Collaborative Research: CIRC: ENS: Enabling Detailed, Open-Source Accelerator Modeling

About This Grant

The rapid end of Moore’s Law and Dennard’s Scaling has driven computing systems, from smartphones to supercomputers, to embrace heterogeneous architectures for continued efficiency gains. As specialized, compute-intensive workloads such as machine learning become increasingly prominent, there is a critical need for accurate, open-source simulation tools to model and evaluate the next generation of hardware accelerators. However, the pace of innovation in accelerator architectures, particularly graphics processing units (GPUs), has outstripped the capabilities of existing public simulation frameworks, limiting the research community’s ability to explore new ideas and validate results. This project addresses these challenges by enhancing the widely used Accel-Sim simulation infrastructure, enabling detailed, validated modeling of modern and future accelerators. The proposed enhancements will empower a broad community of researchers to advance innovations in computer architecture, improve system efficiency, and support the development of emerging applications that rely on high-performance accelerators. This award will significantly extend Accel-Sim’s capabilities through three major technical thrusts. First, the project will modernize and expand Accel-Sim’s performance and energy models to support the latest GPU architectures (including NVIDIA’s Ampere, Hopper, and Blackwell), incorporating features such as transformer engines, sparse tensor cores, and support for asynchronous execution. Second, the project will broaden the diversity of accelerators and workloads modeled by Accel-Sim, adding support for GPUs from additional vendors (such as AMD) and integrating with broader system simulation frameworks. Third, the project will develop advanced workload sampling and telescopic level-of-detail modeling to enable scalable, accurate simulation of long-running, compute-heavy workloads. These enhancements will be delivered as robust, open-source tools, accompanied by extensive documentation, community outreach, and training resources to ensure broad accessibility and long-term sustainability. Collectively, these efforts will provide the research community with essential infrastructure to drive the next decade of accelerator innovation and foster a more collaborative ecosystem for computer systems research. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.