Performance analysis of Arm-based processors across multiple compilers for HPC workloads.

Arm-based processors are becoming prominent in the high-performance computing (HPC) field. Notable examples include Fujitsu’s A64FX, Amazon Web Services’ Graviton series, and Nvidia’s Grace processor. While the Arm Scalable Vector Extension (SVE) has been implemented in open-source-software compilers such as GNU and LLVM, the practical performance implications for real HPC applications remain insufficiently explored. We conducted performance analyses of various HPC workloads using multiple compilers, including both open-source and proprietary ones, across four Arm-based CPUs: A64FX, Graviton3E, Graviton4, and Grace. To enable consistent profiling across heterogeneous Arm CPUs, we also developed a wrapper for the Linux perf profiling utility that provides a unified interface. This tool enables us to uniformly analyze the key metrics such as SVE use and cache behavior, and it facilitated systematic comparisons of performance across architectures and compilers. Our analyses revealed several performance-related limitations in the current state of the tested compilers, while also providing guidance for the tuning of real applications. These findings provide practical insights for optimizing real HPC workloads with the HPC potential of Arm-based processors and also highlight critical factors that must be addressed to achieve optimal performance when porting applications from other architectures.