Project Details
Projekt
Balancing computations in near-memory and in-memory heterogeneous systems
Subject Area
Software Engineering and Programming Languages
Computer Architecture, Embedded and Massively Parallel Systems
Computer Architecture, Embedded and Massively Parallel Systems
Term
since 2022
Project identifier
Deutsche Forschungsgemeinschaft (DFG) - Project number 502388442
Compute-in-memory (CIM) and compute-near-memory (CNM) systems, collectively referred to as CINM, overcome the limitations of traditional Von Neumann architectures by significantly enhancing performance and energy efficiency. Despite substantial technological advancements, efficiently utilizing these systems remains a major challenge. In Phase I of SPP-2377, we addressed their programmability issues by developing Cinnamon, a novel compiler technology that integrates optimizations and techniques tailored for various standalone CIM and CNM architectures. However, the CINM landscape and the requirements of emerging applications have become increasingly heterogeneous in recent years. Modern systems now range from domain-specific accelerators with dedicated multiply-accumulate units and analog components to memory-integrated general-purpose CPUs—each presenting unique trade-offs in accuracy, reliability, and data sharing. This project (HetCIM-II) builds upon our strong foundation and deep understanding of these technologies and aims to address these emerging challenges by developing an infrastructure for modeling heterogeneous CINM systems using real-world hardware and simulation. This infrastructure will enable a detailed analysis of these systems, allowing us to develop novel hardware and software optimizations for their efficient utilization. Novel cost models will be designed for CINM architectures, integrated into the Cinnamon framework, and used to guide offloading and optimization decisions. Furthermore, new application and operation mapping strategies, along with domain- and hardware-aware optimizations, will be developed to fully exploit the capabilities of these systems. For evaluation, in addition to standard CINM benchmark suites, workloads from machine learning, bioinformatics, and high-performance computing (HPC) domains will be used. As in HetCIM-I, all optimization and simulation frameworks will be released as open-source to foster community-driven research and innovation in this field
DFG Programme
Priority Programmes
Subproject of
SPP 2377:
Disruptive Main-Memory Technologies
