Roofline Analysis and Profiling of HPC Applications to Guide System Design

Learn how to use roofline analysis and performance profiling techniques to characterize High Performance Computing (HPC) and AI scientific applications, ranging from memory-intensive to compute-intensive workloads. Discover methods for collecting and analyzing key performance metrics including compute, memory, and communication characteristics from production HPC clusters like Perlmutter. Explore how performance profiles can guide application tuning for specific hardware platforms and inform the design of next-generation systems. Examine practical approaches for using profiling results to design specialized chiplets and integrate them in modular packages to maximize application efficiency. Gain insights into the relationship between application characteristics and hardware design decisions, with real-world examples from Lawrence Berkeley National Laboratory's research on scientific HPC and AI applications.