The High Performance LINPACK for Accelerator Introspection (HPL-AI) benchmark seeks to highlight the convergence of HPC and AI workloads based on machine learning (ML) and deep learning (DL) by solving a system of linear equations using novel, mixed-precision algorithms that exploit modern hardware. While traditional HPC focuses on simulation runs for modeling phenomena in a variety of scientific disciplines, the mathematical models that drive these computations mostly require 64-bit However, the ML/DL methods that fuel advances in AI achieve the desired results at 32-bit or even lower precisions. This lesser demand for working precision fueled a resurgence of interest in new hardware accelerators that deliver a mix of unprecedented performance levels and energy savings to achieve the fidelity in classification and recognition tasks afforded by higher-accuracy formats on traditional hardware.
HPL-AI strives to unite these two realms by connecting its solver formulation to the decades-old HPL framework of benchmarking supercomputers. A number of large-scale HPC installations—including some machines on the TOP500 have now been benchmarked with HPL-AI, starting with Oak Ridge National Laboratory's Summit machine in 2019 and now including RIKEN's Fugaku supercomputer, which achieved 2 Eflop/s in mixed-precision performance. The growing list of machines with both HPL and HPL AI results was 18-entries long as of the November 2021 release of the HPL-AI's bi-annual ranking.
Find out more at https://icl.bitbucket.io/hpl-ai/