Area 3: Systems – Hardware, Networking and Networked Systems
Area 3 looks at the basic hardware and software components of computer systems.
Memory Centric Computing: (PIs: Tahoori, Henkel, Koziolek, Reussner, Sanders) No one can deny the fact that technology scaling and smart computer architectures have enabled an explosion of computing performance since the introduction of the first microprocessor by Intel in 1971 with only 2300 transistors to the latest Apple M1 processor in 2022 with more than 114 billion transistors. This has enabled many big data and artificial intelligence applications. According to the International Energy Agency (IEA), ICT energy consumption makes up about 3% of the worldwide electricity production in 2021 with an annual increase of 30%. One big contributor to the computation energy is the data transfer energy between the processor and memory subsystems, the so-called memory wall. The emerging data-intensive and big data applications require huge data transfers between processors and memories, in which data transfer energy dominates data processing energy by a factor of 100 to 1000. Computing in Memory (CIM) is an emerging concept based on the tight integration of traditionally separated memory elements and combinational circuits, allowing minimizing time and energy needed to move data across digital architectures, promoting green and sustainable cloud computing and AI. In order to fully benefit from memory-centric computing, breakthroughs in various levels of the computing stack are needed. In this project, we take a holistic approach by considering all the following aspects: (1) Analog computing for CIM by developing primitives for basic logic, arithmetic and search operations in the memory array. (2) Micro-architecture design space exploration and optimization for CIM-based computing architectures. (3) CIM compiler design and optimization. Performance modeling, simulation, and optimization of the application software running on CIM-enabled computing architecture. (4) Programming languages and models for memory-centric processing. (5) Memory-centric algorithm design and optimization for data-intensive applications. (6) CIM-based hyperdimensional computing for AI applications.
Networking and Networked Systems
Self-organized Networks: (PIs: Hartenstein, Zitterbart) Networked systems and applications are ubiquitously accompanying us in our daily life, be it professional or private. Thus, networks and networked systems are de facto critical infrastructures for the functioning of our digitalized society calling for a highly resilient operation. This is in stark contrast to the fact that their controllability becomes increasingly challenging due their ever growing size and complexity overwhelming administrators that are in charge of their operation. Self-organized, autonomous networks are a vision overcoming this situation. Thus, in this subtopic we will research appropriate fundamental mechanisms for a resilience-by-design approach for networks and networked systems. We will explore the highly decentralized peer-to-peer paradigm at different systems levels and design suited networking mechanisms that support higher-level requirements, e.g., regarding scalable, robust, and efficient multicast and anycast communication patterns. With respect to networking, we will design link-layer peer-to-peer approaches that provide connectivity and routing at large scale, address light-weight, fault tolerant service discovery in such a large scale setting and investigate flat as well as structured large-scale approaches. For networked systems, we will address evaluation and optimization of consensus mechanisms in case of failures, dependability properties of second layer approaches, i.e., approaches that build on top of distributed ledgers, and design and analysis of methods that are available under network partition.
|The Science of Decentralized Messaging||
|Memory-Centric Computing Architecture||