Performance Evaluation of LAPACK Using SVE Optimized BLAS Kernels

Optimized mathematical libraries tailored to hardware architectures are critical for maximizing performance in scientific and AI-driven applications. The ARM Scalable Vector Extension (SVE) introduces architectural features such as vector-length-agnostic execution and predicate-driven control, enabling efficient vectorization on modern ARM platforms. This paper evaluates the performance impact of SVE-optimized Level 1 and Level 2 BLAS routines within the OpenBLAS library, used as the computational backend for core LAPACK routines. Experiments conducted on A64FX-based systems demonstrate performance gains ranging from 1.8× to 3.6× across LAPACK functions, and up to 2.0× in real-world applications including LAMMPS and scikit-learn’s Truncated SVD. These results validate the effectiveness of SVE in accelerating linear algebra operations and underscore the importance of architecture-aware kernel optimizations for scalable, portable high-performance computing on emerging ARM platforms.