thermal management of high-performance computing systems. The MULTITHERMAN research team has developed the first scalable model predictive controller for Intel Many-core Processors. In addition, researchers of the MULTITHERMAN laboratory made significant contributions to the thermal analysis and modeling of three-dimensional memory-on-logic systems, and demonstrated highly accurate modeling results on a world-first 3D-IC prototype developed by STMicroelectronics and CEA-LETI.
Third, MULTITHERMAN has achieved major milestones in the area of energy-efficient High-performance computing systems, where a strategic cooperation has been established with CINECA, the only Tier-0 computing facility in Italy. MULTITHERMAN researchers have built a complete characterization and energy analysis infrastructure for Eurotech’s EURORA supercomputer, which achieved top ranking in the GREEN500 list of most energy-efficient supercomputer in 2013, and they co-designed (in cooperation with E4) the D.A.V.I.D.E supercomputer, which achieved the 18th position in GREEN500 in 2018.