Home Computing Researchers develop open-source mixed-precision benchmark tool for supercomputers

Researchers develop open-source mixed-precision benchmark tool for supercomputers

by Amelia Ramiro

Researchers at the Oak Ridge National Laboratory (ORNL) have developed OpenMxP, a mixed-precision performance benchmarking tool, in preparation for the launch of the world’s first exascale supercomputer, Frontier. OpenMxP is a cross-platform benchmarking software package that allows for testing lower-precision performance on GPU-accelerated supercomputers. It implements the HPL-MxP benchmarking task, which is the industry standard for measuring mixed-precision performance. OpenMxP was recognized as a finalist in the Software/Services category of the 2023 R&D 100 Awards for its utility in supercomputer assessment and operation.

Mixed-precision arithmetic, which involves using 16 or 32-bit calculations rather than the standard 64-bit double-precision arithmetic, can offer faster speeds while still maintaining the required levels of accuracy for data science and AI applications. However, there has been a lack of widely available, open-source software for testing lower-precision performance at extreme scale on GPU-accelerated supercomputers.

The development of OpenMxP was driven by the need to benchmark Frontier, which is powered by AMD CPUs and GPUs. Previously, benchmarking codes were developed and run by GPU chip vendors on previous supercomputers. However, since Frontier used new-generation chips, new benchmarking codes needed to be developed. The team at ORNL began developing OpenMxP in April 2021 and used the Summit supercomputer, a very fast but not exascale system, as a test bed until Frontier was ready.

When Frontier was ready to run in May 2022, OpenMxP achieved an initial mixed-precision benchmark of 6.86 exaflops (6.86 billion billion floating point operations per second), putting it at the top of the 2022 HPL-MxP list. One year later, it reached 9.95 exaflops for another first-place finish. OpenMxP was also used by the European High-Performance Computing Joint Undertaking’s LUMI supercomputer, which came in second place to Frontier in the June 2023 ranking.

OpenMxP not only provides benchmarking capabilities but also offers insights into how GPU/CPU supercomputers are operating, allowing for improvements in performance by identifying areas where programming changes can lead to computational speed enhancements. It also demonstrates the advantages of using GPU-equipped systems capable of mixed-precision calculations, which can significantly speed up simulations of large systems of equations. The software package can also be used as a tool for solving science and engineering problems at previously unattainable speeds and energy efficiency.

Overall, OpenMxP plays a crucial role in pushing the boundaries of supercomputing by providing an open-source benchmarking tool and solver that can aid in solving large-scale science problems. It enables researchers to build upon and modify the software to address their specific scientific challenges. With its extreme scale and size, OpenMxP helps accelerate scientific advancements and enables the tackling of complex problems in a faster and more efficient manner.

You may also like